NobleX-AI/vi-stt-mimi-vsv1100

VietSpeech (VSV-1100): Vietnamese speech as Mimi codes with timed transcripts

1,133.7 hours, 1,005,381 clips, 14,527,644 words, 27 channels of Vietnamese speech from VietSpeech / VSV-1100 (MMLab-UIT) (Vietnamese social media speech), 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, fosd, vivoice, bud500.

• 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).
• Fused machine labels. Every chunk was transcribed by Qwen3-ASR-1.7B (Apache-2.0) and gipformer-1.5 (68M RNNT, MIT); the label is their word-level vote together with the corpus' own transcript, weighted 1.5 (it outvotes one teacher). gipformer wins ties, Qwen's foreign words, casing and punctuation are kept where the vote kept its words, and chunks on which all voters disagree were dropped. On five public Vietnamese test sets (FLEURS, Common Voice 17, Bud500, GigaSpeech 2, LSVSC) the same fusion scores about 5% mean WER, against 7.0% for Qwen3-ASR alone and 7.8% for PhoWhisper-large: better than either teacher, but not error-free.
• 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 vsv1100/<upstream parquet file>), 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-vsv1100", 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 VietSpeech (VSV-1100, MMLab-UIT: Pham Quang Nhut, Duong Pham Hoang Anh, Nguyen Vinh Tiep), licensed under the Apache License 2.0, and is released under the same license (full text in LICENSE). Changes: the audio was segmented, encoded as Mimi codes (no audio is distributed), and re-transcribed as described above.

Transcripts were produced with Qwen3-ASR-1.7B (Apache-2.0) and gipformer-1.5 (MIT), credited here with thanks.

Citation

Please cite the upstream dataset:

@misc{VietSpeech,
    author = {Pham Quang Nhut and Duong Pham Hoang Anh and Nguyen Vinh Tiep},
    title = {VietSpeech: Vietnamese social voice dataset},
    url = {https://github.com/NhutP/VietSpeech},
    year = {2024}
}