ViInstruct

ViInstruct is a large-scale Vietnamese dataset for instruction-based text-to-speech (TTS). Each utterance is paired with one free-form Vietnamese instruction that describes how it is spoken: speaker gender, age group and regional accent, plus speaking rate, pitch, loudness, expressiveness and (for the emotion subset) emotion.

• 1,224,945 utterance–instruction pairs, about 1,718 hours of speech
• 6 source corpora covering 3 tasks: general speech, regional accent, emotion
• Instructions written by an LLM (Qwen3.5-122B-A10B) from acoustic and speaker attributes, average 22.1 words
• Paper: ViInstruct: A Large-Scale Vietnamese Dataset for Instruction-Based Text-to-Speech Synthesis, APSIPA ASC 2026
• Demo with audio samples: https://thangquang09.github.io/viinstruct-demo/

This repository is an annotation layer, not an audio corpus. It holds the instructions, attribute labels, transcripts and split definitions, keyed to the identifiers of the source corpora. Get the audio from each original corpus (see Source corpora) and join on dataset + source_id / audio_path.

Example

Field	Value
caption_full	Giọng nữ trưởng thành miền Nam, nói nhanh, to, giọng cao và đều đều.
English gloss	Adult female voice from the South, fast, loud, high-pitched and monotone.
kw_gender / kw_age / region	nữ / người trưởng thành / south
kw_speaking_rate / kw_pitch / kw_loudness_lufs / kw_pitch_cv	nhanh / cao / to / đều đều

from datasets import load_dataset

ds = load_dataset("thangquang09/ViInstruct")
print(ds["train"][0]["caption_full"])

Dataset structure

Splits

Source	Task	Train	Validation	Test	Total
Vivoice	general	523,169	29,065	29,065	581,299
Dolly	general	526,067	29,210	29,232	584,509
ViMD	accent	10,087	1,293	1,410	12,790
LSVSC	accent	37,181	3,800	3,008	43,989
VNEMOS	emotion	189	11	11	211
ViSEC	emotion	1,932	107	108	2,147
Total		1,098,625	63,486	62,834	1,224,945

ViMD and LSVSC keep their official splits. Dolly is split speaker-disjointly by voice_id. Vivoice, VNEMOS and ViSEC have no official split or stable speaker id, so they use a stratified 90/5/5 split at the sample level.

Fields

All descriptor fields hold the Vietnamese keyword that was given to the LLM. An empty string means the attribute does not apply (for example kw_emotion outside the emotion subset).

Column	Description	Values
id	Unique id, {dataset}_{source_id}
source_id	Utterance id in the source corpus
dataset	Source corpus	vivoice, dolly, vimd, lsvsc, vnemos, visec
task	Subset	general, accent, emotion
split	Split	train, val, test
transcript	Text of the utterance
caption_full	The natural-language instruction
kw_gender	Gender	nam (male), nữ (female)
kw_age	Age group	trẻ em (child), thanh thiếu niên (teen), người trưởng thành (adult), cao tuổi (senior)
kw_accent	Regional accent	giọng miền Bắc / giọng miền Trung / giọng miền Nam
kw_accent_province	Province (ViMD only, 63 provinces)
region	Machine-readable accent region	north, central, south
kw_speaking_rate	Speaking rate, 7 levels	rất chậm, chậm, hơi chậm, vừa phải, hơi nhanh, nhanh, rất nhanh
kw_pitch	Mean pitch, 7 levels (binned per gender–accent group)	rất trầm, trầm, hơi trầm, trung bình, hơi cao, cao, rất cao
kw_loudness_lufs	Loudness, 5 levels	rất nhỏ, nhỏ, vừa, to, rất to
kw_pitch_cv	Pitch variation (expressiveness), 5 levels	đều đều, ít biểu cảm, biểu cảm vừa phải, sinh động, rất biểu cảm
kw_emotion	Emotion (emotion subset only)	trung lập (neutral), vui vẻ (happiness), buồn bã (sadness), tức giận (anger), sợ hãi (fear)
audio_path	Path of the audio file relative to the root of its source corpus

Attribute distribution

• Gender: 53.8% male, 46.2% female
• Age: 96.6% adult, 1.8% children, 1.1% teens, 0.4% seniors
• Accent: 59.6% North, 23.9% Central, 16.5% South
• Emotion (2,358 clips): anger, fear, happiness, neutral, sadness. Fear comes only from VNEMOS, whose anxiety class is mapped to fear.
• Median duration 4.31 s; mean pitch about 137 Hz (male) and 231 Hz (female); speaking rate about 4.57 syllables/s; loudness about −21.4 LUFS.

