whisper-small-pashto

This model is a fine-tuned version of openai/whisper-small for Pashto speech recognition, trained on Common Voice Scripted Speech 27.0 - Pashto (train split).

Results

Model	Test set	Clips	WER raw	WER normalized
openai/whisper-small	Common Voice 27.0 Pashto	1,500	106.0	104.2
whisper-small-pashto	Common Voice 27.0 Pashto	5,000	33.1	27.8
openai/whisper-small	Common Voice Spontaneous Speech 5.0 Pashto	76	93.4	91.4
whisper-small-pashto	Common Voice Spontaneous Speech 5.0 Pashto	76	40.3	33.0

Normalized WER ignores punctuation, diacritics and letters Pashto writers use interchangeably (ي/ی, ك/ک, ګ/گ, Arabic-Indic digits), so two spellings of the same words do not count as an error. The untuned model often writes Pashto in Latin letters or adds words, which puts its WER above 100. The spontaneous-speech set is unscripted speech the model was not trained on.

During training, on 500 dev clips: loss 0.3107, raw WER 29.10.

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: Use OptimizerNames.ADAMW_TORCH_FUSED 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: 500
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.8218	0.1350	1000	0.4323	39.4820
0.6801	0.2699	2000	0.3679	32.7877
0.5982	0.4049	3000	0.3349	29.2206
0.5473	0.5398	4000	0.3197	29.9291
0.5596	0.6748	5000	0.3107	29.0985

Framework versions

• Transformers 5.21.0
• Pytorch 2.11.0+cu128
• Datasets 4.8.5
• Tokenizers 0.23.3