Voice Intelligence Releases

1.0 TBD

Feature

Initial release — on-device speaker (voice biometric) enrollment and recognition, built to the same architecture as Shiny.FaceIntelligence: a dependency-free core plus swappable embedder and store packages composed with AddVoiceIntelligence.

Feature

IVoiceIntelligence — Enroll / Recognize over mono PCM sample buffers, CreateEnrollment, GetAll and Forget, keyed on a caller-chosen PersonIdentifier. The library never opens a microphone; capture stays with the app.

Feature

VoiceEnrollmentSession — guided enrollment that measures each recording (duration, speech level, clipping, SNR) and keeps going until the voiceprints agree, with distance gates derived from the match threshold, bad-first-clip recovery, and best-subset storage when it runs out of attempts.

Feature

Shiny.VoiceIntelligence.Onnx — speaker embedder supporting both raw-waveform and 80-bin filterbank model inputs, auto-detected from the model’s declared input rank, with a kaldi-compatible filterbank front end.

Feature

Shiny.VoiceIntelligence.DocumentDb / .DocumentDb.Sqlite — vector storage and nearest-neighbour search over Shiny.DocumentDb, with the vector dimension read from the embedder.

Feature

Shiny.VoiceIntelligence.Maui — VoiceEnrollmentView, a guided enrollment control with prompt list, countdown, dB-scaled VU meter marking the acceptance threshold, and an attempt ceiling that still stores the best usable subset. Records through the app-supplied IVoiceRecorder seam.

Feature

AOT- and trim-compatible across every package, with a source-generated JsonSerializerContext for the stored document type.

Feature

Automatic iOS / Mac Catalyst linker fix for ONNX Runtime’s _RegisterCustomOps symbol, named so it coexists with the Face Intelligence equivalent in the same app.