[Task] speaker-embedding model support #288

Overview

Speaker embedding models encode a variable-length speech segment into a fixed-size vector that captures the speaker's vocal identity. These embeddings are the backbone of speaker verification (is this the claimed speaker?) and identification (which speaker is this?).

Agent Scenarios

• Voice authentication agent: verify a user's identity by comparing a live utterance embedding against a stored voiceprint before granting access
• Personalization agent: adapt a voice assistant's responses or TTS voice profile based on recognized speaker identity
• Fraud detection agent: flag calls where the speaker embedding diverges from the account holder's enrolled voiceprint
• Speaker-aware RAG agent: retrieve documents personalized to a known speaker's history or preferences, identified from an audio query

ModelKit Integration

Models must pass the full wmk pipeline on all EPs:

wmk config → wmk build (ONNX export) → wmk perf → wmk eval

Acceptance Criteria

• pyannote/wespeaker-voxceleb-resnet34-LM
• pyannote/embedding