PhanithLIM/whisper-small-aug-28-april-lightning-v1

Whisper Small Khmer

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

• Loss: 0.0776
• WER: 78.5976

Model description

Fine-tuned for automatic speech recognition (ASR) in Khmer using a combination of public and custom datasets.

Intended uses & limitations

More information needed

Training and evaluation data

Includes:

• PhanithLIM/ams-speech-dataset
• openslr/openslr
• google/fleurs
• PhanithLIM/kh-wmc
• PhanithLIM/wmc-international-news
• PhanithLIM/rfi-news-dataset
• PhanithLIM/aakanee-kh
• rinabuoy/khm-asr-open
• seanghay/khmer_grkpp_speech
• seanghay/khmer_mpwt_speech
• seanghay/km-speech-corpus

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• optimizer: AdamW (betas=(0.9, 0.999), eps=1e-08)
• lr_scheduler_type: constant
• lr_scheduler_warmup_steps: 1000
• num_epochs: 10
• max_steps: 56940
• logging_steps: 500
• save_steps: 500
• eval_steps: 500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	WER
0.235	1.0	5694	0.1025	80.7038
0.0872	2.0	11388	0.0852	80.3682
0.0636	3.0	17082	0.0789	79.8482
0.0494	4.0	22776	0.0776	78.5976

Framework versions

• Transformers 4.51.3
• PyTorch 2.5.1+cu121
• Datasets 3.5.1
• Tokenizers 0.21.0