Source corpora

Corpus	Task	Content	Kept clips	Kept hours	Upstream license
viVoice	general	read and spontaneous speech from public video channels	581,299	779	CC BY-NC-SA 4.0 (gated)
Dolly-Audio	general	TTS-synthesized speech, 484 voice presets	584,509	808	CC BY-NC-SA 4.0 (gated)
ViMD (Dinh et al., EMNLP 2024)	accent	read speech, 63 provinces	12,790	54	CC BY-NC-ND 4.0
LSVSC (Tran et al., Electronics 2024)	accent	read speech with gender / age / region labels	43,989	75	no license published
VNEMOS (Quang Anh et al., ICDV 2024)	emotion	film and TV recordings	211	0.35	no license published
ViSEC (Thanh et al., ICASSP 2024)	emotion	acted speech from public video	2,147	1.12	CC BY 4.0

1,583,772 raw clips (2,223 h) were reduced to 1,224,945 (1,718 h) by quality filtering and deduplication. Dolly is synthetic and makes up 47.7% of the pairs.

How the dataset was built

1. Acoustic features and quality filtering. Duration, speaking rate (syllables per second from the transcript), pitch statistics with CREPE, loudness in LUFS, and speech proportion with Silero VAD. Clips are kept when duration ≥ 2.0 s (≥ 1.0 s for ViSEC), SNR ≥ 15 dB, clipping ratio < 0.01 and speech proportion ≥ 0.30.
2. Speaker attributes. Ground-truth labels are kept whenever the source corpus provides them. Otherwise a wav2vec2 gender classifier, a ViMD-trained wav2vec2 accent classifier (94.6% macro-F1) and a fine-tuned MERaLiON age classifier (teens / adults / seniors) predict them. ViMD transcripts come from PhoWhisper.
3. Quantile binning. Continuous features become Vietnamese keywords with quantile bins: 7 levels for pitch and speaking rate, 5 for loudness and pitch variation. Pitch is binned per gender–accent group. Bin edges are computed on Vivoice and reused for the other corpora.
4. Instruction generation. Qwen3.5-122B-A10B writes one instruction per utterance from the transcript, the shuffled attribute tags and three few-shot examples. Synonym substitution, random attribute order, varied few-shot examples and a per-utterance temperature in [0.3, 0.7] increase lexical variety.

Instruction quality audit

All 1,224,945 instructions were matched against their own tags with a Vietnamese descriptor lexicon (text only, no listening):

• 81.4% of instructions realize every tag in words; the rest paraphrase outside the lexicon.
• When a slot is realized, it names the conditioning level at least 96.6% of the time (loudness is the lowest; gender and accent are ≥ 99.99%).
• 99.7% of disagreements are one bin apart.
• No emotion word appears in the 1,222,587 instructions without an emotion tag.
• Known defect: 4.7% of instructions render người trưởng thành (adult) as lớn tuổi (elderly).

This audit checks tag → text only. Whether the tags match the audio rests on the ground-truth labels and the classifiers above.

Benchmark

The paper adapts two baselines to Vietnamese (Parler-TTS and CapSpeech NAR, both with a frozen ViT5-large instruction encoder) and evaluates them on a general full-tag set (1,228 clips) and an accent set (1,710 clips).

Metric	Reference audio	Parler-TTS	CapSpeech NAR
CER (%) ↓	0.96	3.19	0.85
WER (%) ↓	2.15	4.80	1.55
UTMOSv2 ↑	2.95	2.54	2.65
Gender macro-F1 (%) ↑		84.83	100.00
Age macro-F1 (%) ↑		49.94	52.82
Accent macro-F1, general set (%) ↑		48.56	77.48
Accent macro-F1, accent set (%) ↑		38.82	72.37
Acoustic bin accuracy (%) ↑		36.46	66.55

Chance level is about 33% macro-F1 for the three-way age and accent tasks and 14–20% for bin accuracy.

Limitations

• Age labels are heavily adult and partly come from a coarse three-class classifier; age adherence of both baselines is close to chance.
• The emotion subset is small (2,358 clips, 1.5 h) and has no reliable automatic metric.
• The accent subset is regionally imbalanced toward Northern speech.
• Pitch is summarized by scalar statistics that do not separate lexical tone from paralinguistic prosody.
• Nearly half of the pairs come from synthesized speech (Dolly).
• Predicted attributes and the benchmark are automatic. No human listening study has been done yet.

License

The annotation layer (instructions, attribute labels and splits) is released for non-commercial research use only. Each subset also remains governed by the terms of its source corpus, listed in Source corpora. The upstream terms are mutually incompatible or unstated, so this repository does not redistribute any audio. See LICENSE.md.

Ethics

No new speech was recorded and no attempt was made to identify speakers. Where ground-truth labels are absent, gender, age and accent are model predictions, not self-reported identity, and should not be read as factual claims about individuals. The instructions inherit that uncertainty.

Citation

@inproceedings{ly2026viinstruct,
  title     = {{ViInstruct}: A Large-Scale {Vietnamese} Dataset for Instruction-Based Text-to-Speech Synthesis},
  author    = {Ly, Thang and Do, Tri-Nhan and Nguyen, Tu-Anh and Mac, Dang-Khoa},
  booktitle = {Proc. Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC)},
  year      = {2026}
}

Please also cite the source corpora you use.