VietENT-Speech — synthetic Vietnamese ENT clinical speech

41,816 synthetic 16 kHz utterances of Vietnamese ear-nose-throat clinical speech: 20,908 unique sentences, each rendered once clean and once through a measured recording-channel model, across 106 cloned voices.

from datasets import load_dataset
ds = load_dataset("diepduclai/VietENT-Speech", split="train", streaming=True)   # ~7.5 GB, stream it
next(iter(ds))["audio"]

License is CC BY-NC-SA 4.0, and that was not a free choice

31 of the 106 voices are cloned from VIVOS, released under CC BY-NC-SA 4.0. ShareAlike propagates to derivative works, so this corpus inherits both NonCommercial and ShareAlike. Research use only.

voice source	voices	share of corpus	license
ViMedCSS	33	30.1%	CC-BY-4.0 (audio sourced from YouTube)
VIVOS	31	29.1%	CC-BY-NC-SA-4.0
FLEURS	20	19.7%	CC-BY-4.0
VLSP2020	20	19.2%	CC-BY-4.0
project authors	2	2.0%	authors' own voices

The source corpora are released for research and the clones read generic clinical sentences, but the original speakers did not consent to voice cloning specifically. The reference recordings themselves are not redistributed here. To have a voice removed, open a discussion.

Fields

field	description
id	{sentence_id}__{voice_id}
audio	16 kHz mono WAV
text	transcript, sentence-case, numbers written as words
term	the ENT term the sentence was built around
group	clinical category
speaker	voice id; the prefix gives the source corpus (vmc/vv/fl/vl/bs/bn)
cond	raw = clean synthesis, aug = through the channel model

Splits

split	rows	note
train	21,776
validation	1,575
test	2,412	every clip of 6 voices held out entirely

test is speaker-disjoint, not sentence-disjoint, because the question that matters is whether a model reads a patient it has never heard. No sentence appears in two splits.

The channel model

Synthetic speech is too clean, and a model trained on clean audio learns clean audio. aug clips pass through, in physical order: speaking-rate resampling, then a lowpass, then a synthetic exponential-decay room impulse response, then real noise mixed at a measured SNR. The noise bank is extracted from silent stretches of actual ENT consultation recordings, not generated.

Channel weights follow the target deployment: phone recording 50%, MacBook built-in mic 42%, USB desk mic 8%.

Speaking rate is resampled at 1.00/1.15/1.30/1.45x, weighted to a mean of 1.22x. This is not decoration: the synthesis runs at 9.3 characters per second of speech while real Vietnamese medical speech runs at 14.0, and a doctor speaking to a patient is faster still.

Is it hard enough to be worth training on?

A synthetic corpus the target model already reads perfectly teaches nothing. Measured with Qwen/Qwen3-ASR-1.7B on 300 aug clips, against 11 real ENT consultation recordings:

metric	raw	aug	real ENT speech
term recall	75.2%	70.2%	81.5%
number recall	98.5%	97.8%	99.4%
N-WER	3.8%	5.3%	4.6%
deletion rate	0.23%	0.26%	0.30%

Read the N-WER row. The clean synthesis is easier than real speech by 0.77 points — the voice cloning is good, and a model trained on raw alone would be learning an easier world than the one it ships into. The channel model moves it to 5.3%, slightly harder than real. That is what aug is for.

aug is harder than real speech on terminology, which is the axis being taught, and clean on numbers and deletions, which must not be taught wrong. 164 of those 300 sentences had a term misheard.

One caveat the numbers do not show: all 106 voices pass through a single synthesis model, so they share its artifacts no matter how many voices there are. A 6-14 second reference clip carries timbre, not the way a person actually talks — the hesitations, the overlapping turns, the moment someone turns away from the microphone. Voice diversity does not remove model-level bias, and replay is what covers it.

Quality control

Every clip was screened for label-vs-audio duration consistency, speech presence, loudness and clipping; 4 of 41,816 were dropped. Reference-text leakage into the head of a generated clip — an OmniVoice failure mode where part of the voice prompt is read back — measured 3.3% on a 300-clip sample, of which roughly a fifth actually overlap the source reference text. No duplicate audio and no repeated sentence exists in the corpus.

Do not train on this alone — and here is the other half

Synthetic speech drifts a model away from real acoustics. The fix used here is replay: mix in real Vietnamese medical speech at roughly 1.21 hours of this corpus per hour of real speech — a ratio measured in audio hours, not row counts, because clips here average 4.3s against 7.0s for the real corpus and counting rows gives a misleading 2.4:1.

The real half is ViMedCSS train. Its audio is not redistributed here — it is someone else's corpus and you can fetch it yourself. What mix/ gives you is the part that took the work:

• mix/replay_manifest.jsonl — all 11,822 ViMedCSS train segments with the transcript renormalised into this corpus's number convention (14% of labels changed, e.g. "5 alpha reductase" to "năm alpha reductase"). Training on two spellings of one sound teaches a decoder to guess.
• The same file marks 126 segments to drop, each with a reason. 37 carry a full label over audio with no speech at all, 21 carry a label far too short for their audio, 37 are heavily clipped, 19 are cut short, 8 are at the PCM noise floor, and 4 are sentences that appear verbatim under a second video. The first two categories teach hallucination and deletion respectively.
• split is assigned by source video, not by segment. Two segments from one video share a speaker, a room and a topic, so a random split makes a validation set that only measures memorisation.
• mix/mixture.json — the exact hours and ratio.
• replay_prep.py — the script that turns ViMedCSS parquet into 16 kHz wavs with these labels.

A warning if you plan to evaluate on ViMedCSS test: 93.8% of its segments come from videos that also appear in train, 2 segment ids are identical across the two, and 20% of test sentences share a 7-word run with train. Training on ViMedCSS train and scoring on ViMedCSS test is self-grading. This project uses VietMed, its own ENT recordings, and the 100 ViMedCSS test segments whose video never appears in train.

The sentences are written rather than transcribed, so they carry no evidence about real consultation discourse — only about terminology and clinical phrasing.

Companion dataset

diepduclai/VietENT-Text is the text corpus alone, under CC-BY-4.0, with the generator.

Citation

@misc{vietent2026,
  title  = {VietENT: a Vietnamese ENT clinical corpus for speech recognition},
  author = {Diep Duc Lai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/diepduclai/VietENT-Speech}
}