ProgrammableMarketing/aircaps.ai (opens in a new tab)
PDF

Views

  • Plan
  • Roadmap
  • Research

On this page

  • The brief
  • Start here
  • Idea portfolio
  • Effort vs impact
  • This week · 10
  • This quarter · 9
  • Big bets · 3
  • Wildcards · 1
Generated Nov 10, 2026 1 view
  1. Programmable
  2. Playbooks
  3. AirCaps AI
AirCaps AI logo

growth playbook · Nov 10, 2026

aircaps.ai (opens in a new tab)

AirCaps AI growth playbook

Diagnosis

AirCaps AI has the hard part done: models that lead public multi-speaker benchmarks, a low $0.18-per-hour price, open reproduction code and clear docs. The site around them is thin.

It shows no customers, team or company story; it hides the strongest proof the company has (the AirCaps glasses: 5,000+ customers, 40,000+ hours of in-person conversations, WIRED, YC); it gates its #1-ranked Diarization and its streaming model behind sales; and free evaluators cannot opt out of training on the audio they test with.

This playbook adds proof, opens the gated products, removes evaluation friction and makes the benchmarks citable first, then builds SDKs, migration guides, voice-agent plugins, use-case and trust pages, and later bets on on-device deals with hardware makers, a compliance track and an annual real-world speech report.

Growth Score

6/10

Biggest Opportunity

Turn benchmark wins into adoption: show who is behind the models, let developers self-serve everything that is priced, and make switching from an incumbent a one-afternoon job.

Biggest Risk

Deepgram, AssemblyAI and pyannoteAI already have large developer bases, SDKs and voice-agent products, so a five-person team that wins on accuracy can still lose on trust, integrations and response speed.

01

Start here

The three moves to make first. Each one is written up in full below.

  1. 1

    Show the proof aircaps.ai is missing: 40,000+ hours of real conversations and 5,000+ glasses customers from the same team

    Write the one-line proof band and put it under the homepage hero.

    High impact First result in 2 weeks
  2. 2

    Make Diarization self-serve: the #1-ranked model is the only priced one behind "Request access"

    Count how many Diarization access requests came in and how many were refused.

    High impact First result in 1 week
  3. 3

    Let free evaluators opt out of training, because they are testing with their customers' audio

    Check how many free accounts reached the data controls page or contacted support about data use.

    High impact First result in 1 month

02

Your idea portfolio

23 ideas, grouped by when to run them. Open any idea for how to run it: the steps, who to approach, copy you can use, and how to measure it.

  • 8 ideas are marked Unexpected: plays you are unlikely to have tried.

Every idea placed by the effort it takes against the impact it could have. Quick wins sit top left; select a number to open that idea. The filters below narrow the list by status, timing, category and what Programmable can build.

Effort vs impactEach number is an idea. Select one to open it.
Impact →
Quick wins · 6Big bets · 7Fill-ins · 7Reconsider · 3
Effort →

Showing 23 of 23 ideas

This week 10

Quick wins with a clear first action.

  • 01

    aircaps.ai has no customer logos, testimonials, team, about page or company story, and never mentions that the same company makes the AirCaps captioning glasses. Yet the company's own job post says AirCaps has transcribed "40,000+ hours of in-person conversations" in "40+ countries", and aircaps.com shows "5,000+ customers" at 4.9/5, WIRED ("Works almost eerily well"), The New Yorker, Forbes and Y Combinator. A developer comparing a new speech API with Deepgram ("5,000+ organizations") sees strong benchmarks but no answer to "who are these people, and has this ever run in production?" The best answer is on the company's other website.

    High impactLow effort First result in 2 weeks Founder & CEO

    First action

    Write the one-line proof band and put it under the homepage hero.

    How to measure it

    Visitor-to-signup rate

    Now
    –
    Target
    +20%
    By
    60 days

    Steps

    1. 1

      Confirm the claims

      Agree exactly which glasses numbers and press can be reused, and whether the glasses use the same models.

      Founder & CEO 1 day

    2. 2

      Add the proof band

      Under the hero: 40,000+ hours, 40+ countries, 5,000+ customers, press logos, YC F25.

      Founder & CTO 1 day

    3. 3

      Publish /about

      Origin story, founders' audio-AI background, team and contact.

      Founder & CEO 2 days

    4. 4

      Cross-link the sites

      aircaps.com to the API with a short line, and the API footer back to the glasses.

      Founder & CTO 2 hours

    5. 5

      Write the origin note

      "What real conversations taught us" research note, no customer data.

      Founder & CEO 1 week

    6. 6

      Collect API quotes

      Ask the first paying API users for a quote or logo.

      Founder & CEO 2 weeks

    7. 7

      Measure

      Signup rate before and after, and time on /pricing.

      Founder & CTO weekly

    First result
    2 weeks
    Cost
    $0, founder time
    Owner
    Founder & CEO

    Smallest test

    Add the one-line proof band today and compare signup rate for two weeks.

    Why it works

    Add a proof band directly under the homepage hero: "From the team behind AirCaps glasses: 40,000+ hours of in-person conversation transcribed in 40+ countries", with 5,000+ customers, the press logos and YC Fall 2025. Publish a short /about page: why the company started with captions for noisy, face-to-face conversations, the founders' audio-AI background (Yale, Cornell, MIT Media Lab), the team and a contact. Add one research note, "What 40,000 hours of real conversations taught us", describing the conditions (noise, distance, crosstalk) without customer data. Link aircaps.com to the API and back. Then ask the first API users for two or three quotes or logos and add them to the band. Only say the glasses run on the same models if that is true; otherwise say "built by the team that".

    For a new infrastructure vendor, trust is the main reason a developer does not click "Start free" or a buyer does not send audio. The glasses story is real production proof, and it explains the positioning ("hears what other models miss") better than any chart: the models come from a company that had to caption real, messy conversations for paying users.

    Ready to use

    CopyHomepage proof band

    From the team behind AirCaps glasses
    40,000+ hours of in-person conversation transcribed in 40+ countries
    5,000+ customers · Featured in WIRED, The New Yorker and Forbes · Y Combinator F25
    
    We started by captioning real conversations: noisy rooms, people far from the mic, everyone talking at once. That is the audio our models are built for.

    Risks and how to handle them

    • Developers think of AirCaps as a consumer gadget brand

      Lead with the hours and conditions, and keep the glasses as the origin story, not the product.

    • Overstating that the glasses run the API models

      Use "built by the team that" unless the founders confirm the same models.

    Evidence

    • Your websiteHomepage: H1 "Voice AI that hears what other models miss."; "#1 of 15" on VoiceArena's Diarization-Bench, "29%" fewer errors on 5 public meeting benchmarks and "an hour of audio" back in about 10 seconds; "$0.18" per hour with "5 free hours"; use cases include voice agents, voice-agent platforms, meeting notetakers and media and captioning; A5S Streaming "Coming soon". No customer logos, testimonials, team or company story, and no mention of the AirCaps glasses.Source
    • HiringProduct Engineer (San Francisco, $150K-$220K, open on YC's job board): the first product engineer for the AirCaps smart glasses, building on Android XR, iOS and Android with streaming multimodal models. The post states AirCaps is used in "40+ countries", has transcribed "40,000+ hours of in-person conversations", has "75M+ TikTok views and 150K+ followers" and was featured by The New Yorker, WIRED and Forbes.Source
    • AirCaps glasses siteAirCaps Smart Glasses (from $659) show live captions, translation and meeting notes; "5,000+ customers" with a 4.9/5 rating; press from WIRED ("Works almost eerily well"), The New Yorker and Forbes; "backed by Y Combinator"; the footer links to AirCaps AI and its API docs.Source
    • Your websiteSitemap (13 URLs, checked November 9, 2026): the homepage, /models and four model pages, /research and four research notes, /pricing and /contact. No about, blog, customers, use-case, comparison, security or benchmarks page; the footer's "Benchmarks" link points to the homepage anchor "/#benchmarks". The 52 docs pages are listed separately in /docs/llms.txt.Source
    • Competitor researchDeepgram's homepage: "$200" free credit with no card required, "650K+ developers" and "5,000+ organizations", named customer quotes from voice-agent and contact-centre companies, and a Voice Agent API.Source
    • Y CombinatorAirCaps AI on YC: Fall 2025 batch, team of 5, San Francisco, founded 2024, listed under Developer Tools, Speech Recognition and API; the founders studied computer science at Yale and Cornell, and the CEO researched audio AI at the MIT Media Lab.Source
  • 02

    Diarization has a public price ($0.08 per hour) and is one of the four headline models, but the pricing page says "Enabled per account on request" and the API returns "403 model_access_denied" until sales switches it on. A developer who signs up to test the model AirCaps is proudest of hits a wall on their first call. pyannoteAI, the diarization specialist, offers a free trial with no card.

    High impactLow effort First result in 1 week Founder & CTO

    First action

    Count how many Diarization access requests came in and how many were refused.

    How to measure it

    Week-one diarization activation

    Now
    –
    Target
    25% of new accounts
    By
    60 days

    Steps

    1. 1

      Review requests

      How many asked, how many were approved, and why any were refused.

      Founder & CEO 1 hour

    2. 2

      Pick the rule

      Everyone, work emails, or after the first top-up.

      Founder & CTO 1 day

    3. 3

      Switch it on

      Default access within existing limits.

      Founder & CTO 1 day

    4. 4

      Fix the error

      A 403 message that says how to unlock access now.

      Founder & CTO 2 hours

    5. 5

      Update pages

      Pricing, models page and quickstart.

      Founder & CTO 2 hours

    6. 6

      Watch usage

      Diarization jobs per new account and abuse.

      Founder & CTO weekly

    First result
    1 week
    Cost
    Compute for free hours
    Owner
    Founder & CTO

    Smallest test

    Enable Diarization for all new signups for two weeks and compare activation.

