GPA is a unified transformer model that performs automatic speech recognition, text-to-speech synthesis, and voice conversion using a single compact architecture.
The tool addresses the inefficiency of maintaining separate models for different audio tasks by consolidating automatic speech recognition, text-to-speech, and voice conversion into one autoregressive transformer. This unified approach reduces model size while maintaining near state-of-the-art performance on both ASR and TTS. The model uses an audio-language framework where a single set of weights handles speech understanding and generation tasks through a shared transformer backbone.
Developers should choose this tool when model size and inference efficiency matter more than task-specific optimization, particularly for edge deployment scenarios or resource-constrained environments. The project suits applications requiring multiple audio capabilities without the overhead of managing separate models. The tool offers quantization options including INT8 and INT4 variants for further size reduction, and provides ONNX runtime support for deployment through CLI tools, FastAPI services, or browser-based interfaces. A standalone TTS runtime with voice cloning support is available for users focused primarily on synthesis tasks.
The project maintains active development with recent releases introducing higher-quality decoder options alongside existing quantized variants, selectable at runtime. ONNX runtime support has been added as a deployment pathway alongside the core model. The codebase includes comprehensive documentation covering quick starts, deployment guidance, and evaluation benchmarks across different model versions. Development activity shows ongoing refinement of both the core unified model and specialized extraction variants for specific use cases.
