VoiceToText 0.2.0

VoiceToText

A C# speech-to-text library with a provider abstraction layer. Supports multiple STT backends through a common interface, designed for reuse across transcribers, Unity games, desktop apps, and more.

Packages

Prerequisites

• .NET 10 SDK (10.0.100 or later)
• A speech model — Whisper models auto-download; Vosk models must be downloaded manually

Usage

Register services

services.AddVoiceToText()
    .AddVoskRecognizer(opts => opts.ModelPath = "models/vosk-model-small-en-us")
    .AddNAudioMicrophone();

Batch transcription

await using var recognizer = serviceProvider.GetRequiredService<ISpeechRecognizer>();
await using var stream = File.OpenRead("audio.wav");
var result = await recognizer.TranscribeAsync(stream);
Console.WriteLine(result.Text);

Real-time streaming

var streaming = serviceProvider.GetRequiredService<IStreamingRecognizer>();
streaming.FinalResultReceived += (_, e) => Console.WriteLine(e.Text);
await streaming.StartAsync();
// Push audio chunks...
await streaming.StopAsync();

Restricting the vocabulary (Vosk)

For command-style input, limit Vosk to the words you expect. Accuracy goes up a lot and recognition gets faster. Set a default on the options, or pass phrases per session, which lets the vocabulary change between push-to-talk presses:

services.AddVoiceToText()
    .AddVoskRecognizer(opts =>
    {
        opts.ModelPath = "models/vosk-model-en-us-0.22-lgraph";
        opts.Grammar = ["roll", "pitch", "yaw", "port", "starboard", "ninety", "degrees"];
    });

await streaming.StartAsync(new RecognizerOptions { Phrases = commandWords });

Speech outside the grammar comes back as [unk] (turn off with AllowUnknownWords = false). The model must support runtime grammars: the small and "lgraph" models do; the big static-graph models do not.

Console Sample

The included console sample supports file transcription and live microphone input.

# Transcribe a WAV file (defaults to Whisper)
dotnet run --project samples/VoiceToText.Samples.Console -- hello-world.wav

# Live microphone with Vosk
dotnet run --project samples/VoiceToText.Samples.Console -- --mic --vosk --model vosk-model-small-en-us-0.15

# Live microphone with Whisper (default)
dotnet run --project samples/VoiceToText.Samples.Console -- --mic

Options:

Project Structure

src/
  VoiceToText/                    Core abstractions (zero STT dependencies)
  VoiceToText.Vosk/               Vosk provider (true streaming)
  VoiceToText.Whisper/            Whisper.net provider (batch, best accuracy)
  VoiceToText.Audio.NAudio/       Windows microphone capture
samples/
  VoiceToText.Samples.Console/    Push-to-talk console demo
tests/
  VoiceToText.Tests/              Unit tests (26 tests)

Architecture

Core Interfaces

• ISpeechRecognizer — batch/file transcription (TranscribeAsync, TranscribeSegmentsAsync)
• IStreamingRecognizer — real-time streaming with PushAudio() + PartialResultReceived/FinalResultReceived events
• IAudioSource — audio input abstraction (microphone, file) with DataAvailable event

Audio Pipeline

All audio is normalized to 16kHz mono 16-bit PCM. The AudioFormatConverter utility handles stereo-to-mono, resampling, and PCM/float conversion.

Target Frameworks

• net10.0 — primary target
• netstandard2.1 — Unity 2021+ and broad compatibility

Building

dotnet build
dotnet test
dotnet pack    # produces 4 .nupkg files

Contributing

Feature branches off main. PRs welcome.

License

MIT