    Why it works

    Turn Diarization on for every account, inside the shared 5 free hours and default concurrency limits, so the first request works. If capacity is the worry, enable it automatically after the first credit top-up or for work-email signups, and keep "Request access" only for higher limits. Change the 403 error message to explain how to get access instantly, and replace the button with "Start free".

    Every gate between signup and first result loses developers, and the gated product is the one with the strongest benchmark story. Self-serve access also shows which teams want diarization alone, a good signal for sales.

    Risks and how to handle them

    • Capacity spikes

      Keep the default 5-job concurrency and queue limits.

    Evidence

    • Your websitePricing page: A5S Async "$0.18" per audio hour, Speaker ID "$0.02", Diarization "$0.08" but "Enabled per account on request" behind a "Request access" button; A5S Streaming "Coming soon"; "5 free hours" shared across all models; prepaid credits from $5; volume pricing "starts at 10,000 hours a month" through sales. Its comparison table (checked November 5, 2026) shows AirCaps at $0.18 and 21.5% multi-speaker error (cpWER) against AssemblyAI $0.28 / 32.0%, ElevenLabs $0.27 / 40.1% and Deepgram $0.39 / 66.7%, among eight providers.Source
    • Your websiteDocs (52 pages listed in llms.txt): the quickstart calls the REST API with plain Python requests and cURL, with no SDK package; code samples cover Python, Node.js and cURL; diarization returns "403 model_access_denied" unless enabled for the account; a "Use with AI" page offers an MCP server for coding assistants. There are no migration guides, "switch from" pages or framework integrations.Source
    • Your websiteModels page: A5S Small V2 runs in "Your VPC, on-premises, air-gapped, or on device" with 26.27% WER on noisy, far-field audio and no public price; Diarization is "Enabled on request"; A5S Streaming is "Coming soon, early access for select organizations".Source
    • Competitor researchpyannoteAI's homepage: Precision-3 diarization with "Up to 52% lower error than leading alternatives", the open-source pyannote.audio library with "over 1 billion downloads", a free trial with no card, and on-prem deployment.Source
  • 03

    AirCaps' data docs say customer data is used "by default" to train and benchmark models, and free accounts cannot opt out; only paid accounts can, in the console. But the free 5 hours exist for evaluation, and the teams AirCaps targets (meeting notetakers, contact centres, healthcare, legal) evaluate on their own customers' recordings. Those evaluators either test on public clips that do not show AirCaps' edge, or stop at the data page.

    High impactLow effort First result in 1 month Founder & CEO

    First action

    Check how many free accounts reached the data controls page or contacted support about data use.

    How to measure it

    Free-to-paid conversion

    Now
    –
    Target
    +15%
    By
    90 days

    Steps

    1. 1

      Decide the default

      Opt-out switch for all, or opted out during the trial.

      Founder & CEO 1 day

    2. 2

      Ship the switch

      Data controls for free accounts.

      Founder & CTO 1 day

    3. 3

      Offer an opt-in

      Bonus credit for accounts that choose to share.

      Founder & CEO 1 day

    4. 4

      Update the docs

      Data & privacy, limits and pricing FAQ.

      Founder & CTO 2 hours

    5. 5

      Say it at signup

      One line in the console and on /pricing.

      Founder & CTO 2 hours

    6. 6

      Track

      Opt-in rate and free-to-paid conversion.

      Founder & CEO monthly

    First result
    1 month
    Cost
    $0
    Owner
    Founder & CEO

    Smallest test

    Default new free accounts to opted out for one month and compare paid conversion.

    Why it works

    Give every account the training opt-out switch, or default free accounts to opted out for the first 5 hours. If training data matters, ask instead: an explicit "help improve the models" choice with a credit bonus for those who opt in. State the policy in one line on the pricing page and in the signup flow, next to the 30-day retention default.

    It removes the most common legal question in an evaluation before anyone has to ask it, so evaluators can test with the noisy, real audio where AirCaps wins. An explicit opt-in still collects training data, with cleaner consent.

    Risks and how to handle them

    • Less training data

      Use the credit-for-opt-in offer and keep paid volume terms as they are.

    Evidence

    • Your websiteData & privacy docs: customer data is used "by default" to train and benchmark models; free accounts cannot opt out, paid accounts can opt out in Settings, Data controls, and there is no per-request opt-out. Zero data retention is Enterprise only; a DPA is available through sales; data is processed in the United States; the page makes no SOC 2 or HIPAA claim and lists no sub-processors; speaker profiles should be treated as biometric data.Source
    • Your websiteA5S Async v1 research note (November 2026): 5 public meeting datasets, 226 recordings and 59 hours; macro cpWER of 21.51% against 30.40% for the next-best system; "#1 of 15" on VoiceArena's Diarization-Bench and first in all 10 condition groups; on clean single-speaker speech 6.28% WER, slightly behind the best rival's 5.90%; median 9.8 seconds per hour of audio; code, raw outputs and scores on GitHub; it invites teams building for "meetings, contact centers, media, healthcare, legal, or education". No customers are named.Source
    • Your websiteTalk to sales page: "We reply within 2 business days", then an eval on a sample of the buyer's audio, then pricing. The form asks for current provider, needs (async, streaming, diarization, speaker ID, an eval, partnership), audio per month (from 1k-10k hours up to 1M+ hours) and deployment (cloud API, VPC, on-premise). There is no booking link.Source
  • 04

    Every account gets 5 free hours shared across all models, which at $0.18 an hour is $0.90 of A5S Async. Deepgram's homepage offers $200 of free credit with no card. A team evaluating a meeting product needs dozens of real recordings, and the AirCaps research note itself used 59 hours across five datasets. Five hours is enough to try the API, not to compare it.

    Medium impactLow effort First result in 1 month Founder & CEO

    First action

    Look at how many free accounts used all 5 hours and then stopped.

    How to measure it

    Trial-to-paid conversion

    Now
    –
    Target
    +25%
    By
    90 days

    Steps

    1. 1

      Measure the cliff

      Accounts that used all 5 hours and stopped.

      Founder & CTO 1 day

    2. 2

      Set the rule

      50 hours for work emails, 5 for others.

      Founder & CEO 1 day

    3. 3

      Add the cap

      Per-domain limits and fraud checks.

      Founder & CTO 2 days

    4. 4

      Rewrite the offer

      Pricing page and console copy.

      Founder & CEO 2 hours

    5. 5

      Point to the harness

      Link the benchmark repo for testing on their own audio.

      Founder & CTO 2 hours

    6. 6

      Compare

      Paid conversion by trial size.

      Founder & CEO monthly

    First result
    1 month
    Cost
    About $9 of list-price usage per qualified account
    Owner
    Founder & CEO

    Smallest test

    Give 50 hours to the next 100 work-email signups and compare.

    Why it works

    Keep 5 hours for anonymous signups, but give accounts with a work email 50 free hours (about $9 at list price) on any model, labelled "enough to benchmark on your own audio". Prompt them to run the open benchmark harness on their recordings. Watch abuse with per-domain caps.

    AirCaps wins when buyers test on hard audio, so the trial should be sized for a real test. The cost is small next to a single volume deal, which starts at 10,000 hours a month.

    Risks and how to handle them

    • Free-hour farming

      Work-email check, per-domain caps and no reset.

    Evidence

    • Your websitePricing page: A5S Async "$0.18" per audio hour, Speaker ID "$0.02", Diarization "$0.08" but "Enabled per account on request" behind a "Request access" button; A5S Streaming "Coming soon"; "5 free hours" shared across all models; prepaid credits from $5; volume pricing "starts at 10,000 hours a month" through sales. Its comparison table (checked November 5, 2026) shows AirCaps at $0.18 and 21.5% multi-speaker error (cpWER) against AssemblyAI $0.28 / 32.0%, ElevenLabs $0.27 / 40.1% and Deepgram $0.39 / 66.7%, among eight providers.Source
    • Competitor researchDeepgram's homepage: "$200" free credit with no card required, "650K+ developers" and "5,000+ organizations", named customer quotes from voice-agent and contact-centre companies, and a Voice Agent API.Source
    • Your websiteA5S Async v1 research note (November 2026): 5 public meeting datasets, 226 recordings and 59 hours; macro cpWER of 21.51% against 30.40% for the next-best system; "#1 of 15" on VoiceArena's Diarization-Bench and first in all 10 condition groups; on clean single-speaker speech 6.28% WER, slightly behind the best rival's 5.90%; median 9.8 seconds per hour of audio; code, raw outputs and scores on GitHub; it invites teams building for "meetings, contact centers, media, healthcare, legal, or education". No customers are named.Source
  • 05

    The homepage sells to voice agents and voice-agent platforms, but A5S Streaming is "Coming soon" on the homepage, pricing and models pages, and the streaming guide says it is "available on request to select organizations" through the general sales form. There is no waitlist, so real-time demand is invisible and arrives mixed in with every other sales inquiry.

    Medium impactLow effort First result in 1 week Founder & CEO

    First action

    Add the waitlist form to the pricing page's streaming card.

    How to measure it

    Qualified waitlist signups

    Now
    –
    Target
    100
    By
    60 days

    Steps

    1. 1

      Write the form

      Email, use case, latency, languages, hours.

      Founder & CEO 2 hours

    2. 2

      Place it

      Homepage, pricing, models page and streaming guide.

      Founder & CTO 2 hours

    3. 3

      Route it

      Separate list from general sales.

      Founder & CTO 1 hour

    4. 4

      Onboard batches

      Weekly early-access invites by fit.

      Founder & CTO weekly

    5. 5

      Keep them warm

      Email each new streaming result.

      Founder & CEO monthly

    6. 6

      Read the data

      Rank use cases for the launch.

