A large multilingual corpus for pretraining TTS/STT models, gathered and normalized from 230+ public sources. ~191k hours of audio across 150+ languages, organized into 1,544 dataset configs and tokenized to 34.5B NeuCodec speech tokens (50 Hz, single-codebook) over 111.1M clips.
Each config is one source dataset normalized to rows of {audio_filename, text, speaker}:
audio_filename — clip path inside that config's
_audio.zip (mono MP3).
text —… See the full description on the dataset page: https://huggingface.co/datasets/malaysia-ai/Multilingual-TTS.
