chore: Augmenting current Vietnamese speech dataset

[hahuyhoang411]

Problem

Most of the current Vietnamese dataset has the same pipeline (Youtube -> VAD -> Whisper -> Normalization). This pipeline cleaned all of the punctuation in the transcription makes it lost information.
e.g. Bud500

or gigiaspeech2

Goal

Improve current dataset by adding back nature punctation

Draft solution

• Use Whisperv3 large to transcribe audio in those dataset -> whisper_transcription_format

• Use Llama3.2 8B to use the ground_truth and whisper_transcription_format to reformat the ground_truth

• Note: We only use the structure of WhisperLarge not the label of it

• Architecture:

Tasklist

• Transcribe bud500, gigaspeech refined vi using Whisper large v3
• Setup distil label pipeline
• QA dataset

[bachvudinh]

Need to discuss more on the pipeline cause:

• result from Whisper large V3 is bad for bud500 audio.
• Lllama tend to complete, modify the label instead of correcting typo only