Text Cleaning for ASR: The Case of Turkish

Text cleaning is an essential component of natural language processing that helps prepare training data for automatic speech recognition (ASR) systems. It involves transforming raw data into a "cleaner" version, closer to the actual phonetics of what was said. This process is language-dependent and requires a multi-step processing pipeline to ensure accurate transcriptions. In Turkish text cleaning, challenges include handling the apostrophe, consonant assimilation, vowel harmony rules, and processing currencies and numbers. Text cleaning is crucial for ASR training as it helps improve the accuracy of transcriptions by ensuring a good match between phonetics and phonetic transcription.