Mantis
Model Description
Mantis is an audio-based emotion recognition model designed for customer service intelligence. It classifies emotional states from speech audio using a HuBERT + CNN hybrid architecture, enabling real-time sentiment monitoring in call center environments.
Model Architecture
- Architecture: HuBERT (feature extractor) + CNN (classifier head)
- Framework: PyTorch
- Task: Audio Emotion Classification
- Input: Raw audio waveforms / mel spectrograms
- Output: Emotion class (e.g., neutral, happy, angry, sad, frustrated)
Training Details
- Dataset: Trained on emotion speech datasets (e.g., RAVDESS, IEMOCAP, or proprietary customer service audio)
- Approach: HuBERT pre-trained representations fed into a custom CNN classifier
- Fine-tuning: End-to-end fine-tuning for customer service emotion categories
Performance
Evaluated on held-out emotion speech samples with strong accuracy across key emotion classes relevant to customer service.
Files
| File | Description |
|---|---|
emotion_model.pth |
Final trained HuBERT-CNN emotion recognition model |
Usage
import torch
from huggingface_hub import hf_hub_download
# Download model
model_path = hf_hub_download(repo_id='devanshty/Mantis', filename='emotion_model.pth')
# Load model (adjust to your model class)
model = torch.load(model_path, map_location='cpu')
model.eval()
# Run inference on audio features
# (preprocess audio to match training pipeline)
Download & Use
from huggingface_hub import hf_hub_download
model_path = hf_hub_download(repo_id='devanshty/Mantis', filename='emotion_model.pth')