Fish Speech — Multilingual TTS for 80+ Languages

What it is

Fish Speech is an open-source text-to-speech system that supports over 80 languages. It uses a 4B parameter dual-AR (autoregressive) model architecture for high-quality speech synthesis. Key features include voice cloning from short audio samples, emotional and stylistic control through 15K+ expression tags, and real-time inference speeds.

It targets developers building multilingual voice applications, content creators needing voiceovers in multiple languages, and researchers working on speech synthesis.

How it saves time or tokens

Fish Speech consolidates multilingual TTS into a single model. Instead of using different TTS models for different languages, you use one model that handles 80+ languages. Voice cloning from a short sample eliminates the need for extensive voice recording sessions. The emotional control tags let you adjust tone and expression without re-recording.

How to use

1. Install:

pip install fish-speech

1. Generate speech:

fish-speech tts 'Hello, this is Fish Speech!' --output hello.wav

1. Voice cloning:

from fish_speech import FishSpeech

model = FishSpeech()
audio = model.generate(
    text='This is a cloned voice speaking.',
    reference_audio='reference.wav',
)
audio.save('cloned.wav')

1. Run with Docker:

docker pull fishaudio/fish-speech
docker run -p 7860:7860 fishaudio/fish-speech

Example

Related on TokRepo

• AI tools for voice -- voice AI and TTS tools
• AI tools for content -- content creation tools

Common pitfalls

• The 4B model requires a GPU with at least 8GB VRAM. Smaller model variants are available for lower-end hardware but with reduced quality.
• Voice cloning quality depends on the reference audio. Clean recordings of 10-30 seconds produce the best results. Background noise and multiple speakers in the reference degrade cloning quality.
• Expression tags require familiarity with the tag vocabulary. Refer to the documentation for the full list of supported emotional and stylistic tags.