Awesome Vietnamese Speech Datasets

A curated, machine-readable catalogue of publicly available Vietnamese speech datasets — for ASR, TTS, speech translation, speaker recognition, dialect identification, and emotion recognition.

Why this exists

Vietnamese speech datasets are scattered across VLSP challenge pages, Hugging Face mirrors, personal GitHub repos, and appendices of papers. Finding them is annoying. Finding out whether you can actually use them — how clean the labels are, whether the licence permits shipping a product — is worse.

This list tries to answer three questions for every entry:

1. What is it? Hours, speech type, domain, sample rate.
2. Can I trust the labels? Human-transcribed, pseudo-labelled, or a mix — and what the known quality problems are.
3. Can I use it? Licence, commercial use, and whether you need to register.

Everything lives in datasets.yaml. The tables below are generated from it, so the list is also a dataset you can query:

import yaml

data = yaml.safe_load(open("datasets.yaml"))

# Everything I can legally ship in a commercial product
usable = [d for d in data["datasets"] if d["commercial_ok"] is True]

# Everything with human labels, over 100 hours, for ASR
big_clean = [
    d for d in data["datasets"]
    if "asr" in d["task"] and d["labels"] == "human" and (d["hours"] or 0) > 100
]

Start here

If you want to…	Use
Smoke-test an ASR pipeline in 10 minutes	VIVOS (15h, clean, tiny)
Pre-train a general Vietnamese ASR model	PhoAudiobook + viVoice + Bud500
Fine-tune for a call-centre / voicebot	VietSuperSpeech (conversational)
Train or clone a TTS voice	viVoice (multi-speaker)
Ship something commercially	Common Voice (CC0) — and read every other licence twice
Benchmark against published numbers	VLSP2020-Test-T1, VLSP2023-PublicTest
Evaluate speech-to-text translation	FLEURS (parallel)
Test robustness across dialects	ViMD

A word on licences. Most Vietnamese speech corpora are research-only, and several of the largest are pseudo-labelled from YouTube audio — which raises questions the licence field alone does not settle. If you are building a product, treat the Comm. column as a starting point for your own legal review, not as an answer.

23 datasets — roughly 11,455 labeled hours catalogued.

Legend: Comm. = usable commercially (✅ yes / ❌ no / ❓ unclear — always read the licence yourself). Labels ⚠️ pseudo = machine-generated transcripts, expect systematic errors. ¹ = figures come from a secondary source and need confirming (help wanted).

Large-scale general ASR

Dataset	Hours	Type	Domain	Labels	Licence	Comm.	Access
GigaSpeech 2 (Vietnamese subset)	6039h	mixed	general	⚠️ pseudo	unknown	❓	🟢 open
PhoAudiobook ¹	1494h	audiobook	literature	human	unknown	❓	🟢 open
VietSpeech (VSV-1100)	1100h	spontaneous	social media, general	mixed	unknown	❓	🟢 open
viVoice ¹	1000h	mixed	general	⚠️ pseudo	unknown	❓	🟢 open
Bud500 (VIET_BUD500)	500h	spontaneous	general	⚠️ pseudo	Apache-2.0 (disputed)	❓	🟢 open
YODAS / YODAS2 (Vietnamese subset) ¹	?	mixed	general	mixed	CC-BY-3.0	❓	🟢 open
LSVSC (Large-scale Vietnamese Speech Corpus) ¹	?	mixed	general	human	unknown	❓	🟢 open

Conversational / spontaneous

Dataset	Hours	Type	Domain	Labels	Licence	Comm.	Access
VietSuperSpeech	267.39h	conversational	call-center, chatbot, everyday conversation	⚠️ pseudo	unknown	❓	🟢 open

Read / studio speech (clean baselines)

Dataset	Hours	Type	Domain	Labels	Licence	Comm.	Access
VIVOS	15h	read	general	human	CC-BY-NC-SA-4.0	❌	🟢 open
FOSD (FPT Open Speech Dataset) ¹	?	read	general	human	unknown	❓	🟢 open
InfoRe (Dataset 1 & 2) ¹	?	read	general	human	unknown	❓	🟢 open
FLEURS (Vietnamese)	?	read	general	human	CC-BY-4.0	✅	🟢 open
Mozilla Common Voice (Vietnamese)	?	read	general	human	CC0-1.0	✅	🟢 open

