Persian FastPitch

Persian FastPitch is a fast and efficient TTS model adapted to generate mel-spectrograms from Persian text, using the FastPitch architecture. FastPitch, originally developed by NVIDIA, offers a significant speed advantage over Tacotron-based models by directly predicting pitch and leveraging parallelized training.

This implementation is based on NVIDIA's FastPitch repository, adapted for Persian TTS. Changes made include language-specific adjustments and customizations to handle the unique aspects of Persian phonetics.

Key Modifications

1. Prepare Persian Data: Collect Persian audio files and generate corresponding phoneme sequences for each. This uses phonemes instead of text for better accuracy, handling challenges like missing vowels in Persian script.
2. Resolve Colab Error in data_function.py: Minor edits to address Colab compatibility issues. See issue #1016 for more details.
3. Update cleaners.py in common/text/: Adapt character handling to Persian phonemes.
4. Customize Training Parameters: Modify scripts/train.sh and train.py to fit Persian data.
5. Adjust Inference Parameters: Update scripts/inference_example.sh for Persian-specific inference.

---

How to Use

1. Clone this Repository

   git clone https://github.com/Adibian/Persian-FastPitch.git
   cd Persian-FastPitch
   

2. Install Requirements

   pip install -r requirements.txt
   

3. Add Your Data

   • Place audio files in wavs/
   • Add training and validation phoneme transcriptions to filelists/
   • Add test phoneme transcriptions to phrases/

4. Extract Pitch

   • Run this command to extract pitch values for the audio files

       python prepare_dataset.py \
           --wav-text-filelists filelists/audio_text_train.txt \
                                filelists/audio_text_val.txt \
           --n-workers 16 \
           --batch-size 1 \
           --dataset-path 'wavs/' \
           --extract-pitch \
           --f0-method pyin

5. Install Additional Dependencies

   • Run the following to install NVIDIA Apex and download CMUdict:

        git clone https://github.com/NVIDIA/apex
        cd apex; pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
        bash scripts/download_cmudict.sh
     

6. Train the Model

   • Use this command to start training. Checkpoints will be saved in output/:

   bash scripts/train.sh
   

7. Download WaveGlow Vocoder

   • WaveGlow is required to convert mel-spectrograms into audio. Download it using:

     bash scripts/download_waveglow.sh
     

8. Run Inference

   • To synthesize audio for a test file in phrases/, run:

     bash scripts/inference_example.sh

   • The synthesized audio will be saved in output/audio_test_file/.