      Founder & CEO monthly

    First result
    1 week
    Cost
    $0
    Owner
    Founder & CEO

    Smallest test

    Add the form to the pricing page only and count signups for two weeks.

    Why it works

    Put a two-field waitlist (work email, use case) where "Coming soon" appears, with optional latency need, languages and hours per month. Show the queue honestly ("we are onboarding teams weekly"), let waitlisters into early access in batches, and send them each streaming benchmark as it ships.

    A waitlist turns a dead end into pipeline and gives the team real data on which voice-agent use cases to support first. Teams that wait get onboarded warmer.

    Risks and how to handle them

    • Long waits sour leads

      Give a date range and weekly batches.

    Evidence

    • Your websiteHomepage: H1 "Voice AI that hears what other models miss."; "#1 of 15" on VoiceArena's Diarization-Bench, "29%" fewer errors on 5 public meeting benchmarks and "an hour of audio" back in about 10 seconds; "$0.18" per hour with "5 free hours"; use cases include voice agents, voice-agent platforms, meeting notetakers and media and captioning; A5S Streaming "Coming soon". No customer logos, testimonials, team or company story, and no mention of the AirCaps glasses.Source
    • Your websiteStreaming guide: A5S Streaming v1 is "available on request to select organizations" through the contact page; a WebSocket API taking 16 kHz PCM, with speaker labels, 40 custom terms and 30-minute sessions; no voice-agent framework integrations are mentioned and no price is shown.Source
    • Your websiteModels page: A5S Small V2 runs in "Your VPC, on-premises, air-gapped, or on device" with 26.27% WER on noisy, far-field audio and no public price; Diarization is "Enabled on request"; A5S Streaming is "Coming soon, early access for select organizations".Source
  • 06

    Benchmarks are AirCaps' whole argument, but the footer's "Benchmarks" link points to a homepage anchor and the sitemap has no benchmarks page. The results are spread across four research notes and three GitHub repos with 2, 1 and 0 stars. Developers, analysts and AI assistants answering "most accurate speech-to-text for meetings" have no single URL to cite.

    High impactLow effort First result in 1 month Founder & CTO

    First action

    Collect every published table into one page draft.

    How to measure it

    Referring domains to /benchmarks

    Now
    –
    Target
    20
    By
    90 days

    Steps

    1. 1

      Gather tables

      From the four research notes.

      Founder & CTO 1 day

    2. 2

      Build the page

      One leaderboard per task, with dates.

      Founder & CTO 2 days

    3. 3

      Show the weak spot

      Clean single-speaker speech results too.

      Founder & CTO 1 hour

    4. 4

      Link the code

      Repos, manifests and raw outputs.

      Founder & CTO 2 hours

    5. 5

      Index it

      Sitemap, dataset markup, footer link.

      Founder & CTO 2 hours

    6. 6

      Re-run

      Quarterly, with a changelog.

      Founder & CTO quarterly

    First result
    1 month
    Cost
    $0, plus API costs for re-runs
    Owner
    Founder & CTO

    Smallest test

    Publish the A5S Async table as /benchmarks this week.

    Why it works

    Publish /benchmarks: one leaderboard per task (speaker-attributed transcription, diarization, speaker ID, noisy streaming), each with the date, datasets, metric, every provider tested, the one place AirCaps is behind (clean single-speaker speech), and links to code and raw outputs. Add a changelog and re-run quarterly. Mark it up as a dataset and add it to the sitemap.

    A dated, honest leaderboard is the page people link to in comparisons and the page AI assistants quote. Including the one weak spot makes the wins more believable.

    Risks and how to handle them

    • Rivals dispute results

      Publish method, versions and dates, and invite reruns.

    Evidence

    • Your websiteSitemap (13 URLs, checked November 9, 2026): the homepage, /models and four model pages, /research and four research notes, /pricing and /contact. No about, blog, customers, use-case, comparison, security or benchmarks page; the footer's "Benchmarks" link points to the homepage anchor "/#benchmarks". The 52 docs pages are listed separately in /docs/llms.txt.Source
    • GitHubThree public benchmark repos (a5s-async-v1-benchmark, a5s-small-v2-asr-benchmark, AirCaps-Diarization-Benchmark) with 2, 1 and 0 stars.Source
    • Your websiteA5S Async v1 research note (November 2026): 5 public meeting datasets, 226 recordings and 59 hours; macro cpWER of 21.51% against 30.40% for the next-best system; "#1 of 15" on VoiceArena's Diarization-Bench and first in all 10 condition groups; on clean single-speaker speech 6.28% WER, slightly behind the best rival's 5.90%; median 9.8 seconds per hour of audio; code, raw outputs and scores on GitHub; it invites teams building for "meetings, contact centers, media, healthcare, legal, or education". No customers are named.Source
  • 07

    The pricing page already has the comparison buyers search for: AirCaps at $0.18 and 21.5% multi-speaker error against AssemblyAI ($0.28, 32.0%), ElevenLabs ($0.27, 40.1%) and Deepgram ($0.39, 66.7%), checked November 5, 2026. But it sits in one table on one page, and the sitemap has no comparison or alternatives pages. "Deepgram alternative" and "AssemblyAI vs" searches come from teams already shopping.

    High impactLow effort First result in 1 month Founder & CEO

    First action

    Write the Deepgram page from the existing table.

    How to measure it

    Comparison-page signups

    Now
    –
    Target
    50 a month
    By
    90 days

    Steps

    1. 1

      Pick rivals

      The three named most in the sales form's current-provider field.

      Founder & CEO 1 day

    2. 2

      Check facts

      Current prices, features and dates.

      Founder & CEO 2 days

    3. 3

      Write pages

      Fair tables, trade-offs and method link.

      Founder & CEO 1 week

    4. 4

      Add migration

      A short code snippet per rival.

      Founder & CTO 2 days

    5. 5

      Publish

      Linked from /pricing and the sitemap.

      Founder & CTO 1 day

    6. 6

      Refresh

      Each time prices change.

      Founder & CEO quarterly

    First result
    1 month
    Cost
    Founder time
    Owner
    Founder & CEO

    Smallest test

    Publish the Deepgram page and track signups for a month.

    Why it works

    Create one page per main rival (Deepgram, AssemblyAI, ElevenLabs, pyannoteAI for diarization): price like for like, multi-speaker error, diarization, speaker ID, streaming status, deployment options, languages and SDKs, with the date checked and a link to the method. Be honest about where the rival leads (streaming maturity, SDKs, voice-agent products). End each page with a migration snippet and the free trial.

    Comparison searches are high intent, and a fair page that admits trade-offs is the one developers trust and AI assistants cite. AirCaps' price and error gap is large enough to win on facts.

    Risks and how to handle them

    • Out-of-date claims

      Date every table and link the method.

    Builds on

    Evidence

    • Your websitePricing page: A5S Async "$0.18" per audio hour, Speaker ID "$0.02", Diarization "$0.08" but "Enabled per account on request" behind a "Request access" button; A5S Streaming "Coming soon"; "5 free hours" shared across all models; prepaid credits from $5; volume pricing "starts at 10,000 hours a month" through sales. Its comparison table (checked November 5, 2026) shows AirCaps at $0.18 and 21.5% multi-speaker error (cpWER) against AssemblyAI $0.28 / 32.0%, ElevenLabs $0.27 / 40.1% and Deepgram $0.39 / 66.7%, among eight providers.Source
    • Your websiteSitemap (13 URLs, checked November 9, 2026): the homepage, /models and four model pages, /research and four research notes, /pricing and /contact. No about, blog, customers, use-case, comparison, security or benchmarks page; the footer's "Benchmarks" link points to the homepage anchor "/#benchmarks". The 52 docs pages are listed separately in /docs/llms.txt.Source
    • Competitor researchDeepgram's homepage: "$200" free credit with no card required, "650K+ developers" and "5,000+ organizations", named customer quotes from voice-agent and contact-centre companies, and a Voice Agent API.Source
    • Competitor researchAssemblyAI's homepage: "Voice AI infrastructure for developers", with a Voice Agent API, real-time transcription and Guardrails for PII redaction.Source
  • 08

    The Talk to sales page promises "We reply within 2 business days" and has no booking link, even though it collects exactly what is needed to qualify a lead on the spot: current provider, needs, audio per month (up to 1M+ hours) and deployment. A team sending 50k hours a month is worth more than the whole self-serve base, and it is usually talking to two other vendors that week.

    High impactLow effort First result in 1 week Founder & CEO

    First action

    Add the booking link to the form's thank-you state.

    How to measure it

    Form-to-meeting time

    Now
    –
    Target
    Under 1 business day
    By
    30 days

    Steps

    1. 1

      Set the rule

      Which answers count as volume leads.

      Founder & CEO 1 hour

    2. 2

      Add booking

      Calendar link on the thank-you state.

      Founder & CTO 2 hours

    3. 3

      Alert

      Instant message to a founder for volume answers.

      Founder & CTO 2 hours

    4. 4

      Auto-reply

      Ask for a sample file and the current provider's output.

      Founder & CEO 1 hour

    5. 5

      Update the promise

      Same business day for volume leads.

      Founder & CEO 15 minutes

    6. 6

      Track

      Reply time and meetings booked.

      Founder & CEO weekly

    First result
    1 week
    Cost
    $0 to $15 a month
    Owner
    Founder & CEO

    Smallest test

    Add the booking link for 10k+ hour answers this week.

    Why it works

    Show a calendar link right after the form for anyone who picks 10k hours or more, or VPC/on-premise, and alert a founder instantly for those answers. Change the promise to "same business day" for volume leads, and send an automatic reply asking for a sample file so the eval can start before the call.

    Speed to first reply is one of the strongest drivers of B2B win rates, and the eval on the buyer's own audio is AirCaps' best closing tool. Starting it before the call shortens the cycle.

