You can deploy this model and start earning money today!

Booberjacob/
wav2vec2-lg-xlsr-en-speech-circumplex-emotion-recognition

Name: wav2vec2-lg-xlsr-en-speech-circumplex-emotion-recognition
Availability: InStock

safetensors

wav2vec2

apache-2.0

Views

No views yet

This model's structure was modified from ehcalabres/wav2vec2-lg-xlsr-en-speech-emotion-recognition for a Speech Emotion Recognition (SER) task.

The model was reinitialized and trained to predict sentiment and arousal, ranging from -1 to 1, using the following datasets:

The emotions were translated using the emotions.npz file.

To prevent bias from imbalanced datasets, I used a sampling strategy that draws evenly from each class during training. This involved potentially oversampling smaller classes and undersampling larger ones, ensuring the model learns effectively from all classes despite differences in their original representation.

Example Code

from transformers import AutoModelForAudioClassification, Wav2Vec2FeatureExtractor  
import librosa  
  
model_path = "Booberjacob/wav2vec2-lg-xlsr-en-speech-circumplex-emotion-recognition"  
model = AutoModelForAudioClassification.from_pretrained(model_path)  
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("facebook/wav2vec2-large-xlsr-53")  
  
  
wav_path = "sample.wav"  
wav, _ = librosa.load(wav_path, sr=16000)  
inputs = feature_extractor(raw_speech=wav, sampling_rate=16000, padding=True, return_tensors='pt')  
outputs = model(**inputs)  
sentiment, arousal = outputs['logits'][0]  
print(sentiment, arousal)