feat: More multilingual audio training data

[hahuyhoang411]

Problem

Currently we only have mostly Vietnamese, some Thai, some Singlish.

Suggestion

Process the youtube audio in this dataset: https://huggingface.co/datasets/espnet/yodas2/

Thanks jpc for sharing this dataset

[hahuyhoang411]

Description

This is the preprocessing code for the GigaSpeech2 dataset.

Gist link

Data demo

Problem

• The original GigaSpeech2 dataset contains over 7M raw Vietnamese audio files.
• Audio files are stored in .tar.gz archives.
• Transcriptions are saved separately in .tsv files.

Preprocessing Steps

• Load the dataset and map each audio file to its corresponding transcription using IDs.
• Convert the dataset into a Hugging Face-compatible format for use with the current training code.