Domain-specific & shared tasks

Dataset	Hours	Type	Domain	Labels	Licence	Comm.	Access
VietMed	16h +2200h unlab.	conversational	medical	human	unknown	❓	🟢 open
VLSP 2020 ASR (VinBigData-VLSP2020)	100h	mixed	news, general	human	research-only	❌	🟡 register
VLSP 2021 ASR	280h +400h unlab.	mixed	general	human	research-only	❌	🟡 register
VLSP 2023 ASR ¹	?	mixed	general	mixed	research-only	❌	🟡 register
VLSP 2025 ASR/SER	?	mixed	general, emotion	mixed	research-only	❌	🟡 register

Speech translation

Dataset	Hours	Type	Domain	Labels	Licence	Comm.	Access
PhoST (English-Vietnamese Speech Translation)	508h	mixed	TED talks, general	human	unknown	❓	🟢 open

Dialect, speaker & paralinguistics

Dataset	Hours	Type	Domain	Labels	Licence	Comm.	Access
ViMD (Multi-dialect Vietnamese) ¹	100h	broadcast	general	human	unknown	❓	🟢 open
Vietnam-Celeb ¹	?	spontaneous	entertainment	human	unknown	❓	🟢 open
ViSEC ¹	?	mixed	emotion	human	unknown	❓	🟢 open

TTS-oriented

Dataset	Hours	Type	Domain	Labels	Licence	Comm.	Access
viet-tts-dataset (ntt123)	35.9h	synthesized	literature	human	research-only	❌	🟢 open

Notes on each dataset

The table is deliberately terse. The caveats are where the value is — read these before you download 100 GB.

