whisper-small-clearglobal-hausa-asr-1.0.0

This model is a fine-tuned version of openai/whisper-small on the TWB Voice 1.0 dataset.

It achieves the following results on the internal evaluation set:

• WER: 2.23%
• CER: 3.11%

Training and evaluation data

This model was trained by colleagues from the Makerere University Centre for Artificial Intelligence and Data Science in collaboration with CLEAR Global. We gratefully acknowledge their expertise and partnership.

Model was trained and tested on the approved Hausa subset of TWB Voice 1.0 dataset.

Train/dev/test portions correspond to the splits in this dataset version. Test splits consist of speakers not present in train and dev splits.

We also tested on external datasets: Common voice v17 and Naija Voices test splits.

The evaluation results are as follows:

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• 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_ratio: 0.08
• num_epochs: 50.0
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.53.1
• Pytorch 2.7.1+cu128
• Datasets 4.0.0
• Tokenizers 0.21.2