feat: Add speaker diarization functionality to Qwen3-ForcedAligner

[YuYun329]

Summary

This PR adds speaker diarization functionality to Qwen3-ForcedAligner, enabling speaker identification alongside timestamp prediction.

Key Changes

1. Speaker Diarization Support

• Added speaker field to ForcedAlignItem dataclass to store speaker labels
• Integrated CAM++ model for speaker embedding extraction
• Implemented clustering-based speaker diarization using ClusterBackend

2. New Components

• cluster_backend.py: New module for speaker clustering with spectral clustering algorithm which refered to FunASR;
• CAM++ model integration: Automatic model download from ModelScope or HuggingFace

3. Enhanced Features

• Qwen3ForcedAligner now accepts optional campplus_model parameter
• Automatic model resolution from local path, ModelScope, or HuggingFace
• load_wav() utility for audio preprocessing
• Speaker embedding extraction and clustering pipeline

4. Files Modified

File	Changes
qwen_asr/inference/qwen3_forced_aligner.py	+226 lines - Core speaker diarization logic
qwen_asr/inference/cluster_backend.py	+192 lines - New clustering module
qwen_asr/cli/demo.py	+20 lines - Demo updates
README.md	Documentation updates
pyproject.toml	Dependency updates

Usage Example

python qwen_asr/cli/demo.py --asr-checkpoint Qwen/Qwen3-ASR-1.7B --aligner-checkpoint Qwen/Qwen3-ForcedAligner-0.6B --campplus-model FunAudioLLM/Fun-CosyVoice3-0.5B-2512/campplus.onnx

Dependencies

• onnxruntime - For CAM++ model inference
• scikit-learn - For spectral clustering

Testing

Tested with multi-speaker audio files, successfully identifying and labeling different speakers with accurate timestamps.

---

This feature enhances Qwen3-ASR's capabilities for applications like:

• Meeting transcription with speaker attribution
• Interview analysis
• Podcast/Media content processing
• Call center analytics