KELONMYOSA/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

emotion	precision	recall	f1-score	support
neutral	0.92	0.94	0.93	15886
positive	0.85	0.79	0.82	2481
sad	0.77	0.82	0.79	2506
angry	0.89	0.83	0.86	3072
other	0.99	0.74	0.85	226
accuracy			0.90	24171
macro avg	0.89	0.82	0.85	24171
weighted avg	0.90	0.90	0.90	24171