The model has been trained on the ShEMO dataset, which includes annotated emotional speech in Persian.
It leverages the ECAPA-TDNN architecture, commonly used in speaker recognition and emotion classification tasks.

Model Overview

Architecture: ECAPA-TDNN (Emphasized Channel Attention, Propagation and Aggregation in TDNN)
Features: 80‑dim Mel‑filterbank, input normalization
Loss: Additive Angular Margin (AAM‑Softmax)
Classes: 6 emotions – anger, surprise, happiness, sadness, neutral, fear
Trained on: ShEMO dataset (≈3 k Persian utterances, 87 speakers)

Installation

pip install torch torchaudio speechbrain

(Tested with SpeechBrain ≥1.0 and torch ≥2.0.)

Quickstart Inference

python

from speechbrain.pretrained import EncoderClassifier

classifier = EncoderClassifier.from_hparams(
    source="mobina1380/speechbrain-persian-ser",
    savedir="tmp_cache",          
)

# Classify a WAV file
emotion = classifier.classify_file("your_audio.wav")
print("Predicted emotion:", emotion)

No manual Git‑LFS, no snapshot_download calls—just one import and one from_hparams.

File Structure

README.md                  
hyperparams.yaml           
embedding_model.ckpt       
classifier.ckpt            
normalizer.ckpt            
label_encoder.txt

Limitations

📉 Not highly accurate for subtle or noisy emotions