Release Notes - February 28, 2025

Overview

This release introduces a fully functional MLX-Audio package with text-to-speech capabilities, complete with testing infrastructure and CI/CD integration via GitHub Actions.

New Features

• Text-to-Speech Generation: Added complete generation pipeline with audio output functionality
• Audio Joining: New functionality to join multiple audio segments
• Model Quantization: Added support for model quantization to improve performance
• GitHub Actions: Implemented CI/CD workflows for automated testing and deployment

Improvements

• Kokoro MLX porting: Completed refactoring of the entire model to MLX framework:
  • Text encoder with BERT implementation
  • Decoder with improved audio quality
  • Duration, indices, and alignment target prediction
  • Custom Bidirectional LSTM, Weight norm for CNNs, AdaLayerNorm and Generator layers
• SafeTensors Support: Added working implementation for SafeTensors format
• Pipeline Structure: Restructured the generation pipeline for better maintainability

Bug Fixes

• Fixed model loading mechanism
• Resolved issues with text encoder LayerNorm operation
• Fixed generator functionality
• Addressed issues in LSTM and AdaLayerNorm implementations
• Refactored and fixed ConvWeight component

Full Changelog: https://github.com/Blaizzy/mlx-audio/commits/v0.0.1