    Risks and how to handle them

    • Founder time

      Only volume answers get the booking link.

    Evidence

    • Your websiteTalk to sales page: "We reply within 2 business days", then an eval on a sample of the buyer's audio, then pricing. The form asks for current provider, needs (async, streaming, diarization, speaker ID, an eval, partnership), audio per month (from 1k-10k hours up to 1M+ hours) and deployment (cloud API, VPC, on-premise). There is no booking link.Source
    • Y CombinatorAirCaps AI on YC: Fall 2025 batch, team of 5, San Francisco, founded 2024, listed under Developer Tools, Speech Recognition and API; the founders studied computer science at Yale and Cornell, and the CEO researched audio AI at the MIT Media Lab.Source
  • 09

    AirCaps' Hugging Face organisation has no public models, but its datasets are used: VOX1O-MUSAN has about 47.1k downloads, and the A5S Small V2 benchmark and noisy-speech datasets about 1.2-1.3k each. The people downloading noisy, overlapping speech data are speech engineers, the exact buyers of a noisy-audio API, and the GitHub benchmark repos they would find next have 0-2 stars.

    Medium impactLow effort First result in 2 weeks Founder & CTO

    First action

    Add the block and tracked link to the VOX1O-MUSAN card.

    How to measure it

    Research-channel signups

    Now
    –
    Target
    40 a month
    By
    60 days

    Steps

    1. 1

      Write the block

      Score, benchmark link, free hours.

      Founder & CTO 1 hour

    2. 2

      Update cards

      All five datasets.

      Founder & CTO 2 hours

    3. 3

      Update READMEs

      The three GitHub benchmark repos.

      Founder & CTO 1 hour

    4. 4

      Build a Space

      Drop a clip, see who spoke when.

      Founder & CTO 3 days

    5. 5

      Track links

      UTM tags per card and repo.

      Founder & CTO 1 hour

    6. 6

      Review

      Signups by source.

      Founder & CEO monthly

    First result
    2 weeks
    Cost
    $0, plus Space compute
    Owner
    Founder & CTO

    Smallest test

    Add a tracked link to the most downloaded dataset card.

    Why it works

    Add a short block to every dataset card: what AirCaps scored on it, a link to the benchmark page and repo, and "Try A5S on this audio: 5 free hours" with a tracked link. Add a Hugging Face Space where visitors drop a noisy clip and see a diarized transcript without signing up, then prompt for an API key. Add the same block to the GitHub READMEs.

    It turns existing research reach into product discovery at zero cost, with an audience that already cares about the problem AirCaps solves.

    Risks and how to handle them

    • Cards read as ads

      Keep the block short and lead with the benchmark result.

    Builds on

    Evidence

    • Hugging FaceNo public models; five datasets, including VOX1O-MUSAN (about 47.1k downloads), the A5S Small V2 benchmark dataset (about 1.29k) and mega-asr-noise-a5sv2 (about 1.25k).Source
    • GitHubThree public benchmark repos (a5s-async-v1-benchmark, a5s-small-v2-asr-benchmark, AirCaps-Diarization-Benchmark) with 2, 1 and 0 stars.Source
  • 10

    Google PageSpeed gives aircaps.ai 90 on mobile and 100 on desktop, so it is fast. Two details stand out: the main content (the H1) appears at 2.7 s on mobile against the 2.5 s target, and Google Fonts make up 277 KB of a 916 KB page. Accessibility scores 94, with ARIA roles missing required attributes and headings out of order. There is no real-visitor data yet, so these are lab results only.

    Low impactLow effort First result in 1 day Founder & CTO

    First action

    Subset and self-host the fonts.

    How to measure it

    Mobile LCP and accessibility

    Now
    –
    Target
    LCP under 2.5 s, accessibility 100
    By
    2 weeks

    Steps

    1. 1

      Subset fonts

      Only the weights and characters used.

      Founder & CTO 2 hours

    2. 2

      Self-host

      Preload with font-display swap.

      Founder & CTO 1 hour

    3. 3

      Fix ARIA

      Required attributes on each role.

      Founder & CTO 1 hour

    4. 4

      Fix headings

      Sequential heading order.

      Founder & CTO 1 hour

    5. 5

      Caption the video

      Captions on the homepage video.

      Founder & CEO 1 hour

    6. 6

      Re-test

      PageSpeed mobile and desktop.

      Founder & CTO 15 minutes

    First result
    1 day
    Cost
    $0
    Owner
    Founder & CTO

    Smallest test

    Self-host the fonts and re-run PageSpeed.

    Why it works

    Self-host and subset the fonts (one or two weights, preloaded, with font-display swap) to bring the H1 under 2.5 s. Fix the ARIA roles and heading order so the site scores 100 on accessibility, and caption any video on the page. Then say so in the footer: an accessibility statement fits a company that started with captioning.

    It is an hour of work, it removes the only lab warning on the page, and an accessible site is a small but credible signal from a company whose origin is access to conversation.

    Risks and how to handle them

    • Font change alters the look

      Keep the same family, fewer weights.

    Evidence

    • PageSpeedGoogle PageSpeed (mobile, tested November 2026): performance 90 (desktop 100); the main content, the H1, appears at 2.7 s against a 2.5 s target; Google Fonts are 277 KB of a 916 KB page; accessibility 94 with two issues, "[role]s do not have all required [aria-*] attributes" and headings out of sequence; there is no real-visitor (field) data yet.Source
    • AirCaps glasses siteAirCaps Smart Glasses (from $659) show live captions, translation and meeting notes; "5,000+ customers" with a 4.9/5 rating; press from WIRED ("Works almost eerily well"), The New Yorker and Forbes; "backed by Y Combinator"; the footer links to AirCaps AI and its API docs.Source

This quarter 9

The core plays for the next three months.

  • 11

    AirCaps' sales process is "an eval on a sample of your audio", done by hand after a reply that can take two business days, and the sales form already asks for the buyer's current provider. The open benchmark code exists on GitHub. With five people, every manual eval is founder time, while self-serve developers never see the side-by-side that wins volume deals.

    High impactMedium effort First result in 2 months Founder & CTO

    First action

    Wrap the existing scoring code as an internal service.

    How to measure it

    Comparison-to-paid rate

    Now
    –
    Target
    30%
    By
    2 quarters

    Steps

    1. 1

      Wrap the scorer

      Benchmark cpWER code as a service.

      Founder & CTO 1 week

    2. 2

      Design the upload

      Audio plus rival or reference transcript.

      Founder & CTO 1 week

    3. 3

      Build the report

      Side-by-side errors and speaker swaps.

      Founder & CTO 2 weeks

    4. 4

      Shareable link

      Read-only report for managers.

      Founder & CTO 3 days

    5. 5

      Route to sales

      Large-volume reports alert a founder.

      Founder & CEO 1 day

    6. 6

      Measure

      Comparison runs and conversion.

      Founder & CEO monthly

    First result
    2 months
    Cost
    About 4 weeks of engineering
    Owner
    Founder & CTO

    Smallest test

    Run the comparison by hand for five free accounts and see if they convert.

    Why it works

    Add a console page where a user uploads up to 10 files and, optionally, the transcripts their current provider returned (or a reference transcript). AirCaps runs A5S Async and scores both with the same cpWER code as the benchmarks, showing errors, speaker swaps and dropped words side by side. Offer a shareable report link so the developer can show their manager, and a "talk to sales" button with the report attached.

    It automates the sales team's best closing step, scales it to every free account, and produces an internal artefact the buyer forwards for you.

    Risks and how to handle them

    • Format differences between providers

      Start with plain text and speaker-labelled text only.

    • Privacy of uploaded files

      Apply the training opt-out and short retention to comparison uploads.

    Builds on

    Evidence

    • Your websiteTalk to sales page: "We reply within 2 business days", then an eval on a sample of the buyer's audio, then pricing. The form asks for current provider, needs (async, streaming, diarization, speaker ID, an eval, partnership), audio per month (from 1k-10k hours up to 1M+ hours) and deployment (cloud API, VPC, on-premise). There is no booking link.Source
    • GitHubThree public benchmark repos (a5s-async-v1-benchmark, a5s-small-v2-asr-benchmark, AirCaps-Diarization-Benchmark) with 2, 1 and 0 stars.Source
    • Y CombinatorAirCaps AI on YC: Fall 2025 batch, team of 5, San Francisco, founded 2024, listed under Developer Tools, Speech Recognition and API; the founders studied computer science at Yale and Cornell, and the CEO researched audio AI at the MIT Media Lab.Source
  • 12

    AirCaps' docs have 52 pages covering every endpoint, but no migration guides or "switch from" pages. A team already running Deepgram or AssemblyAI has code built around that provider's response shape (words, utterances, speaker fields, webhooks), and rewriting it is the real cost of switching, not the price.

    High impactMedium effort First result in 1 month Founder & CTO

    First action

    Write the field mapping for the most common current provider in sales forms.

    How to measure it

    Accounts over 50 hours in 30 days

    Now
    –
    Target
    +30%
    By
    2 quarters

    Steps

    1. 1

      Rank rivals

      From the current-provider form field.

      Founder & CEO 1 day

    2. 2

      Map fields

      Words, utterances, speakers, timestamps.

      Founder & CTO 3 days

    3. 3

      Write adapters

      Python and Node drop-in functions.

      Founder & CTO 1 week

    4. 4

      Parallel-run checklist

      One week side by side.

      Founder & CTO 1 day

    5. 5

      Publish

      In docs and llms.txt.

      Founder & CTO 1 day

    6. 6

      Link

      From comparison pages and sales emails.

      Founder & CEO 1 day

    First result
    1 month
    Cost
    About 2 weeks of engineering
    Owner
    Founder & CTO

    Smallest test

    Publish one mapping table and adapter for the top rival.

