ONNX Models

You supply the speaker embedding model — an ECAPA-TDNN, x-vector or CAM++ export in ONNX form. Nothing is bundled.

voice.UseOnnxEmbedder(o =>
{
    o.ModelBytesProvider = () => LoadBundledModel("ecapa.onnx");
    o.Dimensions = 512;                          // must match the model's real output width
    o.SampleRate = 16000;
    // o.InputMode = OnnxSpeakerInputMode.Auto;  // Auto | Waveform | Fbank80
});

OnnxEmbedderOptions accepts three model sources in priority order: ModelBytesProvider → ModelBytes → ModelPath. On iOS and Android, bundled assets have no real filesystem path — use the bytes provider. The model loads lazily on first enroll/recognize, so a missing model surfaces there as a FileNotFoundException rather than breaking startup.

Waveform or filterbank — detected from the model

Two model families are common, and they take completely different inputs. OnnxEcapaEmbedder reads the model’s declared input rank when the session loads and picks between them:

Model input	Handling
[batch, samples] (rank 2)	Raw mono waveform, fed straight through
[batch, frames, 80] (rank 3, usually named feats)	80-bin Kaldi filterbank computed first

Set InputMode explicitly if a model’s declared shape lies. A rank-3 model wanting anything other than 80 bins throws NotSupportedException — implement ISpeakerEmbedder yourself and register it with UseEmbedder(...).

The filterbank front end has no cepstral mean normalization, deliberately

KaldiFbank implements kaldi-compatible features with snip_edges=false. Adding CMN makes matching worse against these exports — don’t “fix” it. It’s covered by tests for bin count, frame count, determinism and level dependence, the last of which fails if someone adds CMN.

Dimensions must match the model

This one bites

Dimensions sizes the vector store before the model loads, so it cannot be inferred. Common values:

• 512 — CAM++ / WeSpeaker exports
• 192 — many ECAPA-TDNN exports (this is the option default)

A mismatch throws on the first embed rather than silently corrupting stored voiceprints, but you still have to get it right.

Probe an unfamiliar model before wiring it up:

using var session = new InferenceSession(modelBytes);
foreach (var i in session.InputMetadata)
    Console.WriteLine($"in  {i.Key}: [{string.Join(",", i.Value.Dimensions)}]");
foreach (var o in session.OutputMetadata)
    Console.WriteLine($"out {o.Key}: [{string.Join(",", o.Value.Dimensions)}]");

foreach (var kv in session.ModelMetadata.CustomMetadataMap)
    Console.WriteLine($"{kv.Key} = {kv.Value}");   // WeSpeaker/sherpa-onnx carry output_dim, sample_rate, …

WeSpeaker and sherpa-onnx exports carry framework, output_dim, sample_rate and normalize_samples in their custom metadata map, which answers most of these questions directly.

Sample rate

SampleRate (default 16000) is the rate the embedder expects, and it is what IVoiceIntelligence documents its float[] buffers as being in. Resampling to it is your capture code’s job — see Audio Capture, including why a naive resampler silently ruins matching.

App size and the iOS linker fix

The ONNX Runtime floor (~33 MB iOS arm64, ~17 MB Android arm64-v8a per architecture) and the reduced-build options are the same as for face — see Face Intelligence · Models. Speaker models are small by comparison (typically single-digit MB).

Shiny.VoiceIntelligence.Onnx ships the same automatic iOS / Mac Catalyst linker fix for ONNX Runtime’s _RegisterCustomOps symbol, with its MSBuild target name suffixed _Voice so an app referencing both the face and voice ONNX packages doesn’t hit a duplicate-target-name error. Nothing to configure either way.