Hubert Large Speech Emotion Recognition Russian Dusha Finetuned

🚀 HuBERT fine-tuned on DUSHA dataset for speech emotion recognition in Russian language

This project fine-tunes the HuBERT model on the DUSHA dataset to achieve speech emotion recognition in Russian, improving accuracy and F1 score compared to the baseline.

🚀 Quick Start

The pre-trained model is facebook/hubert-large-ls960-ft. The DUSHA dataset used can be found here.

✨ Features

• Fine-tuning Environment: Fine-tuned in Google Colab using a Pro account with an A100 GPU.
• Layer Freezing: All layers were frozen except the projector, classifier, and all 24 HubertEncoderLayerStableLayerNorm layers.
• Dataset Usage: Half of the training dataset was used.

📦 Installation

There is no specific installation content provided in the original document, so this section is skipped.

💻 Usage Examples

Basic Usage

from transformers import HubertForSequenceClassification, Wav2Vec2FeatureExtractor
import torchaudio
import torch

feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("facebook/hubert-large-ls960-ft")
model = HubertForSequenceClassification.from_pretrained("xbgoose/hubert-speech-emotion-recognition-russian-dusha-finetuned")
num2emotion = {0: 'neutral', 1: 'angry', 2: 'positive', 3: 'sad', 4: 'other'}

filepath = "path/to/audio.wav"

waveform, sample_rate = torchaudio.load(filepath, normalize=True)
transform = torchaudio.transforms.Resample(sample_rate, 16000)
waveform = transform(waveform)

inputs = feature_extractor(
        waveform, 
        sampling_rate=feature_extractor.sampling_rate, 
        return_tensors="pt",
        padding=True,
        max_length=16000 * 10,
        truncation=True
    )

logits = model(inputs['input_values'][0]).logits
predictions = torch.argmax(logits, dim=-1)
predicted_emotion = num2emotion[predictions.numpy()[0]]
print(predicted_emotion)

📚 Documentation

Training Parameters

• 2 epochs
• Train batch size = 8
• Eval batch size = 8
• Gradient accumulation steps = 4
• Learning rate = 5e-5 without warm up and decay

Metrics

Achieved

• Accuracy = 0.86
• Balanced = 0.76
• Macro F1 score = 0.81 on the test set, improving accuracy and F1 score compared to the dataset baseline.

🔧 Technical Details

There is no specific technical details content that meets the requirements in the original document, so this section is skipped.

📄 License

The project is licensed under the apache-2.0 license.