    Why it works

    For each of the top three rivals, publish a guide with a field-by-field mapping table, a drop-in adapter function in Python and Node that converts AirCaps' transcript object into the rival's shape, webhook differences, and a checklist for running both in parallel for a week. Link the guides from the comparison pages and the sales auto-reply.

    Lowering switching cost turns "interesting benchmark" into "we can try it this sprint". Adapters also let teams A/B AirCaps in production without a rewrite.

    Risks and how to handle them

    • Rival formats change

      Version the adapters and test them monthly.

    Builds on

    Evidence

    • Your websiteDocs (52 pages listed in llms.txt): the quickstart calls the REST API with plain Python requests and cURL, with no SDK package; code samples cover Python, Node.js and cURL; diarization returns "403 model_access_denied" unless enabled for the account; a "Use with AI" page offers an MCP server for coding assistants. There are no migration guides, "switch from" pages or framework integrations.Source
    • Your websiteTalk to sales page: "We reply within 2 business days", then an eval on a sample of the buyer's audio, then pricing. The form asks for current provider, needs (async, streaming, diarization, speaker ID, an eval, partnership), audio per month (from 1k-10k hours up to 1M+ hours) and deployment (cloud API, VPC, on-premise). There is no booking link.Source
  • 13

    The quickstart builds requests by hand with Python's requests library and cURL, with no SDK package named anywhere in the docs. Polling, webhooks, file uploads over 1 GB (which need an upload URL) and the 303 redirect on long uploads are all left to the developer. Rivals' developer platforms ship SDKs, and many developers judge an API by its first five lines.

    Medium impactMedium effort First result in 1 month Founder & CTO

    First action

    Generate a Python client from the OpenAPI spec and test it on the quickstart.

    How to measure it

    Time to first transcript

    Now
    –
    Target
    Under 10 minutes
    By
    1 quarter

    Steps

    1. 1

      Generate clients

      From openapi.json, Python and Node.

      Founder & CTO 2 days

    2. 2

      Add helpers

      Wait, large upload, retries, webhooks.

      Founder & CTO 1 week

    3. 3

      Test

      Against every guide in the docs.

      Founder & CTO 3 days

    4. 4

      Publish

      PyPI and npm with READMEs.

      Founder & CTO 1 day

    5. 5

      Rewrite quickstart

      SDK first, REST as a tab.

      Founder & CTO 1 day

    6. 6

      Track

      Downloads and time to first transcript.

      Founder & CEO monthly

    First result
    1 month
    Cost
    About 2 weeks of engineering
    Owner
    Founder & CTO

    Smallest test

    Ship the Python SDK only and compare time to first transcript.

    Why it works

    Generate typed Python and Node SDKs from the existing OpenAPI spec, then add the hand-written helpers that matter: transcribe-and-wait, automatic upload URL for large files, webhook signature checking, retries on 429 with Retry-After, and subtitle download. Publish on PyPI and npm, rewrite the quickstart around them and keep the raw REST examples as a tab.

    SDKs cut time to first transcript, remove the edge cases that cause early failures, and make package downloads a visible adoption metric.

    Risks and how to handle them

    • Maintenance load

      Generate from the spec in CI so SDKs update with the API.

    Evidence

    • Your websiteDocs (52 pages listed in llms.txt): the quickstart calls the REST API with plain Python requests and cURL, with no SDK package; code samples cover Python, Node.js and cURL; diarization returns "403 model_access_denied" unless enabled for the account; a "Use with AI" page offers an MCP server for coding assistants. There are no migration guides, "switch from" pages or framework integrations.Source
    • Your websiteLimits docs: by default 5 async jobs process at once and 5 live streaming sessions per account; files up to 3 hours or 5 GB; streaming is in early access; access to a specific model "may require contacting sales"; higher limits and concurrency come through volume deals with sales.Source
  • 14

    The homepage lists voice agents and voice-agent platforms as use cases, and streaming has a documented WebSocket API, but the streaming guide mentions no voice-agent framework and the feature is early access only. Deepgram and AssemblyAI both market Voice Agent APIs. Voice-agent builders mostly pick speech-to-text from the list of plugins their framework already supports.

    High impactMedium effort First result in 2 months Founder & CTO

    First action

    Ask waitlist signups which framework they use.

    How to measure it

    Plugin streaming hours

    Now
    –
    Target
    2,000 hours a month
    By
    2 quarters

    Steps

    1. 1

      Pick frameworks

      From waitlist answers.

      Founder & CEO 1 day

    2. 2

      Build plugins

      Speech-to-text with speaker labels.

      Founder & CTO 3 weeks

    3. 3

      Upstream

      Pull requests to each framework.

      Founder & CTO 2 weeks

    4. 4

      Noisy-room demo

      An example agent in a busy space.

      Founder & CTO 1 week

    5. 5

      Open access

      Early access for plugin users.

      Founder & CEO 1 day

    6. 6

      Measure

      Streaming hours by plugin.

      Founder & CEO monthly

    First result
    2 months
    Cost
    About 1 month of engineering
    Owner
    Founder & CTO

    Smallest test

    Ship one plugin and the noisy-room demo to the waitlist.

    Why it works

    Write and upstream speech-to-text plugins for the two or three open-source voice-agent frameworks developers use most, with AirCaps' edge in front: speaker labels in streaming (who is talking in a group call) and robustness to background noise. Publish an example agent that works in a noisy room, and open early access to anyone who installs the plugin.

    Plugins put AirCaps in the dropdown where the choice is made, and the noisy-room demo shows the exact failure mode voice agents suffer from.

    Risks and how to handle them

    • Streaming capacity

      Keep the 5-session default and raise it for paid accounts.

    Builds on

    Evidence

    • Your websiteStreaming guide: A5S Streaming v1 is "available on request to select organizations" through the contact page; a WebSocket API taking 16 kHz PCM, with speaker labels, 40 custom terms and 30-minute sessions; no voice-agent framework integrations are mentioned and no price is shown.Source
    • Your websiteHomepage: H1 "Voice AI that hears what other models miss."; "#1 of 15" on VoiceArena's Diarization-Bench, "29%" fewer errors on 5 public meeting benchmarks and "an hour of audio" back in about 10 seconds; "$0.18" per hour with "5 free hours"; use cases include voice agents, voice-agent platforms, meeting notetakers and media and captioning; A5S Streaming "Coming soon". No customer logos, testimonials, team or company story, and no mention of the AirCaps glasses.Source
    • Competitor researchDeepgram's homepage: "$200" free credit with no card required, "650K+ developers" and "5,000+ organizations", named customer quotes from voice-agent and contact-centre companies, and a Voice Agent API.Source
    • Competitor researchAssemblyAI's homepage: "Voice AI infrastructure for developers", with a Voice Agent API, real-time transcription and Guardrails for PII redaction.Source
  • 15

    The A5S Async research note invites teams building for "meetings, contact centers, media, healthcare, legal, or education", and the homepage lists meeting notetakers, voice agents and media captioning. But the sitemap has only 13 URLs, with no use-case or industry pages, so a product manager searching "speaker diarization for contact centres" or "captions for sports broadcasts" lands on a rival.

    Medium impactMedium effort First result in 2 months Founder & CEO

    First action

    Draft the meeting-notetaker page using the NOTSOFAR-1 and AMI results.

    How to measure it

    Use-case page conversions

    Now
    –
    Target
    60 a month
    By
    2 quarters

    Steps

    1. 1

      Pick three

      Meetings, media, contact centres first.

      Founder & CEO 1 day

    2. 2

      Match datasets

      The benchmark closest to each setting.

      Founder & CTO 2 days

    3. 3

      Write pages

      Problem, results, options, cost, code.

      Founder & CEO 3 weeks

    4. 4

      Add samples

      A short audio clip and its transcript per page.

      Founder & CTO 1 week

    5. 5

      Publish

      Linked from the homepage use-case section.

      Founder & CTO 1 day

    6. 6

      Expand

      Healthcare, legal, education after the trust page.

      Founder & CEO next quarter

    First result
    2 months
    Cost
    Founder time
    Owner
    Founder & CEO

    Smallest test

    Publish the meeting-notetaker page and track sign-ups.

    Why it works

    Write one page per use case, starting with meeting notetakers, media and captioning, and contact centres: the audio problem in that setting (crosstalk, far-field mics, two-channel calls, crowd noise), AirCaps results on the closest public dataset, the right model and options (subtitles, key terms, Speaker ID), a cost example at typical volume, and a code sample. Add healthcare, legal and education once the trust page exists.

    Use-case pages match how buyers search and let each page show the benchmark that matters to that reader, instead of one general claim.

    Risks and how to handle them

    • Thin pages

      Only publish a page with a real dataset result and sample.

    Builds on

    Evidence

    • Your websiteA5S Async v1 research note (November 2026): 5 public meeting datasets, 226 recordings and 59 hours; macro cpWER of 21.51% against 30.40% for the next-best system; "#1 of 15" on VoiceArena's Diarization-Bench and first in all 10 condition groups; on clean single-speaker speech 6.28% WER, slightly behind the best rival's 5.90%; median 9.8 seconds per hour of audio; code, raw outputs and scores on GitHub; it invites teams building for "meetings, contact centers, media, healthcare, legal, or education". No customers are named.Source
    • Your websiteSitemap (13 URLs, checked November 9, 2026): the homepage, /models and four model pages, /research and four research notes, /pricing and /contact. No about, blog, customers, use-case, comparison, security or benchmarks page; the footer's "Benchmarks" link points to the homepage anchor "/#benchmarks". The 52 docs pages are listed separately in /docs/llms.txt.Source
  • 16

    AirCaps' data docs are clear on retention and encryption, but the site has no security or trust page, makes no SOC 2 or HIPAA claim, lists no sub-processors, and offers the DPA only through sales. The docs also say speaker profiles should be treated as biometric data and that consent is the customer's job. Volume buyers in contact centres, healthcare and legal send a security questionnaire before any eval.

