speech-emotion-recognition-with-openai-whisper-large-v3

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

🎧 Speech Emotion Recognition with Whisper

This project leverages the Whisper model to recognize emotions in speech. The goal is to classify audio recordings into different emotional categories, such as Happy, Sad, Surprised, and etc.

🗂 Dataset

The dataset used for training and evaluation is sourced from multiple datasets, including:

• RAVDESS
• SAVEE
• TESS
• URDU

The dataset contains recordings labeled with various emotions. Below is the distribution of the emotions in the dataset:

This distribution reflects the balance of emotions in the dataset, with some emotions having more samples than others. Excluded the "calm" emotion during training due to its underrepresentation.

🎤 Preprocessing

• Audio Loading: Using Librosa to load the audio files and convert them to numpy arrays.
• Feature Extraction: The audio data is processed using the Whisper Feature Extractor, which standardizes and normalizes the audio features for input to the model.

🔧 Model

The model used is the Whisper Large V3 model, fine-tuned for audio classification tasks:

• Model: openai/whisper-large-v3
• Output: Emotion labels (Angry', 'Disgust', 'Fearful', 'Happy', 'Neutral', 'Sad', 'Surprised')

I map the emotion labels to numeric IDs and use them for model training and evaluation.

⚙️ Training

The model is trained with the following parameters:

• Learning Rate: 5e-05
• Train Batch Size: 2
• Eval Batch Size: 2
• Random Seed: 42
• Gradient Accumulation Steps: 5
• Total Train Batch Size: 10 (effective batch size after gradient accumulation)
• Optimizer: Adam with parameters: betas=(0.9, 0.999) and epsilon=1e-08
• Learning Rate Scheduler: linear
• Warmup Ratio for LR Scheduler: 0.1
• Number of Epochs: 25
• Mixed Precision Training: Native AMP (Automatic Mixed Precision)

These parameters ensure efficient model training and stability, especially when dealing with large datasets and deep models like Whisper. The training utilizes Wandb for experiment tracking and monitoring.

📊 Metrics

The following evaluation metrics were obtained after training the model:

• Loss: 0.5008
• Accuracy: 0.9199
• Precision: 0.9230
• Recall: 0.9199
• F1 Score: 0.9198

These metrics demonstrate the model's performance on the speech emotion recognition task. The high values for accuracy, precision, recall, and F1 score indicate that the model is effectively identifying emotional states from speech data.

🧪 Results

After training, the model is evaluated on the test dataset, and the results are monitored using Wandb in this Link.

🚀 How to Use

# Requires: librosa
from transformers import AutoModelForAudioClassification, AutoFeatureExtractor
import librosa
import torch
import numpy as np

model_id = "firdhokk/speech-emotion-recognition-with-openai-whisper-large-v3"
model = AutoModelForAudioClassification.from_pretrained(model_id)

feature_extractor = AutoFeatureExtractor.from_pretrained(model_id, do_normalize=True)
id2label = model.config.id2label

def preprocess_audio(audio_path, feature_extractor, max_duration=30.0):
    audio_array, sampling_rate = librosa.load(audio_path, sr=None)
    
    max_length = int(feature_extractor.sampling_rate * max_duration)
    if len(audio_array) > max_length:
        audio_array = audio_array[:max_length]
    else:
        audio_array = np.pad(audio_array, (0, max_length - len(audio_array)))

    inputs = feature_extractor(
        audio_array,
        sampling_rate=feature_extractor.sampling_rate,
        max_length=max_length,
        truncation=True,
        return_tensors="pt",
    )
    return inputs

def predict_emotion(audio_path, model, feature_extractor, id2label, max_duration=30.0):
    inputs = preprocess_audio(audio_path, feature_extractor, max_duration)
    
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = model.to(device)
    inputs = {key: value.to(device) for key, value in inputs.items()}

    with torch.no_grad():
        outputs = model(**inputs)

    logits = outputs.logits
    predicted_id = torch.argmax(logits, dim=-1).item()
    predicted_label = id2label[predicted_id]
    
    return predicted_label

audio_path = "/content/drive/MyDrive/Audio/Speech_URDU/Happy/SM5_F4_H058.wav"

predicted_emotion = predict_emotion(audio_path, model, feature_extractor, id2label)
print(f"Predicted Emotion: {predicted_emotion}")

🎯 Framework versions

• Transformers 4.44.2
• Pytorch 2.4.1+cu121
• Datasets 3.0.0
• Tokenizers 0.19.1