Skip to content

F5-TTS is an open-source voice cloning tool from Shanghai Jiao Tong University with excellent results. The initial version only supported Chinese and English cloning, but the latest version v1 has expanded to support multiple languages such as French, Italian, Hindi, Japanese, Russian, Spanish, and Finnish.

This article mainly introduces how to install and launch F5-TTS using the official source code, and how to integrate it with the pyVideotrans project. In addition, it will also introduce how to modify the source code to achieve calls within the local area network.

Due to limited energy, I will no longer maintain the previous personal integration package and API interface, but will instead use the official interface to integrate with the pyVideotrans project. The limitation of the official interface is that it can only be called locally and cannot be called within the local area network. Please refer to the local area network usage section of this article for the solution.

Prerequisites

Your system must have Python version 3.10 installed. Although versions 3.11/3.12 may theoretically work, they have not been tested, so it is recommended to use version 3.10.

If Python is not yet installed:

Check if Python is installed:

  • Windows System: Press the Win+R keys, enter cmd in the pop-up window and press Enter. Enter python --version in the opened black window. If 3.10.xx is displayed, it means it has been installed; if it prompts "python is not an internal or external command", it means it has not been installed or Python has not been added to the Path environment variable, and needs to be reinstalled.
  • Mac System: Directly execute python3 --version in the terminal. If 3.10.x is output, it means it has been installed; otherwise, it needs to be installed.

Download F5-TTS Source Code

First, create an empty folder in a suitable location. It is recommended to choose a non-system disk, a location that does not require special permissions, such as drive D, etc. Avoid placing it in directories like C:/Program Files (it is recommended that the location and folders at all levels use names composed of pure numbers or letters) to avoid potential problems. For example, D:/f5/v1 is a good location, while D:/开源 f5/f5 v1 with spaces and Chinese is not recommended.

This article takes installing F5-TTS in the D:/python/f5ttsnew folder of the Windows10 system as an example.

Open the website: https://github.com/SWivid/F5-TTS

As shown in the figure below, click to download the source code:

Download source code zip package

After downloading, unzip the package and copy all the files in the F5-TTS-main folder to the D:/python/f5ttsnew folder, as shown below:

F5-TTS-main folder in the compressed package

Copy to f5ttsnew

Create Virtual Environment

It is strongly recommended to create a virtual environment unless your computer has no other Python projects or AI projects. Virtual environments can effectively avoid many potential errors.

Enter cmd in the address bar of the folder you just created D:/python/f5ttsnew and press Enter (Mac systems please use the terminal to enter this folder).

Execute the following command to create a virtual environment: python -m venv venv. After execution, a folder named venv will be added to the folder.

Next, activate the virtual environment (note the spaces and dots):

  • Windows System: .\venv\scripts\activate
  • Mac System: . ./venv/bin/activate

After the virtual environment is activated, the (venv) text will be added before the command line prompt. Please make sure that all subsequent operations are performed in this virtual environment, and check whether the command line prompt has (venv) before each operation.

The command line has (venv) to represent activation

Install Dependencies

In the terminal with the virtual environment activated, continue to enter the following command (note the spaces and dots):

pip install -e .

Wait for the installation to complete. If CUDA acceleration is required, continue to execute the following command (this is one line of command, do not wrap):

# Install pytorch with your CUDA version, e.g.
pip install torch==2.4.0+cu124 torchaudio==2.4.0+cu124 --extra-index-url https://download.pytorch.org/whl/cu124

Configure Scientific Internet Access Environment

Important Note: F5-TTS needs to download models online from the huggingface.co website. Since this website has been blocked in China and cannot be connected directly, you must configure a scientific Internet access environment and enable global or system proxy before starting.

If the VPN tool you are using provides an HTTP port (as shown in the figure below):

Check if scientific software provides a port

Please enter the following command in the terminal to set the proxy:

  • Windows System: set https_proxy=http://127.0.0.1:10808 (Please replace the port number with the actual port you are using)
  • Mac System: https_proxy=http://127.0.0.1:10808 (Please replace the port number with the actual port you are using)

You can also directly modify the code to set the proxy, avoiding manual input in the terminal each time. Open the F5-TTS root directory/src/f5_tts/infer/infer_gradio.py file and add the following code at the top of the file:

python
import os
os.environ['https_proxy']='http://127.0.0.1:10808' # Fill in according to your actual proxy address

Launch WebUI Interface

After configuring the scientific Internet access environment, enter the following command in the terminal to launch the WebUI:

f5-tts_infer-gradio

The program will automatically download the model for the first time, which may be slow, please be patient. When starting later, the program may still connect to huggingface.co for detection. It is recommended to keep the proxy turned on to avoid errors.

