aicinema69/audio-emotion-detector-try2

🎧 Audio Emotion Detector - Finetuned AST Model

This is a fine-tuned version of MIT/ast-finetuned-audioset-14-14-0.443 for audio emotion classification using a custom labeled dataset of voice samples tagged with 9 emotion classes. The model was trained using Hugging Face Transformers with QLoRA and 4-bit quantization (bitsandbytes) for efficient finetuning on a limited compute setup [[1]].

---

📝 Model Details

• Base Model: MIT/ast-finetuned-audioset-14-14-0.443
• Task: Audio Emotion Classification
• Fine-tuned Model Name: audio-emotion-detector
• Labels:
  • angry (0)
  • apologetic (1)
  • base (2)
  • calm (3)
  • excited (4)
  • fear (5)
  • happy (6)
  • sad (7)
  • surprise (8)
• Quantization: 4-bit (nf4) via bitsandbytes
• Training Framework: Hugging Face Transformers, Trainer API
• Optimization: QLoRA with LoRA rank 64, α=16, dropout=0.1
• Compute: Single GPU with mixed precision disabled (fp16=False)

---

📊 Training Results (2 Epochs)

Step	Training Loss	Validation Loss
25	2.3754	1.9816
200	0.7936	0.5268
375	0.2698	0.3061
450	0.0878	0.1847
550	0.1707	0.1864

• Best Validation Loss: 0.1625
• Metric for Model Selection: Validation Loss (best model loaded at end)

---

📂 Dataset

• Dataset: Emotion TTS Dataset
• Size: Split 80% train / 20% validation
• Sampling Rate: 16kHz
• Max Audio Length: 10 seconds
• Classes: 9 audio-based emotions

---

🛠️ How to Use

from transformers import AutoModelForAudioClassification, AutoFeatureExtractor
import torchaudio

# Load model and feature extractor
model = AutoModelForAudioClassification.from_pretrained("aicinema69/audio-emotion-detector-try2")
feature_extractor = AutoFeatureExtractor.from_pretrained("aicinema69/audio-emotion-detector-try2")

# Load audio (16kHz WAV)
audio_input, sample_rate = torchaudio.load("your_audio_file.wav")

# Preprocess audio
inputs = feature_extractor(audio_input.squeeze().numpy(), sampling_rate=16000, return_tensors="pt")

# Predict
logits = model(**inputs).logits
predicted_class = logits.argmax(-1).item()

# Label mapping
id2label = model.config.id2label
print(f"Predicted Emotion: {id2label[predicted_class]}")
'''



@misc{aicinema69_audio_emotion_detector,
  title={Audio Emotion Detector},
  author={Satyam Singh},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/aicinema69/audio-emotion-detector-try2}}
}