Document DB v12 - Improved Interceptors with Soft Delete Integration, AI protections, & Admin UI with Aspire Integration! How!?

Getting Started


GitHub
Downloads

Frameworks

.NET

.NET MAUI

Operating Systems

Android

iOS

macOS

Windows

Linux

Shiny.VoiceIntelligence is speaker (voice biometric) enrollment and recognition — the audio twin of Shiny.FaceIntelligence, built to the same architecture so the two read alike. An ECAPA-TDNN / x-vector ONNX model turns a spoken utterance into an L2-normalized “voiceprint”; recognition is the same nearest-neighbour vector lookup with a cosine-distance threshold.

It recognizes who is speaking, not what they said. It is text-independent — the words don’t have to match between enrollment and recognition.

Packages

Package	Role
`Shiny.VoiceIntelligence`	Core. `IVoiceIntelligence`, the contracts (`ISpeakerEmbedder`, `IVoiceStore`), `VoiceEnrollmentSession`, and the registration builder. DI abstractions only — no image stack, no audio capture.
`Shiny.VoiceIntelligence.Onnx`	ONNX speaker embedder (`UseOnnxEmbedder`), handling both waveform and filterbank model inputs. Ships the iOS linker fix.
`Shiny.VoiceIntelligence.DocumentDb`	Provider-agnostic vector store over Shiny.DocumentDb.
`Shiny.VoiceIntelligence.DocumentDb.Sqlite`	Turnkey SQLite + sqlite-vec store.
`Shiny.VoiceIntelligence.Maui`	`VoiceEnrollmentView` — the guided enrollment control — plus the `IVoiceRecorder` seam your app implements.

Install

dotnet add package Shiny.VoiceIntelligence.Onnx
dotnet add package Shiny.VoiceIntelligence.DocumentDb.Sqlite
dotnet add package Shiny.VoiceIntelligence.Maui

Register

using Shiny.VoiceIntelligence;
using Shiny.VoiceIntelligence.Onnx;
using Shiny.VoiceIntelligence.DocumentDb.Sqlite;

builder.Services.AddVoiceIntelligence(voice =>
{
    voice.Options.MaxDistance = 0.4f;   // MUST be measured against your model + capture path

    voice.UseOnnxEmbedder(o =>
    {
        o.ModelBytesProvider = () => LoadBundledModel("ecapa.onnx");
        o.Dimensions = 512;   // MUST match the model: 512 for CAM++/WeSpeaker, 192 for many ECAPA exports
        o.SampleRate = 16000;
    });

    voice.UseSqliteStore(o => o.ConnectionString = $"Data Source={Path.Combine(FileSystem.AppDataDirectory, "voices.db")}");
});

AddVoiceIntelligence validates the result and throws at startup if an embedder or a store is missing, naming the method you forgot.

Or generate the full set of files — packages, registration, manifests and permissions — with the builder below:

Shiny.VoiceIntelligence.DocumentDb.Sqlite

Shiny.VoiceIntelligence.Maui

Shiny.VoiceIntelligence.Onnx

The library never opens a microphone

IVoiceIntelligence takes a float[] sample buffer — mono PCM in [-1, 1] at the embedder’s SampleRate (16 kHz by default). Capturing it — microphone, file, network stream — is your app’s job, exactly as the camera is for face recognition.

That keeps the core usable server-side and testable without a mic, and it means the library never fights an app that already owns its audio pipeline. See Audio Capture — the capture path is the part that most often goes wrong, and it goes wrong in ways no threshold can compensate for.

Use it

public class VoiceCheckIn(IVoiceIntelligence voices)
{
    public Task Enroll(string userId, float[] samples) => voices.Enroll(userId, samples);

    public async Task<string?> WhoIsSpeaking(float[] samples)
    {
        var result = await voices.Recognize(samples);
        return result.IsMatch ? result.PersonIdentifier : null;
    }
}

Member	What it does
`Enroll(id, samples)`	Embed the utterance and store the voiceprint under `id`.
`CreateEnrollment(id, options?)`	Start a guided enrollment session that checks each clip and decides when it has enough.
`Recognize(samples)`	Nearest enrolled identity within the threshold, or `RecognitionResult.NoMatch`.
`GetAll()`	Every stored utterance, most recent first.
`Forget(id)`	Delete every utterance for an identity.

Prefer guided enrollment

Enroll stores whatever you hand it, including a clip that was mostly silence or picked up the wrong person. That clip then becomes a template every future match is compared against.

var session = voices.CreateEnrollment("user-118");
while (!session.IsComplete)
{
    Show(session.CurrentPrompt);
    var step = await session.Submit(await recorder.RecordAsync(TimeSpan.FromSeconds(5)));
    Show(step.Hint);           // "" when accepted
}
var result = session.Result!; // already stored; result.Cohesion is the quality number

See Guided Enrollment, or drop VoiceEnrollmentView into a MAUI page and let it drive the whole loop.

Where next

Guided Enrollment — the session, its gates, and why they’re inverted vs face
Recognition & Tuning — thresholds, measurement, anti-spoofing
Audio Capture — the settings that quietly destroy speaker matching
MAUI Control — VoiceEnrollmentView and IVoiceRecorder
ONNX Models — CAM++/WeSpeaker vs ECAPA, input modes, dimensions