Train Your Voice Model

Once you've prepared your high-quality audio segments, you're ready to train your custom voice model.

What you'll get from this step:

• t2s.pt - Text-to-speech model file
• vits.pt - Voice synthesis model file

What You'll Need

We'll use Google Colab for training since it's free and provides the GPU power you need.

Make sure you have:

1. A Google Drive account
2. A Google Colab account

Step 1: Set Up Model Files

First, you'll need access to the pre-trained model files.

1. Add model files to your Drive: Visit the GPT-SoVITS Google Drive Folder
2. Create a shortcut: Click "Add shortcut to Drive" to add these files to your workspace (see image below)

Step 2: Open the Training Notebook

1. Open the template: Go to the Training Notebook
2. Make your own copy: Go to File → Save a copy in Drive
3. Connect your Drive: Click the Drive folder icon to mount Google Drive

4. Grant permissions: Allow Colab to access your Drive files

Click the folder icon on the left sidebar to browse your Drive contents.

Step 3: Upload Your Audio Data

Now you'll upload the audio segments you prepared in the last article.

1. Create a list file: Make a .list file containing your transcriptions from the website (It may have done by the Segment Audio Files with EchoKit's Audio Processor)
2. Organize your files: Put the .list file in the same folder as your audio files (It may have done by the Segment Audio Files with EchoKit's Audio Processor)
3. Compress the folder: Create a zip file of your entire audio folder (It may have done by the Segment Audio Files with EchoKit's Audio Processor)
4. Upload to Colab: Use the upload icon in Colab's file browser to upload your zip folder.
5. Extract the files: Run unzip your-folder.zip in the terminal.

Step 4: Copy Model Files to Your Workspace

The shortcut you created points to the model files, but you need actual copies in your workspace.

1. Open terminal: Click the terminal icon (bottom-left in Colab)
2. Run these commands:

   cd drive/MyDrive
   mkdir -p workspace/GPT-SoVITS/
   cp -r GPT-SoVITS/ workspace/GPT-SoVITS/

Note: This will take a few minutes since the files are several GB in size.

Step 5: Install Dependencies

Run each cell in your notebook in order, ending with:

!pip install -r requirements.txt

Important: When this finishes, you'll be asked to restart the runtime. Click "Restart".

Step 6: Prepare Your Training Data

After restarting, run the cells starting from:

%cd /content/drive/MyDrive/workspace/GPT-SoVITS

Continue until you reach this code block:

s = process_ad_list('/content/train-model/echokit.list', 'echokit', 'en')
with open('/content/train-model/echokit.new.list', 'w') as f:
    f.write(s)

Before running this command, update these values to match your setup:

• File path: /content/train-model/echokit.list → your actual .list file path
• Model name: echokit → your preferred model name
• Language: en → your language code

Step 7: Start the Training Interface

Run this command to launch the web interface:

!export is_share=True && python webui.py

You should see output like:

Running on local URL:  http://0.0.0.0:9874
Running on public URL: https://fbefe7d567a1054c43.gradio.live

Open the public URL in your browser and click the 1-GPT-SOVITS-TTS tab.

Step 8: Configure Training Settings

1. Enter your model details:
   • Model name: Use the same name from Step 6
   • List file path: Path to your .list file from Step 6

2. Run preprocessing steps (click each button and wait for completion):
   • "Open Tokenization & BERT Feature Extraction"
   • "Open Speech SSL Feature Extraction"
   • "Open Semantics Token Extraction"

Step 9: Train Your Models

1. Go to the 1B-Fine-Tuning tab
2. Start training (click each button and wait):
   • "Open SoVITS Training"
   • "Open GPT Training"

Step 10: Test Your Model

Once the training is done, you can start to test your voice-cloned model.

1. Go to the 1C-Inference tab
2. Select your trained models:
   • GPT weight: Choose GPT_weights_v2Pro/[model-name]-e15.ckpt
   • SoVITS weight: Choose SoVITS_weights_v2Pro/[model-name]_e8_s104.pth
3. Click "Open TTS Inference WebUI"

Check your Colab notebook for the inference URL:

Step 11: Test Voice Generation

Open the inference URL in your browser:

1. Upload reference audio: Use one of your training audio files
2. Enter reference text: Type the exact transcription of your reference audio

3. Test generation:
   • Enter text you want to synthesize
   • Select the correct language
   • Click "Start inference"

If you're not satisfied with the results, review your training audio quality and consider retraining.

Step 12: Export Your Models

Once you're happy with the voice quality:

1. Prepare reference audio: Convert your reference audio to 16kHz:

   ffmpeg -i REFERENCE_AUDIO_PATH.wav -ar 16000 -ac 1 output.wav

2. Export the models: Run this command (update the paths to match your setup):

   !python GPT_SoVITS/stream_v2pro.py \
   --gpt_model GPT_weights_v2Pro/echokit-e15.ckpt \
   --sovits_model SoVITS_weights_v2Pro/echokit_e8_s104.pth \
   --ref_audio /content/output.wav \
   --ref_text 'It would be our book of nursery rhymes or numbers or the alphabet.' \
   --output_path /content/models \
   --version v2Pro \
   --device=cuda

Make sure to update:

• gpt_model: Path to your GPT model
• sovits_model: Path to your SoVITS model
• ref_audio: Path to your 16kHz reference audio
• ref_text: Exact transcription of your reference audio
• output_path: Where to save the final model files

Final Output

After export completes, you'll have:

• out_final.wav - Test audio to verify everything works
• t2s.pt - Your text-to-speech model
• vits.pt - Your voice synthesis model

Listen to out_final.wav to confirm your model is working properly.

Next step: Deploy your models to start using your custom voice with

• t2s.pt
• vits.pt
• Reference audio in 16k
• The transcription of the reference audio