Fish Audio STT

FishAudioSTT targets the Fish Audio ASR endpoint (POST https://api.fish.audio/v1/asr). It buffers incoming PCM, uploads it as a WAV once the window fills, and emits the returned text as a final transcript.

Fish transcription is batch-only — there is no streaming socket. This adapter therefore emits one final transcript per ~2 s window and no interim partials, exactly like WhisperSTT. If turn latency is what you are optimising, use a genuinely streaming provider: Deepgram, Soniox, AssemblyAI or Speechmatics.Reach for Fish ASR when you want a single vendor (and a single key) for both directions of the call, or for its language coverage.

Install

pip install "getpatter[fish_audio]"

npm install getpatter

Authentication

export FISH_AUDIO_API_KEY="<your-fish-audio-api-key>"

The same key covers ASR and Fish Audio TTS.

Usage

# Namespaced import (pipeline mode)
from getpatter.stt import fish_audio

stt = fish_audio.STT()                                   # reads FISH_AUDIO_API_KEY
stt = fish_audio.STT(api_key="...", language="it")

# Flat alias (equivalent)
from getpatter import FishAudioSTT

stt = FishAudioSTT()

// Namespaced import (pipeline mode)
import * as fishAudio from "getpatter/stt/fish-audio";

const stt = new fishAudio.STT();                         // reads FISH_AUDIO_API_KEY
const stt2 = new fishAudio.STT({ apiKey: "...", language: "it" });

// Flat alias (equivalent)
import { FishAudioSTT } from "getpatter";

const stt3 = new FishAudioSTT();

Plug it into an agent:

import asyncio
from getpatter import Patter, Twilio, FishAudioSTT, FishAudioTTS

phone = Patter(carrier=Twilio(), phone_number="+15550001234")

agent = phone.agent(
    stt=FishAudioSTT(language="it"),                      # FISH_AUDIO_API_KEY from env
    tts=FishAudioTTS.for_twilio(),                        # same key
    system_prompt="Sei l'assistente di Acme. Rispondi in italiano.",
)

asyncio.run(phone.serve(agent))

// npx tsx example.ts
import { Patter, Twilio, FishAudioSTT, FishAudioTTS } from "getpatter";

const phone = new Patter({ carrier: new Twilio(), phoneNumber: "+15550001234" });

const agent = phone.agent({
  stt: new FishAudioSTT({ language: "it" }),              // FISH_AUDIO_API_KEY from env
  tts: FishAudioTTS.forTwilio({}),                        // same key
  systemPrompt: "Sei l'assistente di Acme. Rispondi in italiano.",
});

await phone.serve({ agent });

Language detection

Leave language unset (None / undefined) and Fish auto-detects. Pass an ISO code to pin it — pinning is usually more accurate and slightly faster on short windows.

stt = FishAudioSTT(language=None)                         # auto-detect
stt = FishAudioSTT(language="it")                         # pinned

const auto = new FishAudioSTT({ language: undefined });   // auto-detect
const pinned = new FishAudioSTT({ language: "it" });      // pinned

Timestamps

ignore_timestamps defaults to True, which Fish documents as the lower-latency path for clips under 30 s. Set it to False to receive per-segment timings.

stt = FishAudioSTT(ignore_timestamps=False)

# Segments arrive on Transcript.words — Fish returns {text, start, end} per segment.
# transcript.words[0] -> {"text": "ciao", "start": 0.0, "end": 0.4}

const stt = new FishAudioSTT({ ignoreTimestamps: false });

// Segments arrive on Transcript.segments — {text, start, end} per segment.
// transcript.segments?.[0] -> { text: "ciao", start: 0, end: 0.4 }

Python surfaces segments on Transcript.words (the shared frozen dataclass’s provider-specific field); TypeScript exposes a dedicated segments array. Same data, same order — the field name differs because the Python Transcript shape is shared across every STT provider.

Buffering and the 1-second floor

Fish rejects clips shorter than 1 second and longer than 60 minutes / 20 MB. The adapter handles both ends:

The steady-state window is ~2 s (64,000 bytes of 16 kHz PCM16), comfortably clear of the floor and half the request count of a 1 s window.
On close() a short tail is padded with digital silence up to the 1 s minimum rather than dropped — otherwise the last words of an utterance would vanish, which is the exact bug that was fixed on the Whisper adapter.
An oversized buffer is truncated to the most recent 20 MB with a warning.

Tune the window with buffer_size_bytes / bufferSize if you want lower latency at the cost of more requests:

stt = FishAudioSTT(buffer_size_bytes=32000)               # ~1 s windows

const stt = new FishAudioSTT({ bufferSize: 32000 });      // ~1 s windows

Failure behaviour

A non-2xx response or an unreachable host is logged at ERROR and yields no transcript — it never raises into the call. A live call degrades to silence on that window rather than dropping.

Options

Python	TypeScript	Default	Notes
`api_key`	`apiKey`	—	Reads `FISH_AUDIO_API_KEY` when omitted.
`language`	`language`	`"en"`	ISO code. Pass `None` / `undefined` to auto-detect.
`ignore_timestamps`	`ignoreTimestamps`	`True`	`False` returns per-segment `start` / `end`.
`buffer_size_bytes`	`bufferSize`	`64000`	Bytes of 16 kHz PCM16 per upload (~2 s).
`base_url`	`baseUrl`	Fish `/v1/asr`	Override for proxying or tests.

Pricing

Fish bills ASR at **

0.36 per audio hour**, rounded up to the nearest second — Patter records `

0.006 / minute`. See Fish Audio pricing and rate limits for the authoritative numbers and the concurrency tiers.

Beta. This provider is validated against the Fish Audio API specification but has not yet been exercised on a live phone call end to end.

Get Started

Setting up Patter

Observability

Integrations

Development

Fish Audio STT

Fish Audio STT

Install

Authentication

Usage

Language detection

Timestamps

Buffering and the 1-second floor

Failure behaviour

Options

Pricing

​Fish Audio STT

​Install

​Authentication

​Usage

​Language detection

​Timestamps

​Buffering and the 1-second floor

​Failure behaviour

​Options

​Pricing

Fish Audio STT

Install

Authentication

Usage

Language detection

Timestamps

Buffering and the 1-second floor

Failure behaviour

Options

Pricing