fish-audio-tts

Send a POST request to https://api.fish.audio/v1/tts with your FISH_AUDIO_API_KEY as a Bearer token and the text in the JSON body. Select the backend with the model HTTP header, and the API returns raw audio bytes you write to an output file.

Create the voice model in the fish.audio playground, copy its model id, and pass it as reference_id in each TTS request. Inline on-the-fly cloning from uploaded reference audio is not supported, so the voice model must exist beforehand.

Use s2.1-pro for the best quality with emotion tags and 80+ languages, s2-pro for stable high quality, or s1 for legacy compatibility. The s2.1-pro-free tier suits drafts only during the promo window and must not serve production or commercial narration.

S2 models (s2-pro, s2.1-pro, s2.1-pro-free) interpret inline tags like [laugh] and [whispers] to change delivery mid-sentence. The s1 model does not interpret them and may read them aloud, so strip tags when targeting s1.

A 401 means the FISH_AUDIO_API_KEY is missing or incorrect, while a 404 indicates the reference_id is wrong or not owned by your account. A 402 signals exhausted account credit, and empty audio usually means empty text input.

Billing is per UTF-8 byte of input text, not per character, so CJK text and emoji cost three to four times an ASCII character of the same visible length. List pricing is $15 per 1M bytes for s1, s2-pro, and s2.1-pro.