Fish Audio

New

Fish Audio generates expressive speech from text, clones voices from short clips and transcribes audio, with a 2M+ community voice library, inline emotion cues and a separate pay-as-you-go API for developers.

fish.audio

Fish Audio

Explore features, practical uses and pricing below.

Fish Audio is a voice platform built on its own speech models. In the browser it works as a text-to-speech studio with a very large community voice library and a cloning tool, and developers get the same models through a REST and WebSocket API. The company also lists speech to text, a voice changer and audiobook tools, but cloning and expressive speech are the main draw.

Cloning a voice or picking one from the library

The homepage says a clone needs about 15 seconds of audio (its FAQ says as little as 10), and the voice cloning docs add that a minute or two of clean, single-speaker speech improves fidelity. Developers can clone instantly by sending reference audio and its transcript with each request, or train a persistent voice model once and reuse its ID. New models are private by default and can be made unlisted or published to the Voice Library. A Professional Voice Clone, available only in the web app, is described as a verified, studio-quality clone its owner can release and earn from.

The Voice Library is the other route. Fish Audio says it holds more than 2 million voices, mostly uploaded by users, so quality and naming vary from one voice to the next. Paid plans add Voice Design, which builds a voice from a description instead of a recording.

How Fish Audio's models handle emotion and language

The recommended text-to-speech model is S2.1 Pro, which the models overview says supports 83 languages with automatic detection. The S2 generation takes natural-language direction for expression, while the older S1 model uses emotion tags in parentheses; S1 is due to be retired on December 31, 2026. The homepage shows bracketed cues such as [whispering], [excited], [laughing] and [pause] written straight into a script.

Transcription runs on separate models. Transcribe-1-pro handles recordings up to 60 minutes, including multi-speaker conversations, and a preview model transcribes live audio over WebSocket. For text streamed from a language model, the TTS WebSocket buffers incoming tokens and starts speaking once it has enough context.

Where it fits and where it doesn't

Fish Audio suits creators making character voices, narration and game or animation dialogue, and developers who want an expressive TTS API with cloning built in. The open-weight Fish Speech models on GitHub also interest researchers. It's a weaker fit for teams that need every voice professionally licensed, because much of the library is community uploaded and a voice's origin should be checked before commercial use. More options are in our voice cloning tools category.

Producing a short audio drama scene

  1. Upload a clean sample for each character you have permission to clone, or pick library voices and note their IDs.
  2. Write the scene with inline cues, for example a whispered line followed by a [pause].
  3. Generate each line, listen for mispronounced names, and regenerate only the lines that miss.
  4. Export the takes and assemble them with music and timing in an editor such as Descript.

Developers can run the same steps through the Python SDK, passing each voice's ID as reference_id.

Plans, API billing and open weights

The pricing page lists a Free tier with 8,000 monthly credits and 500 characters per generation, Plus at $11 a month billed annually or $15 month to month, and Pro at $75 a month billed annually or $100 month to month, with three team seats. Unused minutes don't roll over. The API is billed separately: the developer page lists S2.1 Pro at $15 per million UTF-8 bytes and Transcribe-1 at $0.36 an hour, plus a free S2.1 Pro model for testing that has no SLA and may use requests to improve the model. The Fish Speech repository publishes code and weights under Fish Audio's own community and research licenses, and commercial self-hosting is an enterprise arrangement.

Questions people ask about Fish Audio

Can I use the free plan for YouTube or client work?

No. The pricing FAQ says free-plan output is for personal, non-commercial projects, and commercial use needs a paid plan.

Is API usage included in Plus or Pro?

No. The API has its own pay-as-you-go credit, billed separately from the app subscriptions.

Similar tools

MiniMax Audio is MiniMax's browser studio and API for its Speech models, offering text to speech in 40 languages with sound tags, voice cloning from 10-second samples, voice design from text and long-form async synthesis.
Cartesia offers Sonic text-to-speech and Ink speech-to-text models through a low-latency API, with instant and pro voice cloning, 44-language speech output and Managed Agents for building phone voice agents.
Dub recorded audio and video, refine translated dialogue and voices, or build localized speech workflows with CAMB.AI Studio and APIs.