• GigaSpeech 2 (Vietnamese subset) (2024) · speechcolab/gigaspeech2 · paper
  THE LARGEST Vietnamese speech corpus available: ~7,324 raw hours refined to ~6,039 hours via Noisy Student Training filtering. One of three low-resource languages in GigaSpeech 2, built by automated crawling + transcription + iterative refinement from YouTube. Papers often use a filtered subset ("Sub-GigaSpeech2-Vi"). Entirely PSEUDO-LABELLED — it is the single biggest source of Vietnamese ASR training data and also the single biggest source of systematic label noise. Do not evaluate on it.
• PhoAudiobook (2025) · vinai/PhoAudiobook
  Largest single Vietnamese corpus by hours. Narrated audiobook material — clean and well-articulated, but the register is far from spontaneous speech. Widely used as the pre-training bulk in VLSP 2025 systems. Frequently used as "Sub-PhoAudioBook" (a filtered subset) in papers.
• VietSpeech (VSV-1100) (2024) · NhutP/VietSpeech · paper
  1,100h of Vietnamese "social voice" collected from diverse social media and online sources (Pham Quang Nhut et al., UIT). Broad domain coverage and strong accent diversity make it one of the core corpora in the standard Vietnamese ASR training mix (it is one of the ten corpora behind Zipformer-30M-RNNT-6000h). NOTE — the literature cites this under two names: "VietSpeech" (Pham et al., 2024b) and "VSV-1100" (Vietnamese Social Voice, Nhut et al., 2024) are the SAME dataset; do not count it twice when summing hours. No conversational two-way dialogue focus.
• viVoice (2024) · capleaf/viVoice
  Built for multi-speaker Vietnamese speech synthesis (FPT University HCMC). Multi-speaker coverage makes it the default choice for TTS voice-cloning training; also heavily reused for ASR pre-training.
• Bud500 (VIET_BUD500) (2024) · linhtran92/viet_bud500 · paper
  ~500h from free public audio (podcasts, travel, books, food), covering North/Central/South accents. Prepared by the VietAI research team. CAVEAT — read before using: independent analysis (Vietnamese ASR: A Revisit, 2026) found a substantial fraction fails quality filters, transcripts have inaccuracies, and there is no standardised punctuation, capitalisation, or word-level timestamps. Utterances are short (~3s avg). LICENCE CONFLICT — the HF repo ships an Apache-2.0 licence file (© VietAI Research) while the same dataset card states "Dataset provided for research purposes only". These two statements contradict each other. Ask the authors before shipping anything built on it.
• YODAS / YODAS2 (Vietnamese subset) (2024) · espnet/yodas · paper
  500k+ hours across 100+ languages from YouTube (CMU/ESPnet); Vietnamese is one of the covered languages. Labels come from YouTube subtitles - MANUAL subtitles in some subsets, AUTOMATIC captions of variable quality in others, so check which subset you are pulling. YODAS2 is the long-form (video-level, unsegmented) variant at 24 kHz, useful if you need long-context or your own segmentation. HELP WANTED - nobody has published the Vietnamese-only hour count; if you compute it, please contribute it.
• LSVSC (Large-scale Vietnamese Speech Corpus) (2024)
  Cited across recent Vietnamese ASR surveys (Tran et al., 2024). HELP WANTED — canonical link, hours, and licence need confirming from the primary source.
• VietSuperSpeech (2026) · thanhnew2001/VietSuperSpeech · paper
  52,023 audio-text pairs / 267.39h, explicitly targeting the casual spontaneous register (train 46,822 / 240.67h, dev-test 5,201 / 26.72h). Sourced from four public YouTube channels; 16 kHz mono PCM WAV, 3-30s utterances, fully diacritised. Transcripts are PSEUDO-LABELS produced by Zipformer-30M-RNNT-6000h via sherpa-onnx — the authors themselves warn of systematic errors on disfluencies, overlapping speech, and rare vocabulary. Currently the best public fit for call-centre / voicebot fine-tuning.
• VIVOS (2016) · AILAB-VNUHCM/vivos · paper
  The classic Vietnamese ASR benchmark — 15h of read speech from 65 speakers in a quiet studio (Luong & Vu, 2016). Tiny and clean; still the standard smoke test, but useless as a proxy for conversational performance. Start here when you want a 10-minute end-to-end sanity check of your pipeline.
• FOSD (FPT Open Speech Dataset) (2020) · paper
  Clean studio read-speech corpus released by FPT. A good clean-speech baseline and a common TTS starting point (see the FOSD-Tacotron-2 model dataset). HELP WANTED — confirm exact hours from the primary release.
• InfoRe (Dataset 1 & 2) (2019)
  Two open corpora released by InfoRe Technology; commonly mirrored on the Hub alongside VIVOS. Often the first thing people concatenate with VIVOS for a quick baseline. HELP WANTED — official link and licence.
• FLEURS (Vietnamese) (2022) · google/fleurs · paper
  Google's 102-language benchmark; the Vietnamese split is read parallel translations of FLoRes sentences. Small, but the parallel structure makes it the natural evaluation set for speech-to-text translation, and it is the standard cross-lingual comparison point.
• Mozilla Common Voice (Vietnamese) (2019) · mozilla-foundation/common_voice_17_0 · paper
  Crowdsourced read speech, CC0 — one of the very few Vietnamese speech corpora that is unambiguously usable commercially. Small for Vietnamese and highly variable in recording quality, but the licence makes it disproportionately useful. Version bumps annually; pin the version in your config.
• VietMed (2024) · leduckhai/VietMed · paper
  16h labeled medical speech + 1,000h unlabeled medical + 1,200h unlabeled general. Claimed as the world's largest public medical ASR dataset across seven axes (duration, speakers, diseases, recording conditions, speaker roles, unique medical terms, accents), and the first to cover all ICD-10 disease groups and all accents in a country. Ships with w2v2-Viet and XLSR-53-Viet pre-trained models. The labeled subset is often cited as "VietMed-L".
• VLSP 2020 ASR (VinBigData-VLSP2020) (2020)
  ~250h total, of which ~100h public — roughly 20h read speech + 80h spontaneous speech crawled from open sources, manually transcribed at ~96% accuracy. Test-T1 is one of the two de-facto Vietnamese ASR leaderboards. Restricted to registered participants; skews formal/scripted.
• VLSP 2021 ASR (2021)
  ~280h labeled general-domain data plus ~400h unlabeled audio for the semi-supervised track. Registration required.
• VLSP 2023 ASR (2023)
  Includes a voting-based pseudo-labelled track (VLSP2023-voting-pseudo-labeled) where labels come from consensus across multiple ASR systems — an early precedent for pseudo-label quality control in Vietnamese. VLSP2023-PublicTest is the second de-facto benchmark alongside VLSP2020-Test-T1.
• VLSP 2025 ASR/SER (2025) · paper
  The most recent VLSP speech shared task, combining ASR and speech emotion recognition. The overview paper is worth reading even if you never touch the data: it documents what the winning systems actually did — joint CTC-RNN-T training, and an in-training data-filtering strategy that keeps only samples with 0.3 < CTC confidence < 0.95 and 5% < WER < 40%, discarding both the trivially easy and the corrupted. That recipe is the current best public answer to the pseudo-label noise problem in Vietnamese ASR.
• PhoST (English-Vietnamese Speech Translation) (2022) · paper
  508 audio hours / 331K triplets of (sentence-length audio, English source transcript, Vietnamese target subtitle). The first large-scale English-Vietnamese speech translation study, and still the reference dataset for the direction. IMPORTANT — direction is English audio to Vietnamese text (En->Vi). For Vietnamese audio to English text you are largely on your own: CoVoST 2 does NOT include Vietnamese in either direction, a gap worth knowing before you plan a Vi->En experiment. The authors found cascaded (ASR + MT) still beat end-to-end on this data.
• ViMD (Multi-dialect Vietnamese) (2024) · paper
  Multi-dialect corpus with task, baselines, and challenges (Dinh et al., 2024). The go-to resource for dialect identification and for measuring how badly your ASR model degrades outside the Hanoi/HCMC axis. Long utterances (~19s avg), so watch your max-duration filter.
• Vietnam-Celeb (2023)
  Speaker-recognition corpus in the VoxCeleb tradition (Pham et al., 2023). The default choice for Vietnamese speaker verification / diarization work.
• ViSEC (2025) · paper
  ~5,400 utterances for Vietnamese speech emotion recognition; used in the VLSP 2025 ASR/SER shared task, typically combined with international SER corpora (IEMOCAP, EMODB, RAVDESS, CREMA-D, EmoV-DB).
• viet-tts-dataset (ntt123) (2020) · ntt123/viet-tts-dataset
  35.9h / 5.4GB. Text is public-domain Vu Trong Phung prose; the AUDIO IS SYNTHESIZED by the offline Google TTS engine on Android and is explicitly NOT for commercial use. Useful for bootstrapping a TTS pipeline, but you are training on another model's output — do not expect natural prosody, and do not ship it.

