Local Accent Correction

Installation

1. Prerequisites:

   • Python 3.12+
   • UV package manager (recommended)
   • CUDA-capable GPU (recommended for performance)

2. Install Dependencies:

   uv sync

Model Setup

This pipeline requires a pretrained HiFi-GAN model for waveform synthesis.

1. Create directory:

   mkdir -p pretrained_models/hifigan

2. Download Model: Download generator.ckpt from the SpeechBrain HiFi-GAN HuggingFace repo (or your preferred source).

3. Place File: Save the generator.ckpt file inside pretrained_models/hifigan/.

Usage

Run the main pipeline on an audio file:

uv run main.py --source input_audio.wav

Options

Argument	Description	Default
--source	Path to the source audio file (accented).	Required
--pitch	Pitch transfer strength (0.0 - 1.0).	0.7
--timbre	Timbre transfer strength (0.0 - 1.0).	0.4
--energy	Energy transfer strength (0.0 - 1.0).	0.3
--gender	Force 'male' or 'female' gender for TTS.	Auto-detect
--text	Override transcription with provided text.	Auto-transcribe

Example

uv run main.py --source ishika-before-clean.wav --pitch 0.8 --timbre 0.5

Project Structure

• main.py: Entry point for the pipeline.
• pipeline.py: Core logic for audio processing (DTW, PSOLA, spectral transfer).
• stt.py: Speech-to-Text module using faster-whisper.
• tts.py: Text-to-Speech module using Chatterbox.
• voices/: Directory containing reference voice samples.
• pretrained_models/: Directory for local model weights.

License

[License Information Here]