wav2vec-english-speech-emotion-recognition

Speech Emotion Recognition By Fine-Tuning Wav2Vec 2.0

The model is a fine-tuned version of jonatasgrosman/wav2vec2-large-xlsr-53-english for a Speech Emotion Recognition (SER) task.

Several datasets were used the fine-tune the original model:

• Surrey Audio-Visual Expressed Emotion (SAVEE) - 480 audio files from 4 male actors
• Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) - 1440 audio files from 24 professional actors (12 female, 12 male)
• Toronto emotional speech set (TESS) - 2800 audio files from 2 female actors

7 labels/emotions were used as classification labels

emotions = ['angry' 'disgust' 'fear' 'happy' 'neutral' 'sad' 'surprise']

It achieves the following results on the evaluation set:

• Loss: 0.104075
• Accuracy: 0.97463

Model Usage

pip install transformers librosa torch

from transformers import *
import librosa
import torch

feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("r-f/wav2vec-english-speech-emotion-recognition")
model = Wav2Vec2ForCTC.from_pretrained("r-f/wav2vec-english-speech-emotion-recognition")

def predict_emotion(audio_path):
    audio, rate = librosa.load(audio_path, sr=16000)
    inputs = feature_extractor(audio, sampling_rate=rate, return_tensors="pt", padding=True)
    
    with torch.no_grad():
        outputs = model(inputs.input_values)
        predictions = torch.nn.functional.softmax(outputs.logits.mean(dim=1), dim=-1)  # Average over sequence length
        predicted_label = torch.argmax(predictions, dim=-1)
        emotion = model.config.id2label[predicted_label.item()]
    return emotion

emotion = predict_emotion("example_audio.wav")
print(f"Predicted emotion: {emotion}")
>> Predicted emotion: angry

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 4
• eval_batch_size: 4
• eval_steps: 500
• seed: 42
• gradient_accumulation_steps: 2
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• num_epochs: 4
• max_steps=7500
• save_steps: 1500

Training results