Scalable-Machine-Learning-Lab_2

Text Transcription using Transformers to your Mother Tongue.

Improvements

• model-centric approaches:

1. tune hyperparameter - set warmup_steps=100 and max_steps=1000. We found that the model starts overfitting after 1000 epochs.
2. change model - use openai/whisper-medium instead of openai/whisper-small. We found the eval loss stop decreasing after 1000 epochs which means the performance ceiling of small model has been reached.

• data-centric approach: Increase sampling rate to 24000 to get more details.

App URL: https://huggingface.co/spaces/esnagy/hungarian_speech_transcriber