---
tags:
- f5-tts
- text-to-speech
- voice-cloning
- flow-matching
- zero-shot-tts
license: cc-by-nc-4.0
datasets:
- mythicinfinity/libritts_r
- amphion/Emilia-Dataset
base_model: SWivid/F5-TTS
pipeline_tag: text-to-speech
language:
- en
- zh
---

# 🎙️ Voice Clone Model (F5-TTS Based)

A production-ready **zero-shot voice cloning** model based on the state-of-the-art **F5-TTS** architecture (Flow Matching + Diffusion Transformer).

## 📦 Files in This Repo

| File | Description |
|------|-------------|
| `README.md` | This documentation |
| `config.json` | Model configuration and hyperparameters |
| `train_voice_clone.py` | **Fine-tuning script** — adapt to your own voice data |
| `inference_voice_clone.py` | **Local inference script** — zero-shot voice cloning CLI |
| `voice_clone_f5tts.ipynb` | **📓 Jupyter Notebook** — ready for Colab / Kaggle |

## 🚀 Quick Start Options

### Option 1: Hugging Face Space (No setup)
Try instantly at [rajkr-voice-clone-f5tts-demo.hf.space](https://rajkr-voice-clone-f5tts-demo.hf.space)

### Option 2: Google Colab / Kaggle (Free GPU)
Open the notebook directly:
- **Colab**: [Open in Colab](https://colab.research.google.com/github/rajkr/voice-clone-f5tts/blob/main/voice_clone_f5tts.ipynb) *(upload `voice_clone_f5tts.ipynb` to your Drive first)*
- **Kaggle**: Download `voice_clone_f5tts.ipynb` from this repo → Upload to Kaggle → Enable GPU T4

Or follow the quick steps below:

```python
# 1. Enable GPU: Runtime → Change runtime type → GPU
# 2. Install
!pip install -q f5-tts soundfile

# 3. Download model (~1.3GB)
from huggingface_hub import snapshot_download
snapshot_download("SWivid/F5-TTS", local_dir="./f5tts_model", allow_patterns=["F5TTS_v1_Base/*"])

# 4. Clone a voice
from f5_tts.api import F5TTS
tts = F5TTS(ckpt_file="./f5tts_model/F5TTS_v1_Base/model_1250000.safetensors",
            vocab_file="./f5tts_model/F5TTS_v1_Base/vocab.txt")

wav, sr, _ = tts.infer(
    ref_file="/content/my_voice.wav",     # Upload your audio first
    ref_text="Exact transcript of your audio.",
    gen_text="Say this in the cloned voice!",
    nfe_step=32,
)

import soundfile as sf
sf.write("output.wav", wav, sr)
```

### Option 3: Local Machine (GPU recommended)

```bash
pip install f5-tts soundfile

python -c "
from f5_tts.api import F5TTS
import soundfile as sf

tts = F5TTS()  # Auto-downloads model on first run
wav, sr, _ = tts.infer(
    ref_file='my_voice.wav',
    ref_text='Hello, this is my voice.',
    gen_text='Hello from my local machine!',
)
sf.write('output.wav', wav, sr)
"
```

## Model Description

This repo provides a complete voice cloning pipeline using **F5-TTS v1 Base** (335M parameters), the current best open-source neural TTS model. Clone any voice from just **3-10 seconds** of reference audio.

### Architecture

| Component | Details |
|-----------|---------|
| **Type** | Conditional Flow Matching (CFM) with Diffusion Transformer (DiT) |
| **Params** | 335M |
| **Backbone** | DiT (dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4) |
| **Vocoder** | Vocos (24kHz, 100 mel channels) |
| **Training** | Trained on 95K hours of multilingual speech (Emilia EN+ZH) |
| **Inference** | Zero-shot voice cloning with 3-10s reference audio |
| **RTF** | ~0.15 (6.7x real-time capable) |

## Fine-Tuning Your Own Voice

### Option A: Python Script

```bash
# 1. Prepare your data:
# my_voice/
#   ├── metadata.csv      # format: audio_path|text
#   └── wavs/
#       ├── clip001.wav
#       └── clip002.wav

# 2. Run training
python train_voice_clone.py \
    --hf_dataset mythicinfinity/libritts_r \
    --hf_config clean \
    --hf_split train.clean.100 \
    --epochs 20 \
    --lr 1e-5
```

### Option B: CLI Fine-Tuning (Official F5-TTS)

```bash
pip install f5-tts

# Prepare dataset
python -m f5_tts.train.datasets.prepare_csv_wavs \
    /path/to/my_voice \
    /path/to/prepared_data/MyVoice_custom

# Fine-tune
python -m f5_tts.train.finetune_cli \
    --exp_name F5TTS_v1_Base \
    --dataset_name MyVoice \
    --tokenizer custom \
    --finetune \
    --learning_rate 1e-5 \
    --batch_size_per_gpu 38400 \
    --batch_size_type frame \
    --max_samples 64 \
    --epochs 20 \
    --num_warmup_updates 300 \
    --grad_accumulation_steps 2 \
    --logger tensorboard
```

## Performance

| Metric | Value |
|--------|-------|
| **WER** (test-clean) | ~1.87% |
| **Speaker Similarity** | SIM-o ~0.66 |
| **Real-Time Factor** | 0.15 (6.7x faster than real-time) |
| **Minimum Reference** | 3 seconds |
| **Languages** | English + Chinese (pretrained), adaptable to others |

## References

- [F5-TTS Paper](https://arxiv.org/abs/2410.06885) — *F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching*
- [Official Repo](https://github.com/SWivid/F5-TTS)
- [Original Model](https://huggingface.co/SWivid/F5-TTS)

## Citation

```bibtex
@article{shen2024f5tts,
  title={F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching},
  author={Shen, Yusheng and Wang, Zhijian and Dalmia, Shaylen and Su, Yuchuan and Liu, Zhejian and Marino, Kevin and Zonooz, Bahram and Yao, Zirun and Ma, Xinyin},
  journal={arXiv preprint arXiv:2410.06885},
  year={2024}
}
```

## License

CC-BY-NC-4.0 (non-commercial use)