Fish Audio Realtime STT: what it is and how to try it
2026-10-28 · 4 min read
Fish Audio Realtime STT is a streaming speech-to-text model built for voice agents. Besides turning live speech into text, it decides when the caller has finished talking. It's in preview now as transcribe-stream-1 on Fish Audio's API, and it costs $0.36 per session hour, billed by the second.
The launch post
@FishAudio posted "Introducing Fish Audio Realtime STT (preview)" on October 27, along with a 72-second demo video. The company says the model follows the flow of a conversation instead of only the words: it judges when to answer, when to keep listening and when to stop. It also claims first place out of 20 models on Cekura's speech-to-text benchmark, with a 2.21% overall word error rate.
That benchmark number needs some context. AlphaSignal's write-up puts Zoom Scribe Pro v2.1 second at 2.35%, a gap of 0.14 points that a settings change could close. The test set is 1,000 short public clips plus eight licensed phone recordings, so I'd read the result as a strong showing on a small test and leave it there.
Why turn detection matters more than accuracy
If you've called a business that uses an AI phone agent, you know the two ways they fail. Some cut you off when you stop to think. Others leave a long dead pause after you finish because they're waiting out a silence timer. In both cases the words usually get transcribed fine and the timing is off, and timing is what Fish is going after. The realtime docs describe a default "auto" mode where the server judges when the user has finished. A short pause in the middle of a sentence usually doesn't end the turn. If someone trails off and never picks back up, the turn closes about 2 seconds later. The model also recognizes backchannels, the "mm-hm" and "right" people say while the agent is still talking, so the agent can keep going instead of stopping every time.
What it does and what it costs
| Item | What Fish's docs say |
|---|---|
| Model | transcribe-stream-1 (preview) |
| Connection | WebSocket at wss://api.fish.audio/v1/asr/realtime, from your server |
| Audio in | 16-bit mono PCM, 8 kHz to 48 kHz; phone audio must be decoded first |
| Turn modes | auto (smart), vad (silence timer), manual |
| Speed | first draft text in about 0.5 s; final text about 1.1 s after the speaker stops |
| Limits | 60-minute sessions, 60-second turns |
| Extras | up to 100 custom vocabulary terms, speaker tags, 29 language codes |
| Price | $0.36 per session hour, billed per second |
On price it lands in the middle of the pack. AlphaSignal lists rough hourly rates of $0.12 for Soniox, $0.45 for AssemblyAI and $1.02 for OpenAI's realtime transcription. We wrote about Grok Voice Transcribe 2.0 at 10 cents an hour back in September. Keep in mind that Fish bills from the moment the connection opens, so a socket left idle still runs the meter.
How to try it
- Create a Fish Audio account and an API key. Fish's pricing page shows that new accounts start at 5 concurrent requests, which is enough for testing.
- Get preview access. The docs change that added the realtime pages says the endpoint is enabled per team and points you to support@fish.audio. Until it's on, connecting returns a 403.
- Open a WebSocket from your server with your key in a Bearer header. Send a
configmessage with your audio format and"turn_detection": {"mode": "auto"}, then wait forsession_started. - Stream audio in 80 to 100 ms chunks. Have your agent reply on
user_turn_takenorcommitted, never on adraft. - Add your own vocabulary (product names, street names, staff names) through
context.terms, and feed the agent's own lines back withcontext_addso the model knows what was said to the caller.
There's no SDK yet, so this is developer work, which is probably fine. Most small businesses won't touch a WebSocket anyway. They'll get this through a voice-agent platform that swaps it in behind the scenes.
The bigger picture for voice agents
Fish is better known for speech output. Its Drama 3 voice model came out in mid-October, and it's priced at a $10 promo rate per million bytes of text until November 12. With realtime STT added, Fish now covers both ends of a phone call, a week after ElevenLabs went after the same buyers with Eleven Realtime.
My read: sub-second turn-taking at 36 cents an hour gets cheap AI phone answering much closer to sounding natural. For a local business, the harder call is which calls an agent should take at all. If you're weighing that against a human service, our breakdown of answering service cost shows what the people-based options charge. And if you want help working out where a voice agent fits in your intake and follow-up, New Face Design's free process audit maps it out before you spend anything.