The pseudo-label problem

Several of the biggest Vietnamese corpora (Bud500, GigaSpeech2-Vi, viVoice, VietSuperSpeech) have machine-generated transcripts. This is not a reason to avoid them — it is how Vietnamese ASR got to where it is — but it has consequences worth stating plainly:

• Errors are systematic, not random. Teacher models fail predictably on disfluencies, overlapping speech, rare vocabulary, and strong accents. Your student model will inherit exactly those blind spots.
• A model trained on pseudo-labels from model X cannot be fairly evaluated against model X. Keep human-labelled test sets (VLSP, VIVOS, FLEURS) for evaluation.
• Filter aggressively. The VLSP 2025 systems that did best filtered by CTC confidence and WER thresholds, keeping only moderately difficult samples and discarding both the trivially easy and the obviously corrupted.

Related lists

• awesome-vietnamese-nlp — broader Vietnamese NLP resources (text, models, tools); this list is the speech-focused, structured companion.
• SEACrowd — Southeast Asian multilingual/multimodal data hub; several Vietnamese speech corpora are mirrored there with a common loader.

Contributing

Additions, corrections, and especially verifications are very welcome — see CONTRIBUTING.md. Entries marked ¹ have figures taken from secondary sources and need someone to confirm them against the primary release. That is the single most useful thing you can do for this repo.

Citation

@misc{awesome_vietnamese_speech_datasets,
  title  = {Awesome Vietnamese Speech Datasets},
  author = {Le, Thi Lien and contributors},
  year   = {2026},
  url    = {https://github.com/lienlt10/awesome-vietnamese-speech-datasets}
}

Licence

The catalogue (this list, datasets.yaml, and the scripts) is released under CC0-1.0 — public domain. The datasets themselves are under their own licences, listed above. Do not assume CC0 applies to them.