f5-tts-mlx-quantized

HF: Mode weight

F5 TTS — MLX

This repo is a fork of original f5-tts-mlx implementation but a quantized flow-matching model that is only 223MB in size. The repo is meant to be used as a component of my blog post on low VRAM voice generator

Implementation of F5-TTS, with the MLX framework.

F5 TTS is a non-autoregressive, zero-shot text-to-speech system using a flow-matching mel spectrogram generator with a diffusion transformer (DiT).

This repo attempted to reduce the VRAM usage of the original model so that it can be easily deployed on any kind of Apple Device with ease (with MLX). The result as can be seen, is still very usable.

Demo:

4bit (223MB)

https://github.com/user-attachments/assets/406b4624-8f7c-48a4-a35d-2108fb081744

Original (1.35GB)

https://github.com/user-attachments/assets/c8b6f7c0-65ab-4950-ac96-b10608954174

Installation

pip install f5-tts-mlx-quantized

Basic Usage

python -m f5_tts_mlx.generate --text "The quick brown fox jumped over the lazy dog."

You can also use a pipe to generate speech from the output of another process, for instance from a language model:

mlx_lm.generate --model mlx-community/Llama-3.2-1B-Instruct-4bit --verbose false \
 --temp 0 --max-tokens 512 --prompt "Write a concise paragraph explaning wavelets." \
| python -m f5_tts_mlx.generate

Voice Matching

If you want to use your own reference audio sample, make sure it's a mono, 24kHz wav file of around 5-10 seconds:

python -m f5_tts_mlx.generate \
--text "The quick brown fox jumped over the lazy dog." \
--ref-audio /path/to/audio.wav \
--ref-text "This is the caption for the reference audio."

You can convert an audio file to the correct format with ffmpeg like this:

ffmpeg -i /path/to/audio.wav -ac 1 -ar 24000 -sample_fmt s16 -t 10 /path/to/output_audio.wav

See here for more options to customize generation.

From Python

You can load a pretrained model from Python:

from f5_tts_mlx.generate import generate

audio = generate(text = "Hello world.", ...)

Pretrained model weights are also available on Hugging Face.

Appreciation

Lucas Newman for original implementation of F5 TTS on MLX

Yushen Chen for the original Pytorch implementation of F5 TTS and pretrained model.

Phil Wang for the E2 TTS implementation that this model is based on.

Citations

@article{chen-etal-2024-f5tts,
      title={F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching}, 
      author={Yushen Chen and Zhikang Niu and Ziyang Ma and Keqi Deng and Chunhui Wang and Jian Zhao and Kai Yu and Xie Chen},
      journal={arXiv preprint arXiv:2410.06885},
      year={2024},
}

@inproceedings{Eskimez2024E2TE,
    title   = {E2 TTS: Embarrassingly Easy Fully Non-Autoregressive Zero-Shot TTS},
    author  = {Sefik Emre Eskimez and Xiaofei Wang and Manthan Thakker and Canrun Li and Chung-Hsien Tsai and Zhen Xiao and Hemin Yang and Zirun Zhu and Min Tang and Xu Tan and Yanqing Liu and Sheng Zhao and Naoyuki Kanda},
    year    = {2024},
    url     = {https://api.semanticscholar.org/CorpusID:270738197}
}

License

The code in this repository is released under the MIT license as found in the LICENSE file.