F5-TTS Model Inference Guide

Welcome! This guide will walk you through the steps to load and run the F5-TTS model for text-to-speech synthesis using reference audio and text inputs.

Did You Know?

Text-to-speech models like F5-TTS can mimic voice characteristics by analyzing just a few seconds of audio input. This adaptability is paving the way for personalized, AI-driven audio content.

Steps to Run the F5-TTS Model

1. Clone the Repository

Start by cloning the F5-TTS repository to your local environment:

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

2. Download the model weights

copy the download link of the model file and download using wget

wget https://huggingface.co/ModelsLab/F5-tts-brazilian/resolve/main/Brazilian_Portuguese/model_2600000.pt -P ckpts/

3. Install CUDA

Install an appropriate CUDA version compatible with your PyTorch and TorchAudio versions to enable GPU support.

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

4. Install Required Python Packages

Install the required dependencies specified in the requirements.txt file to set up your environment:

pip install -r requirements.txt

5. System Setup: APT Update, FFmpeg, and CUDA

Before running inference, ensure your system has the necessary dependencies:

Update APT Packages and Install FFmpeg

FFmpeg is essential for audio processing tasks. Update your APT packages and install ffmpeg with the following commands:

apt update
apt install -y ffmpeg

6. Run Inference with the F5-TTS Model

With the environment ready, you can now run the inference script. Adjust the paths as needed:

python inference-cli.py \
  
  # Specify the model name to use for inference
  --model "F5-TTS" \
  
  # Path to the model checkpoint file, which contains the saved model weights
  --ckpt_file "path/to/model.pt" \
  
  # Path to the reference audio file. This file is used to capture the speaking style
  # and voice characteristics, which the model will try to mimic.
  --ref_audio "wavs/sample_audio.wav" \
  
  # Reference text associated with the reference audio file.
  # This helps the model understand the speaking style.
  --ref_text "levantara a mão contra ele e o oficial então arrancara da espada e atravessara o de lado a lado estava direito ah" \
  
  # Text that the model will generate speech for. This will be spoken in the style
  # derived from the reference audio and text.
  --gen_text "O Brasil, oficialmente República Federativa do Brasil, é o maior país da América do Sul e da América Latina."