NobleX-AI/vi-stt-mimi-bud500

Bud500: Vietnamese speech as Mimi codes with timed transcripts

459.0 hours, 633,509 clips, 5,783,317 words, 1 channels of Vietnamese speech from Bud500 (VietAI Research) (free public audio: podcasts, travel, books and more), stored as Kyutai Mimi codes with word-timed transcripts.

It is one source of the ~20,000 h on which NobleX-AI trains a streaming Vietnamese speech recognizer (Kyutai's delayed-streams STT, kyutai/stt-1b-en_fr, fine-tuned for Vietnamese). The other sources released so far: yodas2, vsv1100, fosd, vivoice.

• No audio files. Each clip is its Mimi encoding (mimi-pytorch-e351c8d8@125.safetensors of kyutai/stt-1b-en_fr: 12.5 frames per second, 32 codebooks, int16), which is what the STT model reads. The codes decode back to 24 kHz audio (lossy, see Usage).
• The corpus' own transcripts, normalised to the spoken form in lower case (numbers, dates and units read out as words), with casing and punctuation added by the dragonSwing/xlm-roberta-capu tagger (CC-BY-SA-4.0).
• Timing. Every word has its start and end in seconds within its clip (forced alignment).

Files

file	content
transcripts.parquet	one row per clip: clip, channel, offset and n_frames (its rows in codes.npy), duration (s), text (cased, punctuated), words (lower case) with start / end (s)
codes.npy	int16 [total_frames, 32]: the Mimi codes of all clips back to back
clips.npz	per clip: offset, n_frames, dur, channel (index into meta.json channels), src
words.npz	per word: clip, onset, end (s), tok_off, n_tok (its ids in tokens.npy)
tokens.npy	int16 token ids of all words; itos.json maps an id to its string (▁x starts a word with letter x, ##x continues it, other entries are whole syllables)
labels/labels_0.npz	per word: case (0 as is, 1 Capitalised, 2 UPPER, 3 other: the string is in labels/mixed_0.json under the word's index) and punct (0 none, else 1 + index in , . ? : !, the mark after the word)
vocab.json, meta.json	token vocabulary (syllables); channel names (channels are bud500 (a single channel: the corpus has no speaker ids)), hours, counts
MANIFEST.sha256	checksums of every file (sha256sum -c MANIFEST.sha256)

Usage

import numpy as np, pyarrow.parquet as pq
from huggingface_hub import snapshot_download

root = snapshot_download("NobleX-AI/vi-stt-mimi-bud500", repo_type="dataset")
rows = pq.read_table(f"{root}/transcripts.parquet")
codes = np.load(f"{root}/codes.npy", mmap_mode="r")
r = rows.slice(0, 1).to_pylist()[0]
clip_codes = codes[r["offset"]:r["offset"] + r["n_frames"]]         # [T, 32] Mimi codes, 12.5 Hz
print(r["text"], list(zip(r["words"], r["start"], r["end"]))[:5])

# back to audio (pip install moshi)
import torch
from moshi.models import loaders
from huggingface_hub import hf_hub_download
mimi = loaders.get_mimi(hf_hub_download("kyutai/stt-1b-en_fr", "mimi-pytorch-e351c8d8@125.safetensors"), device="cpu")
mimi.set_num_codebooks(32)
with torch.no_grad():
    wav = mimi.decode(torch.from_numpy(clip_codes.astype(np.int64).T[None]))  # [1, 1, T * 1920] at 24 kHz

License

This dataset is a modified version of Bud500, Copyright (c) 2024 VietAI Research (Bud500, VietAI research team). The upstream card lists both CC-BY-NC-SA-4.0 (metadata) and Apache-2.0 (text); we apply the stricter one and release this dataset under CC-BY-NC-SA-4.0: non-commercial use only, share-alike. Upstream: dataset provided for research purposes only. Changes: the audio was encoded as Mimi codes (no audio is distributed) and the transcripts were normalised and punctuated as described above.

Casing and punctuation were added with the dragonSwing/xlm-roberta-capu tagger (CC-BY-SA-4.0), credited here with thanks.

Citation

Please cite the upstream dataset:

@misc{Bud500,
  author = {Anh Pham, Khanh Linh Tran, Linh Nguyen, Thanh Duy Cao, Phuc Phan, Duong A. Nguyen},
  title = {Bud500: A Comprehensive Vietnamese ASR Dataset},
  url = {https://github.com/quocanh34/Bud500},
  year = {2024}
}