VieNeu-TTS.cpp

Native C++ runtime for VieNeu-TTS, focused on local Vietnamese text-to-speech inference, voice presets, and zero-shot voice cloning.

Highlights

• Native C++17 implementation with a small CLI and a stable C ABI.
• VieNeu v3 ONNX pipeline for convenient local inference.
• VieNeu v3 native path using llama.cpp for the backbone and native acoustic weights.
• Voice presets through voices_v3_turbo.json.
• Zero-shot voice cloning from a reference WAV file.
• Windows helper scripts for building, downloading runtime assets, and running smoke tests.

Models

Download the model assets from Hugging Face:

• VieNeu v3 Turbo: pnnbao-ump/VieNeu-TTS-v3-Turbo
• MOSS audio tokenizer ONNX codec: OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano-ONNX

The convenience script scripts/run-v3-tts-test.ps1 downloads the required v3 ONNX assets into .models/vieneu-v3-turbo automatically. Native v3 assets such as backbone.gguf, vieneu_v3_heads.npz, and acoustic/vieneu_acoustic_weights.npz can be exported with the scripts in scripts/.

Build

Requirements:

• C++17 compiler: MSVC, GCC, or Clang
• CMake 3.14+
• llama.cpp submodule at third_party/llama.cpp
• ONNX Runtime SDK

On Windows, the easiest path is:

scripts\build-local.ps1

Useful options:

scripts\build-local.ps1 -Clean
scripts\build-local.ps1 -NoPackage
scripts\build-local.ps1 -OnnxRuntimeVersion 1.24.4

Manual CMake build:

cmake -S . -B build -DONNXRUNTIME_ROOT="C:/path/to/onnxruntime" -DVIENEU_LLAMA_DIR="third_party/llama.cpp"
cmake --build build --config Release

Build outputs include:

• vieneu-tts-core: static runtime library
• vieneu-tts: shared library exposing the C ABI
• vieneu-tts-cli: command-line synthesizer
• test-abi-c: C ABI compile test

Quick Start

Run a VieNeu v3 ONNX smoke test:

scripts\run-v3-tts-test.ps1 `
  -Text "Xin chào, đây là bài kiểm tra VieNeu TTS v3 Turbo từ runtime C++." `
  -Output outputs\vieneu-v3-test.wav

Run the CLI directly after building:

build\Release\vieneu-tts-cli.exe `
  --profile vieneu-v3-onnx `
  --model-dir .models\vieneu-v3-turbo `
  --onnx-dir .models\vieneu-v3-turbo\onnx `
  --codec-dir .models\vieneu-v3-turbo\codec `
  --voices-json .models\vieneu-v3-turbo\voices_v3_turbo.json `
  --text "Xin chào, tôi là giọng nói tiếng Việt được tổng hợp bằng C++." `
  --output outputs\hello.wav

Run native v3 voice cloning benchmark with the included reference clips:

scripts\run-v3-native-voice-clone-benchmark.ps1 -NoBuild

Run one reference only:

scripts\run-v3-native-voice-clone-benchmark.ps1 -NoBuild -Ref example_3

Generated WAV files are written to outputs/ by default.

CLI Profiles

vieneu-tts-cli supports these profiles:

• vieneu-v3-onnx: v3 model assets plus ONNX Runtime sessions.
• vieneu-v3-native: v3 model directory with native GGUF/backbone assets and MOSS codec.
• vieneu-v2-turbo: legacy GGUF speech LM path with ONNX encoder/decoder assets.

Common options:

--profile NAME
--model-dir PATH
--onnx-dir PATH
--codec-dir PATH
--voices-json PATH
--ref-audio PATH
--text TEXT
--voice ID
--output PATH
--temperature VALUE
--top-k VALUE
--top-p VALUE
--max-new-frames N
--max-chars N
--threads N

Project Layout

src/
  vieneu/        Stable C ABI and VieNeu v2/v3 runtime orchestration
  codecs/        ONNX Runtime helpers for audio codec models
  backends/      llama.cpp wrappers for GGUF/backbone inference
tools/
  vieneu-tts-cli.cpp
scripts/
  build-local.ps1
  run-v3-tts-test.ps1
  run-v3-native-preset-benchmark.ps1
  run-v3-native-voice-clone-benchmark.ps1
  export-v3-native-assets.py
  export-v3-acoustic-weights.py
examples/
  audio_ref/     Reference clips and transcript manifest for voice cloning
docs/
  v3-native-cpp-pipeline.md

C ABI

The C API is defined in src/vieneu/vieneu_tts.h. It is intended for wrappers and host applications that want to embed VieNeu-TTS.cpp without depending on C++ symbols.

Minimal shape:

#include "vieneu/vieneu_tts.h"
#include <stdio.h>

int main(void) {
    struct vieneu_init_params_v2 init;
    vieneu_init_v2_default_params(&init);
    init.profile = "vieneu-v3-onnx";
    init.model_dir = ".models/vieneu-v3-turbo";
    init.onnx_dir = ".models/vieneu-v3-turbo/onnx";
    init.codec_dir = ".models/vieneu-v3-turbo/codec";
    init.voices_json_path = ".models/vieneu-v3-turbo/voices_v3_turbo.json";

    struct vieneu_context *ctx = vieneu_init_v2(&init);
    if (!ctx) {
        fprintf(stderr, "init failed: %s\n", vieneu_last_error());
        return 1;
    }

    struct vieneu_tts_params_v2 tts;
    vieneu_tts_v2_default_params(&tts);
    tts.text = "Xin chào các bạn.";
    tts.temperature = 0.8f;
    tts.top_k = 25;

    struct vieneu_audio audio;
    if (vieneu_synthesize_v2(ctx, &tts, &audio) != 0) {
        fprintf(stderr, "synthesis failed: %s\n", vieneu_last_error());
        vieneu_free(ctx);
        return 1;
    }

    printf("Generated %d samples at %d Hz\n", audio.n_samples, audio.sample_rate);
    vieneu_audio_free(&audio);
    vieneu_free(ctx);
    return 0;
}

Notes

• Keep large model files out of Git. Store them under .models/ or download them from Hugging Face.
• The included .models/, build/, dist/, and outputs/ directories are local working directories.
• For implementation details of the v3 native path, see docs/v3-native-cpp-pipeline.md.

Credits

Special thanks to pnnbao97 and the original pnnbao97/VieNeu-TTS project for the Vietnamese TTS model and reference implementation that this C++ runtime is based on.

Thanks also to the authors and maintainers of llama.cpp, ONNX Runtime, and MOSS Audio Tokenizer.

License

This project is licensed under the MIT License. See LICENSE for details.