SparkVox

June 9, 2025 ยท View on GitHub

SparkVox is a training framework focused on speech generation, while also supporting a range of related speech tasks, including speaker attribute recognition, emotion recognition, audio codecs, and speech synthesis.

Supported Tasks

  • Speaker Attribute Recognition
    • Age prediction
    • Gender prediction
  • Codec
    • BiCodec
    • BigCodec
  • Speech Synthesis
    • SparkTTS

Project Structure

  • bins:
    • train_pl: The main training entry point for all tasks.
  • egs:
    • task (e.g. codec, speech_synthesis): Example training scripts for each task.
  • sparkvox
    • models: Model implementations for different tasks.
  • tools: Utilities for data processing, model inference, and feature extraction.
  • utils: Common utilities for tasks such as reading and processing audio files, as well as general training tools.

Examples