| license: apache-2.0 | |
| language: | |
| - vi | |
| # VieNeu-Codec: The Heart of VieNeu-TTS v2 | |
| **VieNeu-Codec** is the high-performance audio engine built specifically for the upcoming **VieNeu-TTS v2**. It is a neural audio codec trained on over **20,000 hours** of diverse Vietnamese and English speech data, ensuring state-of-the-art robustness, natural prosody, and crystal-clear audio reconstruction. | |
| This repository provides the optimized ONNX versions of the VieNeu-Codec for production use. | |
| ## 🚀 Key Features | |
| - **24kHz High-Fidelity**: Crystal clear audio reconstruction optimized for the Vietnamese language. | |
| - **Zero-Shot Voice Cloning**: Clone any voice with just 5 seconds of reference audio. | |
| - **Optimized for VieNeu-TTS v2**: Seamlessly integrates with the next-generation LLM backbone of VieNeu-TTS. | |
| - **Two Deployment Modes**: Includes both FP32 (High Quality) and INT8 (High Speed) decoders. | |
| ## 📦 Model Components | |
| - **`vieneu_decoder.onnx`**: (FP32) High-fidelity audio decoder for maximum quality. | |
| - **`vieneu_decoder_int8.onnx`**: (INT8) Quantized decoder for fast CPU inference. | |
| ## 🛠️ Usage | |
| ### Synthesize Speech | |
| Combine the speaker embedding with content tokens from your LLM (VieNeu-TTS v2): | |
| ```python | |
| sess_dec = ort.InferenceSession("vieneu_decoder.onnx") | |
| audio = sess_dec.run(None, { | |
| "content_ids": ids, | |
| "voice": embedding | |
| })[0] | |
| ``` | |
| ## 📄 License & Attribution | |
| Author: **Pham Nguyen Ngoc Bao** | |
| Project: **VieNeu-Codec (for VieNeu-TTS v2)** | |
| Version: 2.0 |