metadata

library_name: peft
language:
  - zh
license: mit
base_model: openai/whisper-large-v3-turbo
tags:
  - wft
  - whisper
  - automatic-speech-recognition
  - audio
  - speech
  - generated_from_trainer
datasets:
  - JacobLinCool/common_voice_19_0_zh-TW
metrics:
  - wer
model-index:
  - name: whisper-large-v3-turbo-common_voice_19_0-zh-TW-lora
    results:
      - task:
          type: automatic-speech-recognition
          name: Automatic Speech Recognition
        dataset:
          name: JacobLinCool/common_voice_19_0_zh-TW
          type: JacobLinCool/common_voice_19_0_zh-TW
        metrics:
          - type: wer
            value: 32.55535607420706
            name: Wer

whisper-large-v3-turbo-common_voice_19_0-zh-TW-lora

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the JacobLinCool/common_voice_19_0_zh-TW dataset. It achieves the following results on the evaluation set:

Loss: 0.1786
Wer: 32.5554
Cer: 8.6009
Decode Runtime: 90.9833
Wer Runtime: 0.1257
Cer Runtime: 0.1534

Model description

This is an open-source Traditional Chinese (Taiwan) automatic speech recognition (ASR) model.

Intended uses & limitations

This model is designed to be a prompt-free ASR model for Traditional Chinese. Due to its inherited language identification (LID) system from Whisper, which supports other Chinese language variants under the same language token (zh), we expect that performance may degrade when transcribing Simplified Chinese.

The model is free to use under the MIT license.

Training and evaluation data

This model was trained on the Common Voice Corpus 19.0 Chinese (Taiwan) Subset, containing about 50k training examples (44 hours) and 5k test examples (5 hours). This dataset is four times larger than the combination of training and validation set (train+validation) of mozilla-foundation/common_voice_16_1, which includes about 12k examples.

Training procedure

Tensorboard

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0002
train_batch_size: 4
eval_batch_size: 32
seed: 42
gradient_accumulation_steps: 8
total_train_batch_size: 32
optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 50
training_steps: 5000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Decode Runtime	Wer Runtime	Cer Runtime
No log	0	0	2.7208	76.5011	20.4851	89.4916	0.1213	0.1639
1.1832	0.1	500	0.1939	39.9561	10.8721	90.0926	0.1222	0.1555
1.5179	0.2	1000	0.1774	37.6621	9.9322	89.8657	0.1225	0.1545
0.6179	0.3	1500	0.1796	36.2657	9.8325	90.2480	0.1198	0.1573
0.3626	1.0912	2000	0.1846	36.2258	9.7801	90.3306	0.1196	0.1539
0.1311	1.1912	2500	0.1776	34.8095	9.3214	90.3124	0.1286	0.1610
0.1263	1.2912	3000	0.1763	36.1261	9.3563	90.4271	0.1330	0.1650
0.2194	2.0825	3500	0.1891	34.6898	9.3114	91.1932	0.1320	0.1643
0.1127	2.1825	4000	0.1838	34.0714	9.1095	90.2416	0.1196	0.1529
0.3792	2.2824	4500	0.1786	33.1339	8.7679	90.9144	0.1310	0.1550
0.0606	3.0737	5000	0.1786	32.5554	8.6009	90.9833	0.1257	0.1534

Framework versions

PEFT 0.13.2
Transformers 4.46.1
Pytorch 2.4.0
Datasets 3.0.2
Tokenizers 0.20.1