F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching

F5-TTS: Diffusion Transformer with ConvNeXt V2, faster trained and inference.

E2 TTS: Flat-UNet Transformer, closest reproduction.

Sway Sampling: Inference-time flow step sampling strategy, greatly improves performance

Installation

Clone the repository:

git clone https://github.com/SWivid/F5-TTS.git
cd F5-TTS

Install torch with your CUDA version, e.g. :

pip install torch==2.3.0+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
pip install torchaudio==2.3.0+cu118 --extra-index-url https://download.pytorch.org/whl/cu118

Install other packages:

pip install -r requirements.txt

Prepare Dataset

Example data processing scripts for Emilia and Wenetspeech4TTS, and you may tailor your own one along with a Dataset class in model/dataset.py.

# prepare custom dataset up to your need
# download corresponding dataset first, and fill in the path in scripts

# Prepare the Emilia dataset
python scripts/prepare_emilia.py

# Prepare the Wenetspeech4TTS dataset
python scripts/prepare_wenetspeech4tts.py

Training

Once your datasets are prepared, you can start the training process.

# setup accelerate config, e.g. use multi-gpu ddp, fp16
# will be to: ~/.cache/huggingface/accelerate/default_config.yaml     
accelerate config
accelerate launch train.py

An initial guidance on Finetuning #57.

Inference

To run inference with pretrained models, download the checkpoints from 🤗 Hugging Face, or automatically downloaded with inference-cli and gradio_app.

Currently support 30s for a single generation, which is the TOTAL length of prompt audio and the generated. Batch inference with chunks is supported by inference-cli and gradio_app.

• To avoid possible inference failures, make sure you have seen through the following instructions.
• A longer prompt audio allows shorter generated output. The part longer than 30s cannot be generated properly. Consider using a prompt audio <15s.
• Uppercased letters will be uttered letter by letter, so use lowercased letters for normal words.
• Add some spaces (blank: " ") or punctuations (e.g. "," ".") to explicitly introduce some pauses. If first few words skipped in code-switched generation (cuz different speed with different languages), this might help.

CLI Inference

Either you can specify everything in inference-cli.toml or override with flags. Leave --ref_text "" will have ASR model transcribe the reference audio automatically (use extra GPU memory). If encounter network error, consider use local ckpt, just set ckpt_path in inference-cli.py

python inference-cli.py \
--model "F5-TTS" \
--ref_audio "tests/ref_audio/test_en_1_ref_short.wav" \
--ref_text "Some call me nature, others call me mother nature." \
--gen_text "I don't really care what you call me. I've been a silent spectator, watching species evolve, empires rise and fall. But always remember, I am mighty and enduring. Respect me and I'll nurture you; ignore me and you shall face the consequences."

python inference-cli.py \
--model "E2-TTS" \
--ref_audio "tests/ref_audio/test_zh_1_ref_short.wav" \
--ref_text "对，这就是我，万人敬仰的太乙真人。" \
--gen_text "突然，身边一阵笑声。我看着他们，意气风发地挺直了胸膛，甩了甩那稍显肉感的双臂，轻笑道，我身上的肉，是为了掩饰我爆棚的魅力，否则，岂不吓坏了你们呢？"

Gradio App

Currently supported features:

• Chunk inference
• Podcast Generation
• Multiple Speech-Type Generation

You can launch a Gradio app (web interface) to launch a GUI for inference (will load ckpt from Huggingface, you may set ckpt_path to local file in gradio_app.py). Currently load ASR model, F5-TTS and E2 TTS all in once, thus use more GPU memory than inference-cli.

python gradio_app.py

You can specify the port/host:

python gradio_app.py --port 7860 --host 0.0.0.0

Or launch a share link:

python gradio_app.py --share

Speech Editing

To test speech editing capabilities, use the following command.

python speech_edit.py

Evaluation

Prepare Test Datasets

1. Seed-TTS test set: Download from seed-tts-eval.
2. LibriSpeech test-clean: Download from OpenSLR.
3. Unzip the downloaded datasets and place them in the data/ directory.
4. Update the path for the test-clean data in scripts/eval_infer_batch.py
5. Our filtered LibriSpeech-PC 4-10s subset is already under data/ in this repo

Batch Inference for Test Set

To run batch inference for evaluations, execute the following commands:

# batch inference for evaluations
accelerate config  # if not set before
bash scripts/eval_infer_batch.sh

Download Evaluation Model Checkpoints

1. Chinese ASR Model: Paraformer-zh
2. English ASR Model: Faster-Whisper
3. WavLM Model: Download from Google Drive.

Objective Evaluation

Some Notes

For faster-whisper with CUDA 11:

pip install --force-reinstall ctranslate2==3.24.0

(Recommended) To avoid possible ASR failures, such as abnormal repetitions in output:

pip install faster-whisper==0.10.1

Update the path with your batch-inferenced results, and carry out WER / SIM evaluations:

# Evaluation for Seed-TTS test set
python scripts/eval_seedtts_testset.py

# Evaluation for LibriSpeech-PC test-clean (cross-sentence)
python scripts/eval_librispeech_test_clean.py

Acknowledgements

• E2-TTS brilliant work, simple and effective
• Emilia, WenetSpeech4TTS valuable datasets
• lucidrains initial CFM structure with also bfs18 for discussion
• SD3 & Hugging Face diffusers DiT and MMDiT code structure
• torchdiffeq as ODE solver, Vocos as vocoder
• mrfakename huggingface space demo ~
• FunASR, faster-whisper, UniSpeech for evaluation tools
• ctc-forced-aligner for speech edit test

Citation

@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},
}

License

Our code is released under MIT License.