    High impactMedium effort First result in 1 quarter Founder & CEO

    First action

    List current sub-processors and publish the DPA as a PDF.

    How to measure it

    Sales cycle length

    Now
    –
    Target
    -25%
    By
    2 quarters

    Steps

    1. 1

      Collect facts

      Hosting, encryption, retention, sub-processors.

      Founder & CTO 2 days

    2. 2

      Publish the DPA

      Downloadable, no sales call.

      Founder & CEO 1 week

    3. 3

      Write /security

      One page with all controls.

      Founder & CTO 3 days

    4. 4

      Consent guide

      Voice biometrics and Speaker ID.

      Founder & CEO 3 days

    5. 5

      Set the roadmap

      Compliance plan with honest dates.

      Founder & CEO 1 week

    6. 6

      Link it

      Footer, pricing and sales emails.

      Founder & CTO 1 hour

    First result
    1 quarter
    Cost
    Legal review of the DPA; audit costs later
    Owner
    Founder & CEO

    Smallest test

    Publish the sub-processor list and DPA link this month.

    Why it works

    Publish /security with everything already true in one place: US processing, TLS and AES-256, per-account isolation, retention options, zero retention for enterprise, the training policy (after idea 3), a downloadable DPA and the sub-processor list. Add a short guide to voice-biometric consent for Speaker ID users. State the compliance roadmap with honest dates (for example a SOC 2 Type I start) and offer a request form for the questionnaire.

    Security review is often the slowest step in a voice-data deal. A clear public page shortens it and filters out buyers who need a certification AirCaps does not yet have.

    Risks and how to handle them

    • Promising certifications too early

      Only state what is done and what has started.

    Builds on

    Evidence

    • Your websiteData & privacy docs: customer data is used "by default" to train and benchmark models; free accounts cannot opt out, paid accounts can opt out in Settings, Data controls, and there is no per-request opt-out. Zero data retention is Enterprise only; a DPA is available through sales; data is processed in the United States; the page makes no SOC 2 or HIPAA claim and lists no sub-processors; speaker profiles should be treated as biometric data.Source
    • Your websiteSitemap (13 URLs, checked November 9, 2026): the homepage, /models and four model pages, /research and four research notes, /pricing and /contact. No about, blog, customers, use-case, comparison, security or benchmarks page; the footer's "Benchmarks" link points to the homepage anchor "/#benchmarks". The 52 docs pages are listed separately in /docs/llms.txt.Source
  • 17

    AirCaps' job post says the glasses "went viral with 75M+ TikTok views and 150K+ followers". The company already knows how to make short videos about speech recognition that people watch, while the API's only channels are research notes, three low-star GitHub repos and LinkedIn and X profiles.

    Medium impactMedium effort First result in 1 month Founder & CEO

    First action

    Record one noisy four-person clip and run it through three APIs.

    How to measure it

    Video-sourced signups

    Now
    –
    Target
    100 a month
    By
    2 quarters

    Steps

    1. 1

      Collect clips

      Consented or public hard audio.

      Founder & CEO 1 week

    2. 2

      Run the rivals

      Same clip through three APIs.

      Founder & CTO 1 day

    3. 3

      Edit

      Side-by-side transcripts with error counts.

      Founder & CEO weekly

    4. 4

      Post

      X, LinkedIn, Shorts and TikTok.

      Founder & CEO weekly

    5. 5

      Link

      Tracked link to /benchmarks and free hours.

      Founder & CTO 1 hour

    6. 6

      Review

      Views to signups by clip type.

      Founder & CEO monthly

    First result
    1 month
    Cost
    Founder time and small API costs
    Owner
    Founder & CEO

    Smallest test

    Post four clips in a month and count signups.

    Why it works

    Make a weekly 30-second clip series for developer audiences on X, LinkedIn, YouTube Shorts and TikTok: one hard real-world clip (a crowded bar, a car, four people talking at once), the transcripts from AirCaps and two rivals scrolling side by side, and the error counts. End with the benchmark link and free hours. Reuse the glasses team's editing style, and keep the audio public or consented.

    Side-by-side transcripts on hard audio are a visual, shareable form of the benchmark, made by a team with proven short-form instincts. They reach developers who will never read a cpWER table.

    Risks and how to handle them

    • Cherry-picking accusations

      Link each clip to the full benchmark and publish the raw outputs.

    Builds on

    Evidence

    • HiringProduct Engineer (San Francisco, $150K-$220K, open on YC's job board): the first product engineer for the AirCaps smart glasses, building on Android XR, iOS and Android with streaming multimodal models. The post states AirCaps is used in "40+ countries", has transcribed "40,000+ hours of in-person conversations", has "75M+ TikTok views and 150K+ followers" and was featured by The New Yorker, WIRED and Forbes.Source
    • GitHubThree public benchmark repos (a5s-async-v1-benchmark, a5s-small-v2-asr-benchmark, AirCaps-Diarization-Benchmark) with 2, 1 and 0 stars.Source
    • Your websitePricing page: A5S Async "$0.18" per audio hour, Speaker ID "$0.02", Diarization "$0.08" but "Enabled per account on request" behind a "Request access" button; A5S Streaming "Coming soon"; "5 free hours" shared across all models; prepaid credits from $5; volume pricing "starts at 10,000 hours a month" through sales. Its comparison table (checked November 5, 2026) shows AirCaps at $0.18 and 21.5% multi-speaker error (cpWER) against AssemblyAI $0.28 / 32.0%, ElevenLabs $0.27 / 40.1% and Deepgram $0.39 / 66.7%, among eight providers.Source
  • 18

    AirCaps has one public price ($0.18 an hour) and then nothing until "volume pricing starts at 10,000 hours a month" through sales. The sales form's lowest band is 1k-10k hours. A startup at 2,000 hours a month ($360) has no reason to commit, no discount to plan around and no path to higher concurrency (5 jobs at once by default) without a sales conversation.

    Medium impactLow effort First result in 1 month Founder & CEO

    First action

    List paying accounts by monthly hours to see where they cluster.

    How to measure it

    Committed revenue share

    Now
    –
    Target
    30%
    By
    2 quarters

    Steps

    1. 1

      Cluster usage

      Paying accounts by monthly hours.

      Founder & CEO 1 day

    2. 2

      Set tiers

      Hours, discount and limits per tier.

      Founder & CEO 2 days

    3. 3

      Build billing

      Monthly prepaid commitments in the console.

      Founder & CTO 1 week

    4. 4

      Raise limits

      Concurrency per tier.

      Founder & CTO 1 day

    5. 5

      Update pricing

      Tiers and calculator.

      Founder & CTO 1 day

    6. 6

      Offer to users

      Email accounts already near a tier.

      Founder & CEO 1 day

    First result
    1 month
    Cost
    Discount margin and billing work
    Owner
    Founder & CEO

    Smallest test

    Offer one 1,000-hour commitment by email to the top ten accounts.

    Why it works

    Add two self-serve commitments on /pricing: for example 1,000 and 5,000 hours a month prepaid at a modest discount, each with higher concurrency, Diarization and streaming access, the training opt-out and a support channel. Keep 10,000+ hours with sales. Show the cost for each tier in the existing calculator.

    Commitments lift revenue predictability and retention, and higher concurrency is the limit a growing team hits first. Publishing it saves founder time on small deals.

    Risks and how to handle them

    • Discounting accounts that would pay list

      Keep discounts small and tie them to commitment length.

    Evidence

    • Your websitePricing page: A5S Async "$0.18" per audio hour, Speaker ID "$0.02", Diarization "$0.08" but "Enabled per account on request" behind a "Request access" button; A5S Streaming "Coming soon"; "5 free hours" shared across all models; prepaid credits from $5; volume pricing "starts at 10,000 hours a month" through sales. Its comparison table (checked November 5, 2026) shows AirCaps at $0.18 and 21.5% multi-speaker error (cpWER) against AssemblyAI $0.28 / 32.0%, ElevenLabs $0.27 / 40.1% and Deepgram $0.39 / 66.7%, among eight providers.Source
    • Your websiteLimits docs: by default 5 async jobs process at once and 5 live streaming sessions per account; files up to 3 hours or 5 GB; streaming is in early access; access to a specific model "may require contacting sales"; higher limits and concurrency come through volume deals with sales.Source
    • Your websiteTalk to sales page: "We reply within 2 business days", then an eval on a sample of the buyer's audio, then pricing. The form asks for current provider, needs (async, streaming, diarization, speaker ID, an eval, partnership), audio per month (from 1k-10k hours up to 1M+ hours) and deployment (cloud API, VPC, on-premise). There is no booking link.Source
  • 19

    The API already exposes account and usage endpoints, and the console has prepaid credits, auto refill and a hard stop ("402 insufficient_balance") when the balance runs out. Nothing on the site suggests any onboarding beyond the quickstart. With only 5 free hours, a developer who runs out mid-test simply sees an error.

    Medium impactLow effort First result in 2 weeks Founder & CTO

    First action

    Write the "no transcript after 24 hours" email.

    How to measure it

    Week-one activation

    Now
    –
    Target
    +20%
    By
    60 days

    Steps

    1. 1

      Define events

      Signup, first job, 80% hours, first 402.

      Founder & CTO 1 day

    2. 2

      Wire the triggers

      From usage and billing events.

      Founder & CTO 3 days

    3. 3

      Write emails

      Four short founder emails.

      Founder & CEO 1 day

    4. 4

      Add the cURL

      A sample file that works first time.

      Founder & CTO 2 hours

    5. 5

      Log replies

      Into a feedback list.

