YourTTS Documentation

MOS and Sim-MOS scores are organized by language in specific directories. You can find the scores for each audio file in the following locations:

• English: EN/ directory
• Portuguese: PT/ directory
• Bilingual (English & Portuguese): EN-PT/ directory

You can recalculate the MOS (Mean Opinion Score) or Sim-MOS (Similarity Mean Opinion Score) along with their respective confidence intervals using the compute_similarity_MOS.py script. You must provide the path to the corresponding CSV file via the --csv_path flag.

# To recalculate MOS for English
python3 compute_similarity_MOS.py --csv_path EN/naturalness-MOS.csv

# To recalculate Sim-MOS for English
python3 compute_similarity_MOS.py --csv_path EN/Sim-MOS.csv

To review the audio samples used for Mean Opinion Score (MOS) and Similarity Mean Opinion Score (Sim-MOS) evaluations, you can download the complete collection from the official release zip file.

https://github.com/Edresson/YourTTS/releases/download/MOS/Audios_MOS.zip

To replicate the training of Experiment 1, you can use the official recipe provided in the Coqui TTS repository. This recipe automates downloading, resampling, extracting speaker embeddings, and training the model without code changes.

Recipe Location: https://github.com/coqui-ai/TTS/blob/dev/recipes/vctk/yourtts/train_yourtts.py

Once your config.json is fully configured with the correct dataset paths, speaker embeddings, and encoder settings, start the training process using the train_tts.py script.

Command:

python3 TTS/bin/train_tts.py --config_path config.json

Tip: Use --restore_path {checkpoint_path} to perform transfer learning from an existing checkpoint to accelerate training.

Before training, you must extract speaker embeddings using the released speaker encoder. This requires a model path, a speaker encoder config, a dataset config, and an output path for the resulting JSON file.

Command:

python3 TTS/bin/compute_embeddings.py --model_path model_se.pth.tar --config_path config_se.json --config_dataset_path  config.json --output_path d_vector_file.json

Required Files:

• model_se.pth.tar: Speaker encoder checkpoint.
• config_se.json: Speaker encoder configuration.
• config.json: Your adjusted dataset configuration.

When setting up training, you must modify the following keys in your config.json file:

• datasets: Update this to point to your specific dataset.
• d_vector_file: Set this to the path of the speaker embedding file generated by compute_embeddings.py (e.g., d_vector_file.json).
• output_path: The directory where checkpoints and training logs will be saved.
• speaker_encoder_config_path: The path to the speaker encoder config used for computing speaker consistency loss (e.g., config_se.json).
• speaker_encoder_model_path: The path to the speaker encoder checkpoint used for computing speaker consistency loss.

If the TTS library is not installed globally, you must manually add its path to sys.path to enable imports from TTS.tts.utils.synthesis, TTS.tts.models, and TTS.config.

import sys
TTS_PATH = "TTS/"
sys.path.append(TTS_PATH)

from TTS.tts.utils.synthesis import synthesis
from TTS.tts.models import setup_model
from TTS.config import load_config
# ... other imports

The following checkpoints are available for use (licensed under CC BY-NC-ND 4.0):

Note: Other experimental checkpoints (Exp 2, 3, 4) are currently not available.

Use the tts command from the Coqui TTS library to perform zero-shot multi-speaker text-to-speech. You must provide the target text, the model name, a reference audio file (--speaker_wav) representing the target speaker's voice, and the language index.

tts --text "This is an example!" --model_name tts_models/multilingual/multi-dataset/your_tts --speaker_wav target_speaker_wav.wav --language_idx "en"