dynann/emotion-speech-recognition

Wav2Vec2 for Emotion Recognition

This model is a fine-tuned version of facebook/wav2vec2-base on the Toronto Emotional Speech Dataset (TESS). It achieves the following results on the evaluation set:

• Accuracy: 85%
• Loss: ~3.76

Model Description

The model classifies audio input into 7 discrete emotions:

• Angry
• Disgust
• Fear
• Happy
• Neutral
• Pleasant Surprise (ps)
• Sad

It uses a custom classification head on top of the frozen Wav2Vec2 base model.

Usage

Note: You must define the custom Wav2Vec2ForEmotionClassification class to load this model.

import torch
import torch.nn as nn
from transformers import Wav2Vec2Processor, Wav2Vec2Model, Wav2Vec2Config

# Define the Custom Model Class
class Wav2Vec2ForEmotionClassification(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.wav2vec2 = Wav2Vec2Model(config)
        self.classifier = nn.Sequential(
            nn.Linear(config.hidden_size, 256),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(256, 128),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(128, config.num_labels),
        )

    def forward(self, input_values, attention_mask=None, labels=None, **kwargs):
        outputs = self.wav2vec2(input_values, attention_mask=attention_mask)
        hidden_states = outputs.last_hidden_state
        pooled_output = torch.mean(hidden_states, dim=1)
        logits = self.classifier(pooled_output)
        
        loss = None
        if labels is not None:
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(logits, labels.view(-1))
            
        return {
            "loss": loss, 
            "logits": logits
        }

# Load Model
model_id = "dynann/emotion-speech-recognition"
config = Wav2Vec2Config.from_pretrained(model_id)
model = Wav2Vec2ForEmotionClassification(config)
model.load_state_dict(torch.hub.load_state_dict_from_url(f"https://huggingface.co/{model_id}/resolve/main/pytorch_model.bin"))

processor = Wav2Vec2Processor.from_pretrained(model_id)

Training Procedure

• Epochs: 10
• Batch Size: 32 (optimized for P100) / 8 (local)
• Learning Rate: 3e-4
• Feature Encoder: Frozen