- DATA CLEANING
1. Removing Erroneous Audio and transcript
line 133
2. Label preprocessing
The transcription of each audio must not have errors, transcripts are normalized(line 106) by removing non-alphabetical characters(line 133 regex) and converting to lowercase.
-DATA TRANSFORMATION
0. Adding Silence (Referenced in
add_silence function)Directly corresponds to enhancing the model's robustness to initial silences in audio inputs.
many audio will have different starting times, by adding silences, the model won’t confuse start time with signal or speech.
1. Resampling Audio to 16kHz (
resample_audio function)line 73 resample_audio(audio, desired_sample_rate) is about resampling the audio to 16kHz if needed cause human voice is typically up to about 8 kHz and all audio needss to be consistent.2.feature extrac