README.md

#Wav2Vec2/Whisper Fine-Tuning

-- Project Status: [Completed]

Project Objective

The objective is to fine-tune state of the art architectures Wav2Vec2 and Whisper on Common Voice Turkish dataset and compare the results

Methods Used

• Machine Learning
• Data Pre-Processing

Libraries

• Pytorch
• Transformers
• Librosa
• Jiwer

Results:

Whisper-small gives a WER of 22% (max-train steps 1500) Wav2Vec2 xls-r-300m gives a WER of 32%(num-epochs:30)