Fish Audio API SDKs for account-owned voice resources and speech generation in JavaScript, Python, Ruby, Go, Java, and PHP on RunAPI.

Create account-owned voice resources and generate RunAPI-managed MP3 or WAV speech. Use s2.1-pro for recommended production TTS with 83-language support and natural-language expression control. s2-pro remains available as a previous-generation option.

npm install @runapi.ai/fish-audio
pip install runapi-fish-audio
gem install runapi-fish-audio
go get github.com/runapi-ai/fish-audio-sdk/go@latest

For Java, install ai.runapi:runapi-fish-audio:0.5.0. The PHP package is released from the split repository at https://github.com/runapi-ai/fish-audio-php.

import { FishAudioClient } from '@runapi.ai/fish-audio';

const client = new FishAudioClient();
const created = await client.createVoice.run({
  name: 'Narrator',
  source_audio_url: 'https://cdn.runapi.ai/public/samples/voice.mp3',
});
const voice = await client.getVoice.run({ voice_id: created.voice.voice_id });
if (voice.voice.state !== 'trained') throw new Error(`Voice is ${voice.voice.state}`);

const result = await client.textToSpeech.run({
  model: 's2.1-pro',
  text: 'Hello from RunAPI [excited]',
  voice_id: created.voice.voice_id,
  output_format: 'wav',
  sample_rate_hz: 44100,
});
console.log(result.audios[0].url);

Use client.listVoices.run({ page_number: 1, page_size: 10 }) to list voice resources created by the current account. Voice state is created, training, trained, or failed; only trained voices can be submitted for speech generation. A returned voice_id is a best-effort reference and may stop working later.

Alternatively, pass request-scoped references, where each item contains a publicly accessible HTTPS audio URL and its exact transcript. Use references when the request must not depend on a previously returned voice_id.

output_format defaults to mp3. MP3 supports 32000 or 44100 Hz and optional bitrate_kbps values of 64, 128, or 192. WAV supports 8000, 16000, 24000, 32000, or 44100 Hz and does not accept bitrate_kbps.

This version does not provide update, delete, revoke, or voice-library management methods. It does not make promises about voice retention or how source audio is used for training.

• s1
• s2-pro
• s2.1-pro

The returned URL points to RunAPI-managed storage. See the Fish Audio model page for current pricing, limits, and usage details.

Licensed under the Apache License, Version 2.0.

For s1, s2-pro, and s2.1-pro, supply at most one references item containing a publicly accessible HTTPS audio URL in audio and a nonempty, verbatim transcript in text. Do not send voice_id with a reference. Omit references when selecting an existing voice. The transcript is forwarded unchanged; callers must ensure it matches the audio.

For local audio or Base64, first upload through POST /api/v1/files, then use the returned url as references[0].audio. Keep only one sample and remove voice_id. The URL must remain valid while the request waits and the audio is fetched. Files URLs expire by default after one hour; storing a reference does not extend that lifetime. For longer waits, use an external HTTPS URL with enough validity. If a URL expires, upload again and submit a new request.