LMSupply.Synthesizer

LMSupply.Synthesizer

Local text-to-speech synthesis using VITS/Piper models.

Features

• Zero-config: Models download automatically from HuggingFace
• English and Korean text out of the box: a built-in English phonemizer (pronunciation dictionary + letter-to-sound rules) and a built-in Korean one (the standard pronunciation's rules over the spelling), in the package and offline, write each voice's own phonemes
• The voice's own contract: phoneme ids, inference defaults, sample rate and speaker count are read from the voice's config
• Phoneme input and a phonemizer seam: pass phonemes (SynthesizePhonemesAsync), or plug in the text-to-phoneme step (IPhonemizer)
• GPU Acceleration: CUDA (NVIDIA), CoreML (macOS)
• Licensing stated per voice: Piper's code is MIT; each voice's own license is in SynthesizerModelInfo.License (the default voice is public domain)
• Multiple Languages: English and Korean (text and phonemes), Chinese (phonemes, or text through your phonemizer)

Quick Start

using LMSupply.Synthesizer;

// Load the default model
await using var synthesizer = await LocalSynthesizer.LoadAsync("default");

var result = await synthesizer.SynthesizeAsync("The weather is beautiful today.");

Console.WriteLine($"Duration: {result.DurationSeconds:F2}s");
Console.WriteLine($"Real-time factor: {result.RealTimeFactor:F1}x");

// Save as WAV file
await File.WriteAllBytesAsync("output.wav", result.ToWavBytes());

Text

Text for the English voices is turned into phonemes by the package: numbers, a few common abbreviations and the punctuation that makes pauses are read, each word is pronounced from the CMU Pronouncing Dictionary or, if it is not there, by letter-to-sound rules, and the result is written in the voice's own accent (American for en-us, British-style for the default voice's en and for en-gb-x-rp). A word with two readings gets its most common one. See the synthesizer guide for what is and is not handled.

Text for the Korean voice is read the same way by a built-in Korean phonemizer: figures as Sino-Korean numbers, Latin letters by their names, and each word by the rules of the standard pronunciation that follow from its spelling (see the guide).

For any other voice, set an IPhonemizer in SynthesizerOptions.Phonemizer (it replaces the built-in one for every voice); the model passes it the phonemes its voice declares (ISynthesizerModel.PhonemeScheme: the type, the phonemizer voice and the language).

using LMSupply.Synthesizer;

var options = new SynthesizerOptions { Phonemizer = new MyPhonemizer() };
await using var synthesizer = await LocalSynthesizer.LoadAsync("chinese", options);

var text = File.ReadAllText("input.txt"); // Chinese text
await synthesizer.SynthesizeToFileAsync(text, "output.wav");

sealed class MyPhonemizer : IPhonemizer
{
    public ValueTask<IReadOnlyList<string>> PhonemizeAsync(
        string text, PhonemeScheme scheme, CancellationToken cancellationToken = default)
    {
        // One string per sentence, in the voice's scheme (here "espeak" / "cmn"); every code point is a phoneme,
        // spaces and clause punctuation included.
        IReadOnlyList<string> sentences = [/* your text-to-phoneme step's output */];
        return new(sentences);
    }
}

Without a phonemizer, text for a voice that is neither English, Korean nor of the text phoneme type (the Chinese voice) throws NotSupportedException; it is never spoken as noise.

Available Models

GPU Acceleration

Do not add ONNX Runtime packages (Microsoft.ML.OnnxRuntime*): LMSupply provisions the runtime itself, and a second copy conflicts with it. ExecutionProvider.Auto uses CUDA when the driver supports CUDA 13 and the CUDA 13 runtime and cuDNN 9 are installed on the machine (ONNX Runtime 1.30 is built for CUDA 13), CoreML on macOS, and the CPU otherwise. On Windows with an AMD or Intel GPU, ONNX sessions run on the CPU (DirectML was removed in 0.67.0). See GPU acceleration.