phuocnguyen90/VieNeu-TTS-v3-Turbo-GGUF

🦜 VieNeu-TTS v3 Turbo GGUF (48 kHz)

This repository contains the official GGUF quantized model package for VieNeu-TTS v3 Turbo (48 kHz), fully integrated and supported natively in audio.cpp (merged in PR #80).

Model Credits & Acknowledgments: This is a GGUF quantization of the original VieNeu-TTS-v3-Turbo model created by pnnbao-ump. All credit for the underlying architecture and model training goes to the original authors.

It packs both the compressed Qwen3 talker backbone and the speech tokenizer/decoder weights into a single optimized 163 MB file, enabling instant load times and ultra-fast, torch-free C++ generation on CPU.

---

📊 Performance Benchmarks (CPU)

The following benchmark compares native Python backends (PyTorch CPU, ONNX CPU int8, and ONNX CPU fp32) against the compiled C++ CPU backend (using 4 threads, generating ~3.5 seconds of high-fidelity 48 kHz audio):

Metric	Native PyTorch CPU	Upstream ONNX CPU (onnx_int8)	Upstream ONNX CPU (onnx_update / FP32)	C++ CPU (audio.cpp)
Model Load Time	9.30 seconds	8.93 seconds	17.04 seconds	Instant (<0.01s)
First Run (Warmup)	1.70 seconds	1.57 seconds	1.40 seconds	0.07 seconds
Inference Time (Avg)	1.58 seconds	1.51 seconds	2.80 seconds	0.89 seconds
Real-Time Factor (RTF)	0.44	0.43	0.80	0.24 (4x faster)
Combined Load+Warmup	11.00 seconds	10.50 seconds	18.44 seconds	0.08 seconds

---

🛠️ Usage

1. Build the C++ CLI Runner

Clone the official audio.cpp repository and build the target binary:

# Clone the upstream repository
git clone https://github.com/0xShug0/audio.cpp.git
cd audio.cpp

# Build the executable
chmod +x scripts/build_linux.sh
./scripts/build_linux.sh --backend cpu --target audiocpp_cli

2. Download the Model File

Download this GGUF model along with its tokenizer configurations using audio.cpp's built-in model manager:

python tools/model_manager.py install vietneu_tts_v3_turbo

Or manually:

mkdir -p models/VieNeu-TTS-v3-Turbo
cd models/VieNeu-TTS-v3-Turbo

wget https://huggingface.co/phuocnguyen90/VieNeu-TTS-v3-Turbo-GGUF/resolve/main/model.gguf
wget https://huggingface.co/phuocnguyen90/VieNeu-TTS-v3-Turbo-GGUF/resolve/main/config.json
wget https://huggingface.co/phuocnguyen90/VieNeu-TTS-v3-Turbo-GGUF/resolve/main/tokenizer.json
wget https://huggingface.co/phuocnguyen90/VieNeu-TTS-v3-Turbo-GGUF/resolve/main/tokenizer_config.json
wget https://huggingface.co/phuocnguyen90/VieNeu-TTS-v3-Turbo-GGUF/resolve/main/special_tokens_map.json
mkdir -p speech_tokenizer
wget -O speech_tokenizer/config.json https://huggingface.co/phuocnguyen90/VieNeu-TTS-v3-Turbo-GGUF/resolve/main/speech_tokenizer/config.json

3. Generate Speech

Run inference via the C++ CLI:

../../build/linux-cpu-release/bin/audiocpp_cli   --task tts   --family vietneu_tts   --model model.gguf   --backend cpu   --voice-ref ../../assets/resources/sample.wav   --reference-text "Some call me nature. Others call me Mother Nature."   --text "sˈin tʃˈaː2w tˈeɜ zˈəːɜj. ɗˈəɪ lˌaː2 bˈaː4n tˈy4 ŋˈiɛ6m."   --out output.wav