      Founder & CEO ongoing

    6. 6

      Measure

      Activation and top-up rates.

      Founder & CTO monthly

    First result
    2 weeks
    Cost
    Email tool, under $50 a month
    Owner
    Founder & CTO

    Smallest test

    Send the 24-hour email by hand to this week's signups.

    Why it works

    Send four short, plain-text emails from a founder based on events: no transcript 24 hours after signup (with a one-line cURL and a sample file), first transcript done (try speaker names with Speaker ID, or subtitles), 80% of free hours used (how to add $5 of credit, or apply for the bigger trial), and first 402 error (one-click top-up). Reply-to goes to a founder, and every reply is logged as product feedback.

    Event-based nudges at the moments developers stall are the cheapest retention tool, and founder replies give the qualitative feedback a five-person team needs.

    Risks and how to handle them

    • Too many emails

      Four in total, each stops once the action is done.

    Builds on

    Evidence

    • Your websiteLimits docs: by default 5 async jobs process at once and 5 live streaming sessions per account; files up to 3 hours or 5 GB; streaming is in early access; access to a specific model "may require contacting sales"; higher limits and concurrency come through volume deals with sales.Source
    • Your websitePricing page: A5S Async "$0.18" per audio hour, Speaker ID "$0.02", Diarization "$0.08" but "Enabled per account on request" behind a "Request access" button; A5S Streaming "Coming soon"; "5 free hours" shared across all models; prepaid credits from $5; volume pricing "starts at 10,000 hours a month" through sales. Its comparison table (checked November 5, 2026) shows AirCaps at $0.18 and 21.5% multi-speaker error (cpWER) against AssemblyAI $0.28 / 32.0%, ElevenLabs $0.27 / 40.1% and Deepgram $0.39 / 66.7%, among eight providers.Source
    • Your websiteDocs (52 pages listed in llms.txt): the quickstart calls the REST API with plain Python requests and cURL, with no SDK package; code samples cover Python, Node.js and cURL; diarization returns "403 model_access_denied" unless enabled for the account; a "Use with AI" page offers an MCP server for coding assistants. There are no migration guides, "switch from" pages or framework integrations.Source

Big bets 3

Larger plays that can change the trajectory.

  • 20

    A5S Small V2 runs "in your VPC, on-premises, air-gapped, or on device" with 26.27% WER on noisy, far-field audio, but it has no public price or dedicated page for device makers. AirCaps itself ships smart glasses that have captioned 40,000+ hours of in-person conversation, and it is hiring an engineer for Android XR, iOS and Android. Few speech vendors also build and ship their own wearable.

    High impactHigh effort First result in 2 quarters Founder & CEO

    First action

    Write down the A5S Small V2 footprint, latency and power on the hardware AirCaps already uses.

    How to measure it

    Hardware pilots

    Now
    –
    Target
    3
    By
    2 quarters

    Steps

    1. 1

      Measure the model

      Footprint, latency and power on target chips.

      Founder & CTO 2 weeks

    2. 2

      Write the note

      Running speech recognition on a wearable.

      Founder & CTO 2 weeks

    3. 3

      Build the page

      On-device offer with specs and licensing.

      Founder & CEO 1 week

    4. 4

      Set licensing

      Per-device or per-unit terms.

      Founder & CEO 2 weeks

    5. 5

      List targets

      Glasses, hearing, earbuds, rooms, cars, robots.

      Founder & CEO 1 week

    6. 6

      Pitch

      Through YC and silicon partners.

      Founder & CEO ongoing

    7. 7

      Run pilots

      Paid pilots with clear success metrics.

      Founder & CTO 1 quarter

    First result
    2 quarters
    Cost
    Founder and engineering time; travel
    Owner
    Founder & CEO

    Smallest test

    Publish the on-device note and pitch five hardware teams.

    Why it works

    Package A5S Small V2 as an embedded offer for smart glasses, hearing and earbud makers, meeting-room hardware, cars and robots: supported chips and memory footprint, latency and power numbers, a reference integration drawn from the AirCaps glasses work, and per-device or per-unit licensing. Publish one technical note on running speech recognition on a wearable, then pitch hardware teams and silicon partners directly, using the founders' network and YC.

    On-device licensing deals are large and sticky, the competition from cloud-first APIs is thin, and AirCaps has unusual proof: its own hardware in daily use.

    Risks and how to handle them

    • Long hardware sales cycles

      Start with paid pilots and small teams that ship fast.

    • Distraction from the API

      Cap it at one founder's time until a pilot signs.

    Builds on

    Evidence

    • Your websiteModels page: A5S Small V2 runs in "Your VPC, on-premises, air-gapped, or on device" with 26.27% WER on noisy, far-field audio and no public price; Diarization is "Enabled on request"; A5S Streaming is "Coming soon, early access for select organizations".Source
    • HiringProduct Engineer (San Francisco, $150K-$220K, open on YC's job board): the first product engineer for the AirCaps smart glasses, building on Android XR, iOS and Android with streaming multimodal models. The post states AirCaps is used in "40+ countries", has transcribed "40,000+ hours of in-person conversations", has "75M+ TikTok views and 150K+ followers" and was featured by The New Yorker, WIRED and Forbes.Source
    • AirCaps glasses siteAirCaps Smart Glasses (from $659) show live captions, translation and meeting notes; "5,000+ customers" with a 4.9/5 rating; press from WIRED ("Works almost eerily well"), The New Yorker and Forbes; "backed by Y Combinator"; the footer links to AirCaps AI and its API docs.Source
  • 21

    The research note invites teams building for healthcare and legal, two markets where speaker attribution matters (who said what in a consultation or deposition) and recordings are often multi-party and noisy. But AirCaps makes no HIPAA claim, offers zero retention only to enterprise, and trains on data by default. Those buyers cannot start an evaluation today.

    High impactHigh effort First result in 2 quarters Founder & CEO

    First action

    Get a legal opinion on what a BAA requires from AirCaps' current setup.

    How to measure it

    Regulated-industry ACV

    Now
    –
    Target
    2 signed design partners
    By
    2 quarters

    Steps

    1. 1

      Scope the gap

      BAA, audit logging and retention requirements.

      Founder & CEO 2 weeks

    2. 2

      Fix defaults

      Zero retention and no training for these accounts.

      Founder & CTO 2 weeks

    3. 3

      Add audit logs

      Access and deletion events.

      Founder & CTO 3 weeks

    4. 4

      Find partners

      One medical scribe and one legal team.

      Founder & CEO 1 month

    5. 5

      Run pilots

      On their audio, with clear metrics.

      Founder & CTO 1 quarter

    6. 6

      Publish results

      Case studies with permission.

      Founder & CEO 1 month

    First result
    2 quarters
    Cost
    Legal and compliance costs; engineering time
    Owner
    Founder & CEO

    Smallest test

    Offer one medical-scribe team a zero-retention pilot under a BAA.

    Why it works

    Build a regulated-industries offer: a signed BAA path, zero retention and no training by default for those accounts, on-prem or VPC deployment with A5S Small V2 or a private A5S Async, an audit log, and a published security review. Start with one design partner each in medical scribing and legal transcription, and publish their results (with permission) as the first case studies.

    Regulated buyers pay more per hour, churn less and value exactly what AirCaps is best at: getting speakers right in messy multi-party audio.

    Risks and how to handle them

    • Compliance costs before revenue

      Sign a design partner before paying for an audit.

    Builds on

    Evidence

    • Your websiteA5S Async v1 research note (November 2026): 5 public meeting datasets, 226 recordings and 59 hours; macro cpWER of 21.51% against 30.40% for the next-best system; "#1 of 15" on VoiceArena's Diarization-Bench and first in all 10 condition groups; on clean single-speaker speech 6.28% WER, slightly behind the best rival's 5.90%; median 9.8 seconds per hour of audio; code, raw outputs and scores on GitHub; it invites teams building for "meetings, contact centers, media, healthcare, legal, or education". No customers are named.Source
    • Your websiteData & privacy docs: customer data is used "by default" to train and benchmark models; free accounts cannot opt out, paid accounts can opt out in Settings, Data controls, and there is no per-request opt-out. Zero data retention is Enterprise only; a DPA is available through sales; data is processed in the United States; the page makes no SOC 2 or HIPAA claim and lists no sub-processors; speaker profiles should be treated as biometric data.Source
    • Your websiteModels page: A5S Small V2 runs in "Your VPC, on-premises, air-gapped, or on device" with 26.27% WER on noisy, far-field audio and no public price; Diarization is "Enabled on request"; A5S Streaming is "Coming soon, early access for select organizations".Source
  • 22

    AirCaps already runs the largest part of this work for each release: eight to fifteen systems across meeting, noisy far-field and overlapping-speech datasets, with prices checked on a date and raw outputs published. Today it appears as product research notes, which rivals and press can dismiss as vendor marketing, and it is not repeated on a schedule.

    High impactHigh effort First result in 1 quarter Founder & CTO

    First action

    Choose the condition set and invite one outside reviewer.

    How to measure it

    Report citations

    Now
    –
    Target
    50 referring domains
    By
    90 days after launch

    Steps

    1. 1

      Fix the method

      Conditions, datasets, metrics and price rules.

      Founder & CTO 2 weeks

    2. 2

      Invite a reviewer

      Academic or third-party benchmark host.

      Founder & CEO 1 month

    3. 3

      Run all systems

      Same audio, same dates.

      Founder & CTO 1 month

    4. 4

      Write the report

      Findings, charts, year-on-year change.

      Founder & CEO 3 weeks

    5. 5

      Release

      Webinar, short videos and press.

      Founder & CEO 1 week

    6. 6

      Repeat

      Same method each year.

