New A5S Async v1: 29% fewer errors on real meetings

Voice AI that hears what other models miss.

Speech-to-text, speaker diarization and speaker ID that stay accurate through noise, distance, crosstalk and many speakers. One API, with an open benchmark behind every result.

Start free Talk to sales

Every account includes 5 free hours

9.8 sto transcribe an hour-long meeting through the API 29%fewer errors on 5 public meeting benchmarks $0.18per hour for transcripts with speaker labels Openbenchmarks: code, raw outputs and scores

Hear the difference.

Real meetings from public benchmarks. Press play, choose another provider, and follow both transcripts word by word against the human reference.

Loading the listening demo…

Benchmarked head to head against

AssemblyAI
Deepgram
ElevenLabs
Google
Meta
Microsoft
NVIDIA
OpenAI
xAI

AI learned to talk. We're teaching it to listen.

Voice agents, notetakers and captions all depend on one step: hearing people correctly. Real speech makes that hard. People interrupt and talk at once, speak through noise and far from the microphone.

Most speech models are graded on clean audio, so their mistakes surface in production, and every step after the transcript inherits them. AirCaps models are built and tested on the hard audio.

Clean audio hides the gap

Every provider looks alike on single-speaker clips. On real meetings with speaker labels, they spread apart.

Clean speech: VoxPopuli, LibriSpeech, Earnings-22Real meetings: 5 public benchmarks

Clean: mean WER on 3 Open ASR Leaderboard sets. Meetings: macro cpWER on NOTSOFAR-1, AMI, CHiME-6, AISHELL-4 and AliMeeting, where a word given to the wrong speaker counts as an error. Method

What other models miss

When people talk over each other, most models drop the words. A5S Async keeps them.

7.9%of words dropped by A5S Async

9.5–39.8%dropped by the 8 other providers

Each bar is 2% of the words people saidLeft out of the transcript

One API. Four models.

Every AirCaps model runs through the AirCaps API and is benchmarked against the leading providers.

Async speech-to-text A5S Async Transcripts with speaker labels for meetings, calls, interviews and media. 29% fewererrors than the next best on 5 meeting benchmarks $0.18/h Streaming speech-to-text A5S Streaming Real-time speech-to-text for live captions and voice agents. 26.27% WERon noisy, far-field audio, ahead of 8 cloud systems $0.30/h Speaker diarization Diarization Who spoke when, across meetings, calls and broadcasts. 12.03% DERlowest of 11 systems across 7 datasets $0.08/h Speaker identification Speaker ID Recognize enrolled voices through noise and music. 3.11% EERwith noise louder than speech $0.02/h

Measured in the open.

Independent leaderboards and public datasets, with the code, raw API responses and scores published for every result.

5 of 5

Public meeting benchmarks where A5S Async v1 has the lowest error: NOTSOFAR-1, AMI, CHiME-6, AISHELL-4 and AliMeeting.

29%

Fewer errors than the next-best model across 5 public meeting benchmarks, in English and Mandarin.

Rerun it

Dataset recipes, provider adapters, raw API responses and scores are on GitHub and Hugging Face.

Benchmarks on GitHub

Built for products that have to listen.

Voice agents

Hear the caller, not the background.

Accurate transcripts through kitchen noise, street noise and crosstalk, so the LLM acts on what the caller actually said.

Voice-agent platforms

One API your customers can switch on.

Streaming transcription, speaker labels and speaker ID behind a single integration.

Meeting notetakers

Every speaker, every aside.

Transcripts with speaker labels that keep the interruptions, asides and quick replies other models drop.

Media and captioning

Captions that keep up.

Captions that hold up when people talk over each other, credited to the right speaker.

Evaluation and benchmarking

Results you can rerun.

Open harnesses, raw outputs and provider adapters for every result we publish.

Something else

Building with voice? Talk to us.

Talk to sales

Human-level voice AI

Voice AI that listens as well as people do.

People follow a conversation through noise, distance and crosstalk without effort. That is the bar for every AirCaps model, and we measure against it in public.

Hears what others miss

Accurate through noise, distant microphones, crosstalk and many speakers.

Knows who's speaking

Speaker labels, diarization and speaker ID from one family of models.

Fast and efficient

Models and inference engine built together: an hour-long meeting comes back in about 10 seconds.

Measured in the open

Code, data recipes, raw outputs and scores for every result we publish.

One API. A few lines of code.

Batch and streaming transcription, diarization and speaker ID over REST and WebSocket.

Speaker labels and word timestamps on every transcript
Live transcripts over WebSocket, with optional live speaker labels
Subtitles, webhooks and names from enrolled voices
No training on your data

Get an API key Quickstart

import os, time, requests

API = "https://api.aircaps.com/v1"
H = {"Authorization": f"Bearer {os.environ['AIRCAPS_API_KEY']}"}

job = requests.post(f"{API}/transcripts", headers=H, json={
    "model": "a5s-async-v1",
    "audio_url": "https://example.com/meeting.mp3",
}).json()

while job["status"] in ("queued", "processing"):
    time.sleep(3)
    job = requests.get(f"{API}/transcripts/{job['id']}", headers=H).json()

for u in job["utterances"]:
    print(f"[{u['start']:.1f}s] {u['speaker']}: {u['text']}")

import asyncio, json, os, websockets

async def main():
    headers = {"Authorization": f"Bearer {os.environ['AIRCAPS_API_KEY']}"}
    async with websockets.connect("wss://api.aircaps.com/v1/realtime",
                                  additional_headers=headers) as ws:
        await ws.send(json.dumps({
            "type": "start", "providers": ["a5sv2"],
            "encoding": "pcm_s16le", "sample_rate": 16000, "channels": 1,
            "language": "en", "diarization": True,
        }))
        # send 20 ms PCM frames, then {"type": "finish"}
        async for message in ws:
            event = json.loads(message)
            if event["type"] == "transcript" and event["is_final"]:
                print(event["text"])

asyncio.run(main())

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/meeting.mp3",
       "webhook_url": "https://your.app/hooks/aircaps"}'

# → {"id": "tr_01k6z...", "status": "queued", ...}
# the webhook fires when the transcript is ready