After the startup is successful, the terminal will display the IP address and port number, as shown in the figure below:

Successful startup when the IP and port are displayed, the first time is very slow

Open the displayed address in the browser, the default is http://127.0.0.1:7860.

webui interface

Connect to pyVideoTrans API

To use F5-TTS in the video translation software, you need to start F5-TTS first and keep the terminal window open.

Then, open the video translation software, and in the menu, select "TTS Settings" -> "F5-TTS API" in order, and fill in the startup address of F5-TTS, the default is http://127.0.0.1:7860. If your startup address is not the default address, please fill it in according to the actual address.

In the "Reference Audio" column, fill in the following content:

Audio file name to use # Corresponding text in the audio file

Note: Please place the reference audio file in the f5-tts folder under the pyVideotrans project root directory. If the folder does not exist, please create it manually. For example, you can name the reference audio file nverguo.wav.

Put the reference audio in the f5-tts folder in the pyVideotrans software, don't get it wrong

The following is a filling example:

Reference audio and text in the reference audio

Re-Recognition?: By default, the reference audio (the subtitles recognized during cloning) will be sent to F5-TTS to avoid F5-TTS starting Whisper for speech recognition, saving time and improving efficiency, but sometimes you may want F5-TTS to re-recognize it, which can improve the cloning quality to a certain extent. You can select this check box at this time, but note that if you select it for the first time, F5-TTS will download the openai-whisper-v3 model online from huggingface.co, please make sure you have scientific Internet access.

Solve Local Area Network Problems

If your F5-TTS is deployed on another computer in the local area network, you need to modify the F5-TTS code to support local area network access.

Open the F5-TTS project directory/src/f5_tts/infer/infer_gradio.py file and add the following code below line 16:

python
# Add LAN start
import os
from pathlib import Path

ROOT=Path(os.getcwd()).as_posix()
TMP=f'{ROOT}/tmp'
Path(TMP).mkdir(exist_ok=True)
os.environ['GRADIO_TEMP_DIR']=TMP
gr.set_static_paths(paths=[TMP,tempfile.gettempdir()])
print(TMP)

## Add LAN end

Schematic diagram of code addition location:Note the code addition location

After saving the modification, restart F5-TTS. Then fill in the IP address and port number after F5-TTS is started in pyVideotrans, such as http://192.168.0.12:7860.

Add Other Languages

If you need to use models in other languages, you also need to modify the F5-TTS project directory/src/f5_tts/infer/infer_gradio.py file.

Find the code around line 59:

python
DEFAULT_TTS_MODEL_CFG = [
    "hf://SWivid/F5-TTS/F5TTS_v1_Base/model_1250000.safetensors",
    "hf://SWivid/F5-TTS/F5TTS_v1_Base/vocab.txt",
    json.dumps(dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)),
]

Code location diagram:

By default, the official Chinese and English models are configured here. If you need to use models in other languages, please modify according to the following instructions. After the modification, you need to restart F5-TTS and make sure that the scientific Internet access environment has been configured so that the program can download new language models online. After the download is successful, clone a timbre through the WebUI for testing first, and then use it through pyVideoTrans.

Important Note: Before use, please make sure that the dubbing text language in pyVideoTrans is consistent with the model language selected in F5-TTS.

The following is the configuration information of each language model:

  1. French:

    python
    DEFAULT_TTS_MODEL_CFG = [
        "hf://RASPIAUDIO/F5-French-MixedSpeakers-reduced/model_last_reduced.pt",
        "hf://RASPIAUDIO/F5-French-MixedSpeakers-reduced/vocab.txt",
        json.dumps({"dim": 1024, "depth": 22, "heads": 16, "ff_mult": 2, "text_dim": 512, "text_mask_padding": False, "conv_layers": 4, "pe_attn_head": 1}),
    ]
  2. Hindi:

    python
    DEFAULT_TTS_MODEL_CFG = [
        "hf://SPRINGLab/F5-Hindi-24KHz/model_2500000.safetensors",
        "hf://SPRINGLab/F5-Hindi-24KHz/vocab.txt",
        json.dumps({"dim": 768, "depth": 18, "heads": 12, "ff_mult": 2, "text_dim": 512, "text_mask_padding": False, "conv_layers": 4, "pe_attn_head": 1})
    ]
  3. Italian:

    python
    DEFAULT_TTS_MODEL_CFG = [
        "hf://alien79/F5-TTS-italian/model_159600.safetensors",
        "hf://alien79/F5-TTS-italian/vocab.txt",
        json.dumps({"dim": 1024, "depth": 22, "heads": 16, "ff_mult": 2, "text_dim": 512, "text_mask_padding": False, "conv_layers": 4, "pe_attn_head": 1})
    ]
  4. Japanese:

