[Task] speech-separation model support #289

Overview

Speech separation models (PixIT and similar) disentangle overlapping speakers from a single-channel mixture into per-speaker audio streams. pyannote/speech-separation-ami-1.0 jointly performs diarization and separation, outputting both speaker segments and cleaned individual audio.

Agent Scenarios

• Multi-speaker transcription agent: cleanly transcribe conference calls or panel discussions where speakers frequently talk over each other
• Hearing assistance agent: enhance a target speaker's voice in a noisy or crowded environment (cocktail-party problem)
• Meeting replay agent: let a user listen to a specific participant's audio track isolated from a recorded meeting
• Broadcast post-production agent: separate host and guest audio in a recorded interview for independent mixing or noise reduction

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/speech-separation-ami-1.0