feat: add CAM++ speaker diarization with sentence-level speaker labels

[yanhuicsdn]

Summary

Adds a speaker diarization module that uses FunASR CAM++ (iic/speech_campplus_sv_zh-cn_16k-common) to assign speaker labels to transcribed segments. Works at the sentence level by splitting VAD segments into small sliding windows (3s / 1.5s shift) before clustering.

Changes

New: core/diarization.py

• SpeakerDiarizer class with two-phase embedding extraction:
  • Phase 1: Split VAD segments into short (3s) overlapping windows
  • Phase 2: Extract CAM++ speaker embeddings per window, then cluster via ClusterBackend
• Majority voting assigns each original VAD segment the label of its majority sub-windows
• Graceful fallback: if funasr is not installed, silently degrades to single-speaker output
• Configurable: merge_thr, window_s, shift_s

Modified: core/auto_model.py

• New AutoModel.transcribe_with_diarization() method that returns:
  • "text": full concatenated transcript with [讲话人N] markers
  • "segments": list of per-segment dicts with start_s, end_s, text, speaker
  • "num_speakers": number of distinct speakers found
• Lazy-initialised SpeakerDiarizer instance; failures caught gracefully

Usage

from core.auto_model import AutoModel

model = AutoModel(model=..., vad_model=...)
result = model.transcribe_with_diarization("meeting.wav")

print(f"Detected {result[\"num_speakers\"]} speakers")
for seg in result["segments"]:
    print(f"[Speaker {seg[\"speaker\"]}] {seg[\"text\"]}")

Dependencies (optional)

• funasr (for CAM++ model)
• soundfile (for audio I/O)

Performance

On a 7-minute podcast (8 VAD segments → 108 sub-windows):

• Transcription: ~50s (MLX)
• Diarization overhead: ~15s (CAM++ sliding window)
• Total: ~65s vs ~50s without diarization (+30%)

The overhead is proportional to the number of sub-windows, not audio duration.