A5S Small V2
Live transcripts that hold up in noise.
Real-time English speech-to-text that runs in your VPC, on premises, air-gapped or on device. Words arrive as people speak, with optional live speaker labels, and stay accurate with distant microphones, music and crosstalk.
Made for live conversation.
Partial and final transcripts, live speakers and key terms, in your own deployment.
Partial and final
Pending words arrive as people speak and may still change. Final words are sent once and never change.
Live speaker labels
Turn on diarization and every utterance and word is labeled by speaker as the conversation unfolds.
Your deployment
Runs in your VPC, on your own servers, air-gapped, or on the device itself.
Key terms
Up to 64 names and terms per session, each with its own strength, to catch jargon and proper nouns.
Noisy, far-field audio
Built for distant microphones, background music and people talking over each other.
Fast on device
Punctuated, cased text, with final segments in 44 ms on an M4 MacBook Air and 50 ms on an iPhone 16.
Ahead where audio is hard.
word error rate across Mega-ASR, AMI, DiPCo and NOTSOFAR: the lowest of 13 streaming systems, ahead of ElevenLabs, Meta, AssemblyAI, OpenAI, xAI, Deepgram and Google.
Where it fits.
Agents that hear people clearly
Accurate words in cars, kitchens, shops and offices, so an agent answers what was actually said.
Live captions
Captions for events, calls and broadcasts as people speak, with speakers marked when you need them.
Real-time assistants
Live notes, prompts and summaries that know who is talking, while the meeting is still going.
Specifications.
More in the model reference.
- Deployment
- VPC, on-premise, air-gapped or on-device
- Audio
- 16 kHz, mono, 16-bit PCM (
pcm_s16le) - Output
- Partial and final utterances with word times; punctuated, cased text
- Speakers
- Optional live speaker labels
- Language
- English
- Key terms
- Names and terms to bias recognition toward
- Price
- Per deployment: talk to sales
Questions.
How do I get it?
A5S Small V2 Streaming is offered for VPC, on-premise, air-gapped and on-device deployments. Talk to sales about your deployment.
Is it in the AirCaps AI API?
No. It runs in your own deployment. The API's live model is A5S Streaming v1, in early access for select organizations.
How accurate is it?
26.27% word error rate across four hard, real-world corpora, the lowest of the 13 streaming systems we tested. The research note has the per-corpus results.
How fast is it on a device?
On an M4 MacBook Air, a final segment arrives 43.99 ms after the speech ends, on average; on an iPhone 16, 50.39 ms. Both run fully on device. See latency.
Which languages does it support?
English.
Do you train on my audio?
No. See Data and privacy in the docs.
More models.
Live transcripts that hold up.
Run A5S Small V2 in your own cloud, on premises or on device. Tell us about your deployment.