F5-TTS — Flow Matching Text-to-Speech

What it is

F5-TTS is a diffusion transformer-based text-to-speech system using flow matching with ConvNeXt V2 architecture. It delivers multi-speaker and multi-style speech synthesis, voice chat powered by Qwen2.5-3B-Instruct, a Gradio web interface for inference and fine-tuning, and CLI inference.

With Triton and TensorRT-LLM optimization, F5-TTS achieves 0.0394 real-time factor on L20 GPU, making it one of the fastest open-source TTS systems available. MIT licensed code with CC-BY-NC pre-trained models.

How it saves time or tokens

F5-TTS provides a complete TTS pipeline in a single package: reference audio input, text input, and high-quality speech output. Traditional TTS setups require assembling multiple components (text normalization, acoustic model, vocoder). F5-TTS handles all stages internally. The Gradio UI enables quick prototyping without writing code. The CLI interface integrates into automation pipelines. Voice cloning from a reference audio sample eliminates the need for training custom voice models.

How to use

1. Install F5-TTS:

pip install f5-tts

1. CLI inference with a reference audio:

f5-tts_infer-cli \
  --model F5TTS_v1_Base \
  --ref_audio ref.wav \
  --ref_text 'Reference text matching the audio' \
  --gen_text 'Text you want to generate as speech'

1. Launch the Gradio web UI:

f5-tts_infer-gradio
# For voice chat with Qwen2.5
f5-tts_infer-gradio --voicechat

Example

Using F5-TTS in a Python script for batch generation:

from f5_tts.api import F5TTS

tts = F5TTS(model_type='F5-TTS', ckpt_file='F5TTS_v1_Base')

# Generate speech from reference audio
wav, sr, _ = tts.infer(
    ref_file='speaker_reference.wav',
    ref_text='This is the reference transcript.',
    gen_text='Generate this text in the same voice.',
    seed=42
)

# Save output
import soundfile as sf
sf.write('output.wav', wav, sr)

# Batch generation with multiple texts
texts = [
    'Welcome to the product demo.',
    'Here are the key features.',
    'Thank you for watching.'
]
for i, text in enumerate(texts):
    wav, sr, _ = tts.infer(
        ref_file='speaker_reference.wav',
        ref_text='Reference text.',
        gen_text=text
    )
    sf.write(f'segment_{i}.wav', wav, sr)

Related on TokRepo

• AI tools for voice — More text-to-speech and voice tools on TokRepo.
• Featured workflows — Discover curated AI tools.

Common pitfalls

• Reference audio quality directly affects output quality. Use clean, noise-free recordings of at least 5 seconds for the best voice cloning results.
• The CC-BY-NC license on pre-trained models restricts commercial use. Train your own models for commercial applications.
• Running F5-TTS without GPU acceleration is very slow. A CUDA-capable GPU is recommended for practical use.