    python
    DEFAULT_TTS_MODEL_CFG = [
        "hf://Jmica/F5TTS/JA_25498980/model_25498980.pt",
        "hf://Jmica/F5TTS/JA_25498980/vocab_updated.txt",
        json.dumps({"dim": 1024, "depth": 22, "heads": 16, "ff_mult": 2, "text_dim": 512, "text_mask_padding": False, "conv_layers": 4, "pe_attn_head": 1})
    ]
  5. Russian:

    python
    DEFAULT_TTS_MODEL_CFG = [
        "hf://hotstone228/F5-TTS-Russian/model_last.safetensors",
        "hf://hotstone228/F5-TTS-Russian/vocab.txt",
        json.dumps({"dim": 1024, "depth": 22, "heads": 16, "ff_mult": 2, "text_dim": 512, "text_mask_padding": False, "conv_layers": 4, "pe_attn_head": 1})
    ]
  6. Spanish:

    python
    DEFAULT_TTS_MODEL_CFG = [
        "hf://jpgallegoar/F5-Spanish/model_last.safetensors",
        "hf://jpgallegoar/F5-Spanish/vocab.txt",
        json.dumps({"dim": 1024, "depth": 22, "heads": 16, "ff_mult": 2, "text_dim": 512, "conv_layers": 4})
    ]
  7. Finnish:

    python
       DEFAULT_TTS_MODEL_CFG = [
        "hf://AsmoKoskinen/F5-TTS_Finnish_Model/model_common_voice_fi_vox_populi_fi_20241206.safetensors",
        "hf://AsmoKoskinen/F5-TTS_Finnish_Model/vocab.txt",
        json.dumps({"dim": 1024, "depth": 22, "heads": 16, "ff_mult": 2, "text_dim": 512, "text_mask_padding": False, "conv_layers": 4, "pe_attn_head": 1})]

You can pay attention to official updates, and other languages can be added in a similar way. Address: https://github.com/SWivid/F5-TTS/blob/main/src/f5_tts/infer/SHARED.md

Common Errors and Precautions

  1. During API use, you can close the WebUI interface in the browser, but you cannot close the terminal window that starts F5-TTS.

    This interface cannot be closed, otherwise the API cannot be called

  2. Can I dynamically switch the model in F5-TTS? No. You need to manually modify the code as described above, and then restart the WebUI.

  3. Frequently occurring this kind of error

    raise ConnectTimeout(e, request=request)
requests.exceptions.ConnectTimeout: (MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /SWivid/F5-TTS/resolve/main/F5TTS_v1_Base/vocab.txt (Caused by ConnectTimeoutError(<urllib3.connection.HTTPSConnection object at 0x000002174796DF60>, 'Connection to huggingface.co timed out. (connect timeout=10)'))"), '(Request ID: 0458b571-90ab-4edd-ae59-b93bd603cdd0)')

Proxy issue, please use scientific Internet access and use a smooth proxy, refer to the above to configure the scientific Internet access environment

  1. How to prohibit connecting to huggingface.co every time?

Please make sure that you have successfully cloned at least once and the model has been downloaded Open F5-TTS root directory/src/f5_tts/infer/utils_infer.py

Search for snapshot_download, find the code line as shown in the figure

Modify to

local_path = snapshot_download(repo_id="nvidia/bigvgan_v2_24khz_100band_256x", cache_dir=hf_cache_dir,local_files_only=True)

Then search for hf_hub_download, find the 2 lines of code as shown in the figure

Modify to

config_path = hf_hub_download(repo_id=repo_id, cache_dir=hf_cache_dir, filename="config.yaml",local_files_only=True)
            model_path = hf_hub_download(repo_id=repo_id, cache_dir=hf_cache_dir, filename="pytorch_model.bin",local_files_only=True)

In fact, the new parameter ,local_files_only=True is added to the place where these 3 lines of code are called. Please make sure that the model has been downloaded locally, otherwise an error that the model cannot be found will be reported.

  1. F5-TTS is normal after deployment, but pyVideotrans returns {detail:"Not found"} when testing
    • Check if other AI projects are occupying the port. Generally, AI projects with interfaces use the gradio interface, which is also 7860 by default. Close the others and restart F5-TTS
    • If pyVideotrans is deployed with source code, please execute pip install --upgrade gradio_client and then retry
    • Restart F5-TTS, use the command f5-tts_infer-gradio --api to start