wav2vec2-xls-r-300m-emotion-ru

Speech Emotion Recognition

The model is a fine-tuned version of facebook/wav2vec2-xls-r-300m for a Speech Emotion Recognition (SER) task.

The dataset used to fine-tune the original pre-trained model is the DUSHA dataset. The dataset consists of about 125 000 audio recordings in Russian with four basic emotions that usually appear in a dialog with a virtual assistant: Happiness (Positive), Sadness, Anger and Neutral emotion.

emotions = ['neutral', 'positive', 'angry', 'sad', 'other']

How to use

Pipeline

from transformers.pipelines import pipeline

pipe = pipeline(model="KELONMYOSA/wav2vec2-xls-r-300m-emotion-ru", trust_remote_code=True)

# The pipeline input can be a file, path or link
result = pipe("speech.wav")
print(result)

[{'label': 'neutral', 'score': 0.00318}, {'label': 'positive', 'score': 0.00376}, {'label': 'sad', 'score': 0.00145}, {'label': 'angry', 'score': 0.98984}, {'label': 'other', 'score': 0.00176}]

AutoModel

import librosa
import torch
import torch.nn.functional as F
from transformers import AutoConfig, Wav2Vec2Processor, AutoModelForAudioClassification

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_name_or_path = "KELONMYOSA/wav2vec2-xls-r-300m-emotion-ru"
config = AutoConfig.from_pretrained(model_name_or_path)
processor = Wav2Vec2Processor.from_pretrained(model_name_or_path)
sampling_rate = processor.feature_extractor.sampling_rate
model = AutoModelForAudioClassification.from_pretrained(model_name_or_path, trust_remote_code=True).to(device)


def predict(path):
    speech, sr = librosa.load(path, sr=sampling_rate)
    features = processor(speech, sampling_rate=sampling_rate, return_tensors="pt", padding=True)

    input_values = features.input_values.to(device)
    attention_mask = features.attention_mask.to(device)

    with torch.no_grad():
        logits = model(input_values, attention_mask=attention_mask).logits

    scores = F.softmax(logits, dim=1).detach().cpu().numpy()[0]
    outputs = [{"label": config.id2label[i], "score": round(score, 5)} for i, score in
               enumerate(scores)]
    return outputs


print(predict("speech.wav"))

[{'label': 'neutral', 'score': 0.00318}, {'label': 'positive', 'score': 0.00376}, {'label': 'sad', 'score': 0.00145}, {'label': 'angry', 'score': 0.98984}, {'label': 'other', 'score': 0.00176}]

Evaluation

It achieves the following results:

• Training Loss: 0.528700
• Validation Loss: 0.349617
• Accuracy: 0.901369