Fish Audio: A Voice AI Platform Built for Cloning and Delivery Pipelines

What Fish Audio actually is

Fish Audio is a voice AI platform. According to its documentation, every core feature is available three ways: in a no-code web app, through a REST API, and via an official SDK. That three-way access matters if you build automations, because it means you can prototype in the browser and then move the exact same capability into code without switching vendors.

The core features are the usual voice-AI set, but worth listing plainly. Text to Speech converts text into speech using the s2.1-pro, s2-pro, and s1 models. Speech to Text transcribes audio with per-segment timestamps. Voice Cloning can clone a voice instantly from a clip or train a persistent model. Realtime Streaming streams audio as it generates, which is the piece you need for voice agents and live apps. There's also a Manage Voices API to list, inspect, update, and delete your voice models.

On the model side, s2.1-pro is the recommended production model, described as improving on s2-pro for quality, latency, and throughput. There's an s2.1-pro-free tier at $0 for testing and smaller businesses, without the TTFA or DPA guarantees of the paid tier. The older s2-pro adds multi-speaker and natural-language expression control, and s1 uses parenthesis-style emotion tags.

The practical claim behind covering this now: in hands-on voice-cloning comparisons, Fish's output tends to sound the most like the actual person, and it's the version that cleared early review rounds on real projects. That's a subjective read, not a benchmark, but it's the reason s2.1-pro going public is interesting.

Why it matters for automation builders

Voice generation is easy to call. Voice *delivery* is the hard part — verifying permission, checking the reference recording, generating, running a fair comparison, and getting sign-off. Fish Audio ships an official coding-agent skill so your agent "writes correct code" against the current API, which is the detail that makes a reliable pipeline realistic rather than a one-off script.

A good pattern doing the rounds breaks the job into five project-scoped subagents that run in order. Whatever coding harness you use (Claude Code, Codex, or similar), the shape is the same:

• test-planner reads the script and permission file, pulls out names, brand terms, numbers, long sentences, pauses, and language switches, then writes a test script. If the reference recording isn't good enough, it produces a recording guide following the official guidance — one speaker, a quiet room, normal pace, roughly half a second between sentences, a couple of 15–20 second clips.
• sample-checker uses ffprobe or ffmpeg to actually read the reference audio and check format, sample rate, duration, clipping, background noise, and whether more than one person is speaking. Crucially, it never claims to have "listened" without calling a tool.
• fish-runner creates a private voice, then generates the test script with s2.1-pro, keeps the first generation, and logs the voice id, model, full request parameters, and any errors to a run file. No silent retries — failures get handed back raw.
• blind-test-maker labels generations, level-matches them with ffmpeg loudnorm, randomizes labels, and builds a scorecard that asks only three questions: is pronunciation right, do pauses sound natural, and which version sounds most like the person. The voice owner answers the last one; the agent does not guess.
• delivery-gate checks the permission file, the run log, and the completed scorecard. If permission, files, and sign-off all line up, it saves the selected audio and writes an approval note. If anything is missing, it writes a blocked note stating exactly what.

The key idea worth stealing: the permission file has to state whose voice it is, where the written permission lives, what content and channels it may be used in, and an expiry date. Anything missing counts as no permission. That's a governance gate baked into the automation, which is exactly where most voice projects get into trouble.

Where it fits versus the alternatives

Fish Audio competes with the well-known voice generators like ElevenLabs and cloud TTS from the big providers. Its differentiators, based on what the docs state, are the s2.1-pro production model, a free tier of the same model for prototyping, realtime streaming for agents, and the official agent skill that keeps generated code aligned with the current OpenAPI spec.

If you already orchestrate work in n8n, Make, or Zapier, you can call the REST API from an HTTP node and slot Fish into an existing content pipeline. But the delivery-gate pattern above is better suited to a coding agent than a low-code flow, because it needs to run ffmpeg, hold state across revisions, and refuse to proceed on missing sign-off. You could pair the two: an agent handles generation and QA, then a low-code workflow handles distribution once an approval file appears.

A fair note of caution: the sound-likeness claim is subjective, the free tier drops the latency and data-processing guarantees, and voice cloning carries real legal and consent risk — which is exactly why the permission gate isn't optional. Treat those as hard requirements, not paperwork.

If you want help wiring Fish Audio into a production voice workflow, browse the provider directory to find someone who can build and maintain it.