wav2vec2-emotion-recognition-ONNX

wav2vec2-emotion-recognition (ONNX)#

This is an ONNX version of Dpngtm/wav2vec2-emotion-recognition. It was automatically converted and uploaded using this Hugging Face Space.

Usage with Transformers.js#

See the pipeline documentation for audio-classification: https://huggingface.co/docs/transformers.js/api/pipelines#module_pipelines.AudioClassificationPipeline

wav2vec2-emotion-recognition (Robust Fine-Tuning)#

This model is a fine-tuned version of facebook/wav2vec2-base-960h for Speech Emotion Recognition (SER).

It has been trained using a Frozen Feature Extractor strategy to preserve the model's acoustic understanding while adapting to emotion detection. This approach ensures stable performance and prevents "Catastrophic Forgetting," achieving nearly 80% accuracy on the validation set.

Update: The "Calm" and "Neutral" classes have been merged to improve classification consistency, resulting in 7 distinct emotion classes.

Model Description#

• Model Architecture: Wav2Vec2 with a frozen CNN feature extractor and a trainable sequence classification head.
• Language: English
• Task: Speech Emotion Recognition (SER)
• Fine-tuned from: facebook/wav2vec2-base-960h
• Class Merging: The "Calm" emotion has been merged into "Neutral" to reduce label ambiguity.

Datasets#

The model was trained on a combined dataset of ~12,000 audio files from:

• TESS
• CREMA-D
• SAVEE
• RAVDESS

Performance Metrics#

Note: Achieved stable convergence using a Cosine Learning Rate Scheduler and Frozen Feature Extractor.

Supported Emotions (7 Classes)#

The model classifies audio into one of the following emotions:

1. 😠 Angry
2. 🤢 Disgust
3. 😨 Fear
4. 😊 Happy
5. 😐 Neutral (includes Calm)
6. 😢 Sad
7. 😲 Surprise

Training Configuration#

The model was fine-tuned using the following "Robust" configuration to ensure stability on Google Colab T4 GPUs:

• Feature Extractor: FROZEN (Locked weights to preserve pre-trained acoustic features)
• Epochs: 10
• Learning Rate: 3e-5
• Scheduler: Cosine Decay (Smooth landing)
• Batch Size: 4 (Physical) / 16 (Effective via Gradient Accumulation)
• Optimizer: AdamW
• Precision: fp32 (Standard precision to prevent gradient underflow)

Limitations#

Audio Requirements:#

• Sampling Rate: 16kHz (Model will hallucinate if input is not 16kHz. Resampling is required.)
• Duration: < 10 seconds recommended (Longer files should be chunked).
• Environment: Best results with clear speech; background noise may reduce confidence.

Demo#

Try the model in your browser: 👉 Hugging Face Space: Audio Emotion Recognition

FINETUNING NOTEBOOOKS#

Notebook

Contact#

• GitHub: DGautam11
• LinkedIn: Deepan Gautam
• Hugging Face: @Dpngtm

For issues and questions, feel free to:

1. Open an issue on the Model Repository
2. Comment on the Demo Space

Usage#

from transformers import AutoModelForAudioClassification, Wav2Vec2Processor
import torch
import torchaudio
import numpy as np

# 1. Load Model and Processor
model_id = "Dpngtm/wav2vec2-emotion-recognition"
model = AutoModelForAudioClassification.from_pretrained(model_id)
processor = Wav2Vec2Processor.from_pretrained(model_id)

# 2. Load Audio
# Replace with your file path
audio_path = "path_to_audio.wav" 
speech_array, sampling_rate = torchaudio.load(audio_path)

# 3. Resample to 16kHz (CRITICAL STEP)
if sampling_rate != 16000:
    resampler = torchaudio.transforms.Resample(sampling_rate, 16000)
    speech_array = resampler(speech_array)
    sampling_rate = 16000

# 4. Handle Stereo (Convert to Mono)
if speech_array.shape[0] > 1:
    speech_array = torch.mean(speech_array, dim=0, keepdim=True)

# 5. Process & Predict
inputs = processor(speech_array.squeeze(), sampling_rate=16000, return_tensors="pt", padding=True)

with torch.no_grad():
    logits = model(**inputs).logits

# 6. Decode Result
predicted_id = torch.argmax(logits, dim=-1).item()
id2label = model.config.id2label
predicted_label = id2label[predicted_id]

print(f"Predicted Emotion: {predicted_label}")