Chore: Multi-voice audio generation

[hahuyhoang411]

Motivation:

• Current dataset is mono-voice (female) -> Model can't answer properly if the sound input is male's voice

Goal:

1. Expand the model with multiple speakers (genders, regions, innotation, etc.)
2. Create v2 dataset: 2B tokens?

Side idea:

1. Overlap audio samples: e.g.
   • Speaker 1: 100 samples
   • Speaker 2: 100 samples (20 samples overlap with speaker 1)

[hungphongtrn]

I created a dataset with unique speaker as a reference for WhisperSpeech here: unique_speaker_audio

Basically, I sample data from openslr/librispeech_asr. For each unique speaker ID, I randomly get 1 sample.

We have 921 unique speakers. You guys can submit your voice to :))