tronghieuit/v-tts

What it solves

V-TTS provides a lightweight, efficient solution for Vietnamese text-to-speech (TTS), specifically addressing the need for high-quality voice synthesis and voice cloning that can run on standard consumer hardware without requiring a GPU.

How it works

The system uses a neural architecture consisting of a synthesizer, a speaker encoder for identity extraction, a style encoder for prosody and emotion, and a prosody predictor for pitch and energy. For zero-shot cloning, it extracts embeddings from a 3-10 second reference audio clip to generate speech in that specific voice. It includes a dedicated Vietnamese phonemizer to support both Northern and Southern dialects.

Who it’s for

• Content creators needing Vietnamese voiceovers without professional recording equipment.
• Developers looking to integrate Vietnamese TTS into applications via a simple Python API or ONNX runtime.
• Windows users who want a plug-and-play experience via a standalone .exe application.

Highlights

• Ultra-lightweight: Only 74.8M parameters, allowing it to run entirely on CPU.
• Zero-Shot Cloning: Can clone any voice using only 3-10 seconds of reference audio without fine-tuning.
• High Performance: Achieves inference speeds 3-4x faster than real-time on CPU.
• Dialect Support: Built-in support for Northern and Southern Vietnamese voices.
• Flexible Deployment: Available as a Windows app, Docker container, Python package, and ONNX exports for Web and Android.