whisper-small-mn-7

This model is a fine-tuned version of openai/whisper-small on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.3061
• Wer: 32.6469
• Cer: 11.2319

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 15000
• mixed_precision_training: Native AMP

Training script

python train.py \
    --train_datasets "mozilla-foundation/common_voice_11_0|mn|train+validation,google/fleurs|mn_mn|train+validation,bayartsogt/ulaanbal-v0||train" \
    --eval_datasets "mozilla-foundation/common_voice_11_0|mn|test" \
    --whisper-size "small" \
    --language "mn,Mongolian" \
    --keep-chars " абвгдеёжзийклмноөпрстуүфхцчшъыьэюя.,?!" \
    --train-batch-size 32 \
    --eval-batch-size 32 \
    --max-steps 15000 \
    --num-workers 8 \
    --version 7 \

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.3416	0.61	1000	0.4335	51.0979	17.8608
0.2266	1.22	2000	0.3383	39.5346	13.6468
0.2134	1.83	3000	0.2994	35.6565	12.1677
0.165	2.43	4000	0.2927	34.1927	11.4602
0.1205	3.04	5000	0.2879	33.5209	11.3002
0.1284	3.65	6000	0.2884	32.7507	10.9885
0.0893	4.26	7000	0.3022	33.0894	11.2075
0.0902	4.87	8000	0.3061	32.6469	11.2319
0.065	5.48	9000	0.3233	32.8163	11.1595
0.0436	6.09	10000	0.3372	32.6852	11.1384
0.0469	6.7	11000	0.3481	32.8272	11.2867
0.0292	7.3	12000	0.3643	33.0784	11.3785
0.0277	7.91	13000	0.3700	33.1877	11.3600
0.0196	8.52	14000	0.3806	33.3734	11.4273
0.016	9.13	15000	0.3844	33.3188	11.4248

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.0+cu117
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2