Whisper-DeepStack Tunisian Arabic Intent Classifier

This model is a small classifier head (~0.8M params) trained on top of frozen Whisper-large-v3 encoder features for Tunisian Arabic intent recognition as part of the NADI 2026 Shared Task.

Architecture

• Encoder: openai/whisper-large-v3 (frozen)
• DeepStack Layers: [10, 21, 31]
• Aggregation: sum
• Validation Accuracy: 0.1513
• Validation Weighted F1: 0.0882

Usage

import torch
from huggingface_hub import hf_hub_download

ckpt = torch.load(
    hf_hub_download("FatimahEmadEldin/whisper-deepstack-tn-intent", "best_classifier.pt"),
    map_location='cpu'
)