Sammaiah/audio-emotion-pipeline-v22

Audio Emotion Pipeline (v22 & v22.2 SOTA)

SOTA multi-modal audio emotion recognition pipeline:

1. Speaker Diarization (pyannote/speaker-diarization-3.1)
2. Text Feature Extraction (microsoft/deberta-v3-large 1024d)
3. Acoustic SER Extraction (iic/emotion2vec_plus_large 1024d @ 50fps)
4. Prosody Extraction (openSMILE 25d eGeMAPSv02 LLD + Speaker Z-norm)
5. Classification (Dynamic Sigmoid Gated Fusion + Self-Attention Pooling)

🚀 Quickstart

from emotion_pipeline.pipeline import EmotionPipeline
import json

# Initialize for v22.2 Conversation Level SOTA (or model_version="v22" for Single Utterance)
pipeline = EmotionPipeline(hf_token="your_hf_token", model_version="v22.2", device="cuda")
result = pipeline.analyze("meeting.wav", max_past=3, max_future=3)
print(json.dumps(result, indent=2))