AutoArk/GPA

What it solves

GPA (General Purpose Audio) addresses the need for a single, unified model capable of handling multiple core audio tasks—Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and Voice Conversion (VC)—rather than relying on separate specialized models for each function.

How it works

GPA uses an auto-regressive transformer architecture to integrate speech understanding and generation. The latest version, GPA-v1.5, provides native PyTorch and Hugging Face workflows for inference and training. For deployment, it offers an ONNX Runtime that supports CLI tools, a FastAPI service, and a browser-based UI. There is also a standalone, lightweight GPA-TTS runtime optimized for local CPU inference on Mac, Linux, and edge devices, utilizing quantization (INT4/INT8) to reduce its footprint.

Who it’s for

This project is for AI researchers and developers who want a unified audio model that performs near-SOTA (State-of-the-Art) performance on ASR and TTS, or those looking for a lightweight, edge-ready voice-cloning TTS system.

Highlights

• Unified Architecture: Combines ASR and TTS into one model (with VC on the roadmap).
• Near-SOTA Performance: Delivers competitive results on ASR and TTS benchmarks compared to larger models.
• Edge-Ready TTS: A standalone GPA-TTS runtime with zero-shot voice cloning and selectable decoder precision (INT8, FP16, FP32).
• Broad Deployment Support: Compatible with torch, vLLM, llama-cpp, sglang, and mlx-lm.