Fish Audio Python SDK for RunAPI

Install runapi-fish-audio and create a FishAudioClient.

created = client.create_voice.run(name="Narrator", source_audio_url="https://cdn.runapi.ai/public/samples/voice.mp3")
voice = client.get_voice.run(voice_id=created.voice.voice_id)
if voice.voice.state != "trained":
    raise RuntimeError(f"Voice is {voice.voice.state}")
result = client.text_to_speech.run(model="s2.1-pro", text="Hello [excited]", voice_id=created.voice.voice_id, output_format="wav", sample_rate_hz=44100)

Pass optional references entries with a publicly accessible HTTPS audio URL and exact transcripts for request-scoped voice matching.

create_voice.run() and text_to_speech.run() return the same typed terminal result whether the request completes directly or is accepted for background processing. To resume an accepted request from its opaque Task Result URL, call subscribe(location) on the same resource.

Use list_voices and get_voice to inspect account-owned voice resources. Only trained voices can be submitted for speech generation; a returned voice_id is a best-effort reference and may stop working later. Update, delete, revoke, or voice-library management methods are not provided; voice retention is not promised.

The output defaults to MP3. Select WAV with output_format; bitrate_kbps applies only to MP3.

Model details and pricing: https://runapi.ai/models/fish-audio/s2.1-pro

Licensed under the Apache License, Version 2.0.

Temporary voice matching

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.