fusion_gttbsc_distilbert-uncased-ft

Ground truth text with prosody encoding and ASR encoding residual cross attention fusion multi-label DAC

Model description

ASR encoder: Whisper small encoder
Prosody encoder: 2 layer transformer encoder with initial dense projection
Backbone: DistilBert uncased
Fusion: 2 residual cross attention fusion layers (F_asr x F_text and F_prosody x F_text) with dense layer on top
Pooling: Self attention
Multi-label classification head: 2 dense layers with two dropouts 0.3 and Tanh activation inbetween

Training and evaluation data

Trained on ground truth.
Evaluated on ground truth (GT) and normalized Whisper small transcripts (E2E).

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0007
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 20
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.41.2
• Pytorch 2.3.0+cu121
• Datasets 2.19.2
• Tokenizers 0.19.1