wav2vec2-large-lv60_phoneme-timit_english_timit-4k_simplified

This model is a fine-tuned version of facebook/wav2vec2-large-lv60 on the TIMIT dataset. It achieves the following results on the evaluation set:

• Loss: 0.2796
• Phone Error Rate: 0.0838 (8.38%)

Model description

Trained on a simplified version of the TIMIT phone set.

Intended uses & limitations

Merged Phonemes

• Based on error analysis for each phoneme from the original TIMIT phoneme set.
• See this repo for detailed analysis.
• ax-h → ax
• axr → er
• ix → ih
• ux → uw
• zh → z
• em → m
• en → n
• eng → ng
• nx → n
• hv → hh

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 16
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 300
• training_steps: 3000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Phone Error Rate
7.3185	1.04	300	3.6437	0.9617
2.5644	2.08	600	0.7668	0.1559
0.6782	3.11	900	0.3794	0.1231
0.4542	4.15	1200	0.3278	0.1164
0.3834	5.19	1500	0.3043	0.1151
0.3407	6.23	1800	0.2872	0.1119
0.3179	7.27	2100	0.2842	0.1110
0.2988	8.3	2400	0.2834	0.1102
0.2834	9.34	2700	0.2826	0.1100
0.2814	10.38	3000	0.2796	0.1100

Framework versions

• Transformers 4.38.1
• Pytorch 2.0.1
• Datasets 2.16.1
• Tokenizers 0.15.2