This model is a fine-tuned version of openai/whisper-tiny on the MINDS-14 (en-US) dataset. It achieves the following results on the evaluation set:
openai/whisper-tiny (39M parameters) fine-tuned for English speech recognition on e-banking voice queries (Unit 5 hands-on exercise of the Hugging Face Audio Course).
Transcription of short American-English customer-service / banking requests (8 kHz telephone audio, resampled to 16 kHz). Trained on only ~450 utterances, so it is a course exercise, not a general-purpose ASR model.
PolyAI/minds14, en-US subset (563 utterances): the first 450
examples for training (445 kept after dropping clips longer than 30 s) and the remaining 113
for evaluation.
Metrics are decimals (not percentages). Wer is the word error rate after Whisper's BasicTextNormalizer
(lower-casing, punctuation removal); Wer Ortho is computed on the raw text. Zero-shot whisper-tiny on the same
evaluation split: normalized WER 0.4014, orthographic WER 0.5416.
Code: https://github.com/emrezorlu1239/course-projects/tree/main/huggingface-audio-course (unit5_whisper_finetune/).
The following hyperparameters were used during training:
| Training Loss | Epoch | Step | Validation Loss | Model Preparation Time | Wer | Wer Ortho |
|---|---|---|---|---|---|---|
| 0.2652 | 3.5714 | 100 | 0.5062 | 0.0015 | 0.3022 | 0.2973 |
| 0.0375 | 7.1429 | 200 | 0.5724 | 0.0015 | 0.3099 | 0.3060 |
| 0.0051 | 10.7143 | 300 | 0.6271 | 0.0015 | 0.3158 | 0.3109 |
| 0.0022 | 14.2857 | 400 | 0.6416 | 0.0015 | 0.2946 | 0.2887 |
| 0.0019 | 17.8571 | 500 | 0.6475 | 0.0015 | 0.2963 | 0.2899 |
model.safetensors 151 MBtraining_args.bin 5.5 kBopenai/whisper-tiny