Introduction

AirCaps has launched A5S Async v1, a speech-to-text model that writes down every word and says who spoke it, and it just took the top spot on VoiceArena's independent Diarization-Bench. The five-person team released it on 9 November 2026 as a hosted REST API at $0.18 per audio hour, with speaker labels, word timestamps and subtitles included, and 5 free hours for new accounts.

If you cut podcasts, caption interviews, or turn recorded table reads into scripts, the hard part is rarely the words. It's the moment two people talk at once and the transcript hands a line to the wrong voice. That is exactly the problem A5S Async is built around.

What Shipped

  • Vendor / model: AirCaps, A5S Async v1 (model ID a5s-async-v1)
  • Modality: speech-to-text for recorded audio, with speaker diarization, word timestamps and SRT or VTT subtitles in one pass
  • Open vs closed: closed model, hosted API only; the benchmark harness is open source under MIT
  • Runnable today: the AirCaps console and REST API, with the first 5 audio hours free
  • Concrete facts: $0.18 per audio hour billed per millisecond; files from 0.1 seconds to 3 hours and up to 5 GB; 9 languages (English, Spanish, French, German, Portuguese, Hindi, Japanese, Korean and Chinese)

The Benchmark Result

VoiceArena scored 15 systems on 139 natural English conversations, about 22 hours of in-person and online talk with 2 to 8 speakers and heavy overlap. The metric is cpWER, a word error rate that also counts a word as wrong when it's credited to the wrong speaker.

  • AirCaps A5S Async v1: 23.88% cpWER, rank 1 of 15
  • NVIDIA Nemotron 3.5 Transcribe: 32.43%
  • ElevenLabs Scribe v2, the previous leader: 35.73%

AirCaps says A5S Async ranks first in all 10 of VoiceArena's subgroups by speaker count, setting and overlap. In its own research note it also ran 5 public meeting datasets (AMI, NOTSOFAR-1, CHiME-6, AISHELL-4 and AliMeeting, three English and two Mandarin) and reports the lowest error on every one, with a macro cpWER of 21.51% against 30.40% for Nemotron 3.5 Transcribe and 32.00% for AssemblyAI Universal-3.5 Pro.

Why It Matters for Audio Makers

Most speech models lose words when people overlap. AirCaps' note puts numbers on it: on the English sets, A5S Async deletes 8.0% of reference words, against 25.8% for AssemblyAI, 18.8% for ElevenLabs and 9.5% for Nemotron. Where half or more of the speech overlaps, the next-best model makes 61% more errors.

For creative work that means cleaner raw material:

  • Podcast and video editors get word-timed, speaker-tagged transcripts for text-based editing and captions
  • Dubbing and voice-cloning pipelines can split a recording into clean per-speaker lines before re-voicing them
  • Interview and documentary teams can search hours of footage by who said what
  • With AirCaps Speaker ID ($0.02 per hour), enrolled voices come back as names instead of speaker_0 and speaker_1

How to Run It

Sign up in the AirCaps console, grab an API key, and send a file by URL. Jobs run asynchronously, so you poll or pass a webhook.

curl https://api.aircaps.com/v1/transcripts \
  -H "Authorization: Bearer $AIRCAPS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "a5s-async-v1",
       "audio_url": "https://example.com/episode.mp3"}'

Each finished job returns utterances and words with speaker labels and start and end times. Ask the same transcript for captions with /v1/transcripts/{id}/subtitles?format=srt. You can also bias recognition toward up to 60 key terms per request in English audio, which helps with character names and product jargon, and turn on smart formatting to write numbers, money and times as digits.

Speed and Price

  • Through the API, AirCaps reports a median of 9.8 seconds to transcribe an hour-long meeting, with only Deepgram faster in its tests at 6.3 seconds
  • On its own inference engine, AirCaps measures 340 times real time on one H100, or about 11 seconds of GPU time for a one-hour meeting
  • Like-for-like pricing with speaker labels and the same add-ons, from AirCaps' 5 November comparison: AirCaps $0.18, ElevenLabs $0.27, AssemblyAI $0.28 and Deepgram $0.39 per audio hour

Limits

  • It's a closed, cloud-only model. There are no weights to run locally
  • The 5-dataset and internal results come from AirCaps itself. The VoiceArena board is independent, and the open benchmark repo lets you rerun the comparison
  • By default, AirCaps says customer audio may be used to train and benchmark its models. Paid accounts can opt out in the console, and Enterprise can get zero data retention, so check this before uploading unreleased material
  • Key-term biasing works on English audio only
  • A real-time streaming version is listed as coming soon, so this release is for recorded files, not live captions

Original Source

A5S Async v1 model page on AirCaps

Research note: A5S Async v1 evaluation and method

Independent leaderboard: VoiceArena Diarization-Bench

Benchmark code: aircaps/a5s-async-v1-benchmark on GitHub

Try it: AirCaps console

Launch thread:

Conclusion

A5S Async v1 isn't a voice generator, but it fixes the step that comes before a lot of voice work: knowing exactly who said each line in a messy, overlapping recording. At $0.18 an hour with labels and subtitles included, it's worth running your noisiest group recording through the free 5 hours and comparing the speaker turns against whatever you use today.

—Anabel ♡