Fish Audio

Freemium

Fish Audio is a voice and text-to-speech platform. It turns text into spoken audio, clones voices from short samples, and transcribes speech back to text.

You get a web interface for generating voices and an API for building apps. There is also an open-source side called Fish Speech.

It sits in the voice and text-to-speech category. That means its job is making computers sound like people.

Open it if you need fast voice cloning or multilingual audio at low cost. Skip it if you need guaranteed polish on every output or strict control over who can clone voices.

It fits people who test before they trust, and who can work around rough edges.

• Fiction producers needing multiple character voices: You get a Story Studio built for multi-speaker narration. You still need to check each voice fits the character.
• Users who browse better than they prompt: You get 2 million community voices with filters for gender, age, and 48 quality markers. You click and listen instead of describing what you want.
• Projects where specific voice archetype matters: You find pre-made voices by type instead of building from scratch. You skip the clone-and-hope cycle.
• Users willing to extensively test before paying: You try voices on the free tier before committing credits. You learn which emotion tags break and which work.
• Those prioritizing emotional expression in long-form content: You add tags like [whispering] or [laughing] to shape delivery. You get variation, but quality shifts by voice.
• Content creators with multilingual needs: You handle mixed English and Spanish in one script. You skip the accent-mismatch problems of English-first tools.
• Developers building TTS apps: You get a free S2.1 Pro API with no hard cap. You integrate faster voice generation without per-request billing surprises.
• Teams needing consistent voice across languages: You clone once and generate in 30+ languages. You keep the same speaker profile instead of recasting per market.
• Cost-conscious users currently on ElevenLabs: You pay roughly an order of magnitude less for API calls. You absorb the trade-off of less polish on some outputs.
• Indie developers building voice agents: You prototype on the free tier with real-time latency. You scale without enterprise negotiating.

These are the parts that keep you opening it after the first try.

• 15-second voice cloning: You upload 10-15 seconds of audio and get a working clone in minutes. You skip the hunt for clean hour-long samples.
• Emotion tags that actually shift delivery: You type [whisper] or [sad] and the voice changes. You learn placement by testing, not from perfect docs.
• Browse-before-generate voice library: You click, listen, and add voices to your collection. You stop burning credits on voices you do not want.
• Multilingual support that mixes languages: You write Spanish and English in one script without switching models. You get one speaker across both.
• Free tier with commercial use: You get roughly 7 minutes monthly to test real projects. You decide if the quality fits before you spend.
• Fast, low-friction API: You get claimed ~150ms latency and streaming responses. You build responsive voice apps without buffering waits.

The path depends on your job, but the core flow is short.

1. Open the web app and pick TTS, voice cloning, or speech-to-text from the main page.
2. For cloning, upload 10-15 seconds of clean audio or record directly.
3. For generation, type your script and add emotion tags like [excited] or [long pause] where delivery should shift.
4. Browse the community library if you want a pre-made voice instead.
5. Generate, listen, and download or send to the API endpoint.

Skip it when these limits block your project.

• Inconsistent voice cloning quality: Users report robotic, stuttery results and voices changing character mid-sentence. You need three or four tries to get usable dialogue.
• Emotion control varies by voice: Tags work well on some voices and break on others. You cannot assume the same script sounds right every time.
• Published clones cannot be edited or deleted: Once you share a voice, it stays public by reported policy. You lose control over your own training data.
• Opaque credit consumption: Some users report 2000 credits for 30 seconds of audio. Your cost estimates drift until you test your exact use case.
• Support and refund issues: Multiple reports cite 17+ day response delays and denied refunds within stated windows. You need patience when things break.
• Legal risk from unauthorized clones: Anyone can upload a 10-second sample of popular characters. You risk using voices built without actor consent.