MossTtsSharp

Warning

I am new to the AI field, and this project has not been extensively tested. It is not recommended for production use. If you have any suggestions or encounter issues, please open an Issue.

An unofficial C# binding for MOSS-TTS, a text-to-speech synthesis system powered by ONNX Runtime.

Project Structure

Project	Description
MossTtsSharp	Core inference library — pipeline, tokenizer, audio codec, ONNX models
MossTtsSharp.Backend	Backend abstraction for platform-specific ONNX Runtime
MossTtsSharp.Cpu	CPU backend
MossTtsSharp.Cuda	CUDA backend (NVIDIA GPU)
MossTtsSharp.DirectML	DirectML backend (Windows GPU)
MossTtsSharp.Demo	CLI tool (mosstts syn / mosstts stream)

Quick Start

Prerequisites

• .NET SDK 8.0 or later
• Git (for downloading models)

1. Download Models

./DownloadModel.ps1

This clones the ONNX model repositories into ./MossTtsModel/.

To specify a custom directory:

./DownloadModel.ps1 -TargetDir "D:\Models\MossTTS"

To skip the overwrite confirmation:

./DownloadModel.ps1 -Force

2. Run CLI Demo

dotnet run --project MossTtsSharp.Demo -- syn \
  -t "千里之行，始于足下。" \
  -p ./path/to/prompt.wav \
  -m ./MossTtsModel

Streaming mode (real-time playback):

dotnet run --project MossTtsSharp.Demo -- stream \
  -t "Hello, world!" \
  -p ./path/to/prompt.wav

3. Set Environment Variable (Optional)

Set MOSSTTS_MODELS_DIR to avoid passing -m every time:

$env:MOSSTTS_MODELS_DIR = "D:\Models\MossTTS"

CLI Reference

mosstts syn     -t <text> -p <audio> [-o <file>] [-m <dir>] [-n <value>]
mosstts stream  -t <text> -p <audio> [-o <file>] [-m <dir>] [-n <value>]

Option	Description
-t, --text	Text to synthesize (required)
-p, --prompt	Audio prompt file (.wav, .mp3) (required)
-o, --output	Output WAV file (if omitted, play directly)
-m, --models-dir	Models root directory (or set $MOSSTTS_MODELS_DIR)
-n, --noise	Fixed noise value for deterministic output (omit for random sampling)

API Usage

using MossTtsSharp.Audio;
using MossTtsSharp.Config;
using MossTtsSharp.Cpu;
using MossTtsSharp.Pipeline;

// 1. Create configuration with CPU backend
var config = new MossConfig
{
    Device = new CpuBackend(),
    ModelsRoot = "./MossTtsModel"
};

// 2. Initialize pipeline
await using var pipeline = await MossTtsPipeline.CreateAsync(config);

// 3. Synthesize (batch)
var (waveform, sampleRate) = pipeline.Synthesize("你好，世界！", "prompt.wav");
AudioFile.Write("output.wav", waveform, sampleRate, MossModelConfig.Channels);

// 4. Or stream chunks in real-time
pipeline.SynthesizeStream("Hello, world!", "prompt.wav", audio =>
{
    // audio is float[] — feed to your audio device
});

Choosing a Backend

// CPU (default, works everywhere)
Device = new CpuBackend()

// NVIDIA GPU (Windows / Linux)
Device = new CudaBackend()

// Windows GPU via DirectML
Device = new DirectMlBackend()

Install the corresponding NuGet package alongside MossTtsSharp:

dotnet add package MossTtsSharp.Cpu       # CPU
dotnet add package MossTtsSharp.Cuda      # CUDA
dotnet add package MossTtsSharp.DirectML  # DirectML

License

This project is licensed under the Apache-2.0 License.

The ONNX models are from MOSS-TTS and are also licensed under Apache-2.0.