This model is a fine-tuned version of openai/whisper-small on the mozilla-foundation/common_voice_11_0 el dataset. It achieves the following results on the evaluation set:

Loss: 0.4805
Wer: 20.6352

Model description

This model is trained for transcription on the Greek subset on mozilla-foundation/common_voice_11_0 interleaved splits train+eval

Intended uses & limitations

This is part of the Whisper Finetuning Event (December 2022)

Training and evaluation data

Training used interleaved splits: train + evaluation. Evaluation was done on the test split. Data was streamed from Hugging Face's Hub.

Training procedure

The script used has been uploaded in the files of this space The command to run it was:

python ./run_speech_recognition_seq2seq_streaming.py \
                --model_name_or_path   "openai/whisper-small" \
                --model_revision       "main" \
                --do_train             True \
                --do_eval              True \
                --use_auth_token       False \
                --freeze_encoder       False \
                --model_index_name     "whisper-sm-el-xs" \
                --dataset_name         "mozilla-foundation/common_voice_11_0" \
                --dataset_config_name  "el" \
                --audio_column_name    "audio" \
                --text_column_name     "sentence" \
                --max_duration_in_seconds 30 \
                --train_split_name    "train+validation" \
                --eval_split_name      "test" \
                --do_lower_case         False \
                --do_remove_punctuation False \
                --do_normalize_eval     True \
                --language              "greek" \
                --task                  "transcribe" \
                --shuffle_buffer_size   500 \
                --output_dir             "./data/finetuningRuns/whisper-sm-el-xs" \
                --per_device_train_batch_size 16 \
                --gradient_accumulation_steps 4  \
                --learning_rate          1e-5 \
                --warmup_steps           500 \
                --max_steps              5000 \
                --gradient_checkpointing True \
                --fp16                   True \
                --evaluation_strategy    "steps" \
                --per_device_eval_batch_size 8 \
                --predict_with_generate  True \
                --generation_max_length  225 \
                --save_steps             1000 \
                --eval_steps             1000 \
                --logging_steps          25 \
                --report_to              "tensorboard" \
                --load_best_model_at_end True \
                --metric_for_best_model  "wer" \
                --greater_is_better      False \
                --push_to_hub            False \
                --overwrite_output_dir    True

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 1e-05
train_batch_size: 16
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 64
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 500
training_steps: 5000
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0024	18.01	1000	0.4246	21.0438
0.0003	37.01	2000	0.4805	20.6352
0.0001	56.01	3000	0.5102	20.8395
0.0001	75.0	4000	0.5296	21.0717
0.0001	94.0	5000	0.5375	21.0253

Here is the summary from the log of the run:

***** train metrics *****
  epoch                    =        94.0
  train_loss               =      0.0222
  train_runtime            = 23:06:13.19
  train_samples_per_second =       3.847
  train_steps_per_second   =        0.06
12/08/2022 11:20:17 - INFO - __main__ - *** Evaluate ***

***** eval metrics *****
  epoch                   =       94.0
  eval_loss               =     0.4805
  eval_runtime            = 0:23:03.68
  eval_samples_per_second =      1.226
  eval_steps_per_second   =      0.153
  eval_wer                =    20.6352
Thu 08 Dec 2022 11:43:22 AM EST

Framework versions

Transformers 4.26.0.dev0
Pytorch 1.13.0
Datasets 2.7.1.dev0
Tokenizers 0.12.1

farsipal/whisper-sm-el-xs

Whisper-Small (el) for Transcription