      Founder & CTO yearly

    First result
    1 quarter
    Cost
    API costs for all systems; about 2 months of team time
    Owner
    Founder & CTO

    Smallest test

    Publish a quarterly update of the existing A5S table with a fixed method.

    Why it works

    Once a year, publish an independent-style report: every major speech API on the same real-world conditions (noise, distance, overlap, accents, languages), cost per accurate hour, speed, and year-on-year change, with all code and outputs. Invite an academic co-author or third-party benchmark host to review the method. Release it with a webinar and a short-video series, and keep AirCaps' own result in the table without special treatment.

    A recurring, method-first report becomes the reference the market cites, earns press and links each year, and keeps AirCaps at the centre of the "real-world accuracy" conversation it wants to own.

    Risks and how to handle them

    • Seen as biased

      Outside reviewer, full method and raw outputs.

    Builds on

    Evidence

    • Your websiteA5S Async v1 research note (November 2026): 5 public meeting datasets, 226 recordings and 59 hours; macro cpWER of 21.51% against 30.40% for the next-best system; "#1 of 15" on VoiceArena's Diarization-Bench and first in all 10 condition groups; on clean single-speaker speech 6.28% WER, slightly behind the best rival's 5.90%; median 9.8 seconds per hour of audio; code, raw outputs and scores on GitHub; it invites teams building for "meetings, contact centers, media, healthcare, legal, or education". No customers are named.Source
    • GitHubThree public benchmark repos (a5s-async-v1-benchmark, a5s-small-v2-asr-benchmark, AirCaps-Diarization-Benchmark) with 2, 1 and 0 stars.Source
    • Your websiteModels page: A5S Small V2 runs in "Your VPC, on-premises, air-gapped, or on device" with 26.27% WER on noisy, far-field audio and no public price; Diarization is "Enabled on request"; A5S Streaming is "Coming soon, early access for select organizations".Source

Wildcards 1

Unexpected experiments, each with a cheap way to test it.

  • 23

    AirCaps' pitch is "every result is backed by an open benchmark", and its harness is public on GitHub, but the repos have 0-2 stars and every published result so far comes from AirCaps itself. The most convincing evidence for a new buyer is a third party's test on audio like theirs, including the cases where AirCaps loses.

    Medium impactLow effort First result in 1 month Founder & CTO

    First action

    Write the bounty rules and a one-page "run it on your audio" guide in the benchmark repo.

    How to measure it

    Published third-party tests

    Now
    –
    Target
    10
    By
    2 quarters

    Steps

    1. 1

      Write rules

      What counts, credit amount, publish win or lose.

      Founder & CEO 1 day

    2. 2

      Simplify the harness

      One command to run on a folder of audio.

      Founder & CTO 1 week

    3. 3

      Announce

      Repo README, X, LinkedIn and research groups.

      Founder & CEO 1 day

    4. 4

      Pay credits

      On publication.

      Founder & CTO ongoing

    5. 5

      Feature results

      On /benchmarks, including losses.

      Founder & CTO monthly

    6. 6

      Respond

      Public notes on fixes for losses.

      Founder & CTO ongoing

    First result
    1 month
    Cost
    About $18 of list-price usage per 100-hour bounty
    Owner
    Founder & CTO

    Smallest test

    Offer the bounty in the benchmark repo README for one month.

    Why it works

    Offer API credit (for example 100 hours) to any developer or team who runs the open harness on their own or public audio, against at least one other provider, and publishes the method and results, win or lose. Feature the best ones on the benchmarks page, and fix and publicly answer any loss. Seed it with a few speech-research groups and developer communities.

    Third-party results are more credible than vendor charts, they spread through the authors' own networks, and losses give the team real product feedback.

    Risks and how to handle them

    • Low-quality write-ups

      Require the method, audio description and raw outputs.

    Builds on

    Evidence

    • GitHubThree public benchmark repos (a5s-async-v1-benchmark, a5s-small-v2-asr-benchmark, AirCaps-Diarization-Benchmark) with 2, 1 and 0 stars.Source
    • Your websiteHomepage: H1 "Voice AI that hears what other models miss."; "#1 of 15" on VoiceArena's Diarization-Bench, "29%" fewer errors on 5 public meeting benchmarks and "an hour of audio" back in about 10 seconds; "$0.18" per hour with "5 free hours"; use cases include voice agents, voice-agent platforms, meeting notetakers and media and captioning; A5S Streaming "Coming soon". No customer logos, testimonials, team or company story, and no mention of the AirCaps glasses.Source
    • Your websiteA5S Async v1 research note (November 2026): 5 public meeting datasets, 226 recordings and 59 hours; macro cpWER of 21.51% against 30.40% for the next-best system; "#1 of 15" on VoiceArena's Diarization-Bench and first in all 10 condition groups; on clean single-speaker speech 6.28% WER, slightly behind the best rival's 5.90%; median 9.8 seconds per hour of audio; code, raw outputs and scores on GitHub; it invites teams building for "meetings, contact centers, media, healthcare, legal, or education". No customers are named.Source
More example playbooks and reviews

More growth playbooks

Other companies we have researched, each read from what they are hiring for.

  • Sezzle growth strategyA public US fintech offering pay-later plans (Pay in 4, Pay in 5, Pay Monthly), subscriptions (Premium, Anywhere), a virtual card, credit…22 growth ideas
  • Tennant Products growth strategyA Colleyville, Texas supplement company selling nutrition products developed under Dr. Jerry Tennant's guidance: Restore protein formulas…21 growth ideas
  • InHawk IT Solutions growth strategySoftware for colleges and universities: student and campus management, digital evaluation of exams, OMR services, fund-based accounting…22 growth ideas
  • Quisitive growth strategyA Microsoft partner for AI, data, security, cloud, managed services and business applications (Dynamics 365, Power Platform, Board…22 growth ideas
  • Warehouse-Lighting.com growth strategyAn online LED lighting distributor and manufacturer based in New Berlin, Wisconsin: industrial, commercial, outdoor, architectural and…22 growth ideas
  • Veriff growth strategyIdentity verification and fraud prevention: document and biometric checks, liveness, biometric authentication, AML screening, proof of…21 growth ideas
  • Rankfast growth strategyA search and performance marketing agency (Rankfast by Coursenator): SEO, AEO and AI search visibility, Google and Meta ads, conversion…22 growth ideas
  • CloseCore growth strategyAn AI-powered financial close management platform: month-end checklists, balance sheet reconciliations, reporting and variance analysis…21 growth ideas

See every playbook and website review

Independent research. Programmable wrote this playbook from public sources, as they stood on November 10, 2026, and cites them throughout. AirCaps AI did not commission, review or endorse it, and Programmable is not affiliated with AirCaps AI. Company and product names are trademarks of their owners, used only to identify them. The ideas are suggestions and opinion, not professional, legal or financial advice, and not statements about how AirCaps AI runs its business.

Something wrong, or do you represent AirCaps AI? Email [email protected] and we will correct it or take it down.

30/60/90 roadmap

What to focus on in each month, and the ideas that carry it.

  1. Days 1 to 3010 ideas

    Proof on the homepage, self-serve Diarization, a training opt-out and bigger trial for evaluators, a streaming waitlist, a real benchmarks page, comparison pages, faster sales replies, Hugging Face links and the accessibility fixes.

  2. Days 31 to 609 ideas

    Run-your-own-audio comparisons, migration guides, SDKs, voice-agent plugins, use-case and trust pages, short-form developer demos, published volume tiers and usage-triggered onboarding.

  3. Days 61 to 904 ideas

    Commit to the big bets: on-device deals with hardware makers, a compliance track for medical and legal transcription, the annual real-world speech report and the benchmark bounty.

The research behind the plan

3 research areas, each summarised in a few lines. Open any of them for the full detail.

Customers

What customers say

Customers value accuracy that holds up in noise, distance and crosstalk. The most cited problem: speaker attribution that breaks when people talk over each other.

  • Speaker labels, timestamps and subtitles included at $0.18 an hour
  • Paying extra for diarization and add-ons
  • In their words: “who spoke when”

Market

Competitive landscape

4 competitors compared: Deepgram, AssemblyAI, ElevenLabs and pyannoteAI.

  • Gap to exploit: Accurate speaker-attributed transcripts for messy in-person and overlapping audio
  • Deepgram: 66.7% multi-speaker error in AirCaps' published comparison
  • AssemblyAI: Higher price ($0.28) and error rate (32.0%) in AirCaps' comparison

Hiring signals

AirCaps' only open role is a Product Engineer for the AirCaps smart glasses apps (Android XR, iOS, Android), not for the API.

  • Product Engineer (AirCaps smart glasses): 40,000+ hours of in-person conversations and 75M+ TikTok views from the glasses. Ideas 1, 17, 20.

From the team that wrote this playbook

All products
  • ProgrammableAgencyWorking AI systems in days, not quarters (opens programmable.agency in a new tab)
  • ProgrammableInfoAn agent that finds, researches and asks firstNew accounts get 100,000 free enrichment credits with code BETA100K (opens programmable.info in a new tab)
  • ProgrammableDigitalScreenshots in, an animated product demo out (opens programmable.digital in a new tab)
  • ProgrammableWatchKnow the day a supplier falls out of compliance (opens programmable.watch in a new tab)
  • ProgrammableTalkGet registrants for your events, and know who’s coming (opens programmable.talk in a new tab)
  • ProgrammableMediaAn AI note taker, interpreter and webinar host for your calls (opens programmable.media in a new tab)
  • ProgrammablePageText what you need, your customer gets a page that works (opens programmable.page in a new tab)
  • ProgrammableCommunityForums with labelled AI members, never empty (opens programmable.community in a new tab)
ProgrammableMarketing
ExamplesLeaderboardsProductsTermsPrivacyCookiesDPALegalContact

© 2026 AI-Powered Growth Intelligence