datasets.yaml

# =============================================================================
# awesome-vietnamese-speech-datasets — single source of truth
# =============================================================================
# README.md is GENERATED from this file. Do not edit the tables in README.md
# by hand — edit here and run `python scripts/gen_readme.py`.
#
# Schema (see scripts/validate.py for the authoritative rules):
#   id            : unique slug, lowercase, [a-z0-9-]
#   name          : display name
#   hours         : float | null   (labeled hours; use `hours_unlabeled` for the rest)
#   task          : list of [asr, tts, s2tt, speaker-id, ser, dialect-id, diarization]
#   speech_type   : read | spontaneous | conversational | audiobook | broadcast | mixed | synthesized
#   domain        : short free text (general, medical, call-center, ...)
#   sample_rate   : int (Hz) | null
#   labels        : human | pseudo | mixed | none
#   license       : SPDX-ish string, or "research-only", or "unknown"
#   commercial_ok : true | false | null   (null = unclear, read the license yourself)
#   access        : open | registration | request
#   url           : primary landing page
#   hf            : Hugging Face dataset id | null
#   paper         : URL | null
#   year          : int
#   verified      : true  = numbers checked against the primary source
#                   false = taken from a secondary source, HELP WANTED to confirm
#   notes         : caveats a practitioner actually needs (quality, gotchas)
# =============================================================================

datasets:

  # ---------------------------------------------------------------------------
  # Large-scale general ASR
  # ---------------------------------------------------------------------------
  - id: phoaudiobook
    name: PhoAudiobook
    hours: 1494
    task: [asr, tts]
    speech_type: audiobook
    domain: literature
    sample_rate: null
    labels: human
    license: unknown
    commercial_ok: null
    access: open
    url: https://huggingface.co/datasets/vinai/PhoAudiobook
    hf: vinai/PhoAudiobook
    paper: null
    year: 2025
    verified: false
    notes: >
      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.

  - id: vivoice
    name: viVoice
    hours: 1000
    task: [tts, asr]
    speech_type: mixed
    domain: general
    sample_rate: null
    labels: pseudo
    license: unknown
    commercial_ok: null
    access: open
    url: https://huggingface.co/datasets/capleaf/viVoice
    hf: capleaf/viVoice
    paper: null
    year: 2024
    verified: false
    notes: >
      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.

  - id: viet-bud500
    name: Bud500 (VIET_BUD500)
    hours: 500
    task: [asr]
    speech_type: spontaneous
    domain: general
    sample_rate: 16000
    labels: pseudo
    license: Apache-2.0 (disputed)
    commercial_ok: null
    access: open
    url: https://huggingface.co/datasets/linhtran92/viet_bud500
    hf: linhtran92/viet_bud500
    paper: https://github.com/quocanh34/Bud500
    year: 2024
    verified: true
    notes: >
      ~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.

  - id: gigaspeech2-vi
    name: GigaSpeech 2 (Vietnamese subset)
    hours: 6039
    task: [asr]
    speech_type: mixed
    domain: general
    sample_rate: null
    labels: pseudo
    license: unknown
    commercial_ok: null
    access: open
    url: https://huggingface.co/datasets/speechcolab/gigaspeech2
    hf: speechcolab/gigaspeech2
    paper: https://arxiv.org/abs/2406.11546
    year: 2024
    verified: true
    notes: >
      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.

  - id: lsvsc
    name: LSVSC (Large-scale Vietnamese Speech Corpus)
    hours: null
    task: [asr]
    speech_type: mixed
    domain: general
    sample_rate: null
    labels: human
    license: unknown
    commercial_ok: null
    access: open
    url: https://github.com/vndee/awsome-vietnamese-nlp
    hf: null
    paper: null
    year: 2024
    verified: false
    notes: >
      Cited across recent Vietnamese ASR surveys (Tran et al., 2024). HELP WANTED —
      canonical link, hours, and licence need confirming from the primary source.

  # ---------------------------------------------------------------------------
  # Conversational / spontaneous
  # ---------------------------------------------------------------------------
  - id: vietsuperspeech
    name: VietSuperSpeech
    hours: 267.39
    task: [asr]
    speech_type: conversational
    domain: call-center, chatbot, everyday conversation
    sample_rate: 16000
    labels: pseudo
    license: unknown
    commercial_ok: null
    access: open
    url: https://huggingface.co/datasets/thanhnew2001/VietSuperSpeech
    hf: thanhnew2001/VietSuperSpeech
    paper: https://arxiv.org/abs/2603.01894
    year: 2026
    verified: true
    notes: >
      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.

  - id: vietspeech
    name: VietSpeech (VSV-1100)
    hours: 1100
    task: [asr]
    speech_type: spontaneous
    domain: social media, general
    sample_rate: null
    labels: mixed
    license: unknown
    commercial_ok: null
    access: open
    url: https://huggingface.co/datasets/NhutP/VietSpeech
    hf: NhutP/VietSpeech
    paper: https://github.com/NhutP/VietSpeech
    year: 2024
    verified: true
    notes: >
      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.

  # ---------------------------------------------------------------------------
  # Read / studio speech (clean baselines)
  # ---------------------------------------------------------------------------
  - id: vivos
    name: VIVOS
    hours: 15
    task: [asr]
    speech_type: read
    domain: general
    sample_rate: 16000
    labels: human
    license: CC-BY-NC-SA-4.0
    commercial_ok: false
    access: open
    url: https://huggingface.co/datasets/AILAB-VNUHCM/vivos
    hf: AILAB-VNUHCM/vivos
    paper: https://aclanthology.org/W16-5207/
    year: 2016
    verified: true
    notes: >
      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.

  - id: fosd
    name: FOSD (FPT Open Speech Dataset)
    hours: null
    task: [asr, tts]
    speech_type: read
    domain: general
    sample_rate: null
    labels: human
    license: unknown
    commercial_ok: null
    access: open
    url: https://data.mendeley.com/datasets/k9sxg2twv4/4
    hf: null
    paper: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7287261/
    year: 2020
    verified: false
    notes: >
      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.

  - id: infore
    name: InfoRe (Dataset 1 & 2)
    hours: null
    task: [asr, tts]
    speech_type: read
    domain: general
    sample_rate: null
    labels: human
    license: unknown
    commercial_ok: null
    access: open
    url: https://github.com/vndee/awsome-vietnamese-nlp
    hf: null
    paper: null
    year: 2019
    verified: false
    notes: >
      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.

  - id: fleurs-vi
    name: FLEURS (Vietnamese)
    hours: null
    task: [asr, s2tt]
    speech_type: read
    domain: general
    sample_rate: 16000
    labels: human
    license: CC-BY-4.0
    commercial_ok: true
    access: open
    url: https://huggingface.co/datasets/google/fleurs
    hf: google/fleurs
    paper: https://arxiv.org/abs/2205.12446
    year: 2022
    verified: true
    notes: >
      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.

  - id: common-voice-vi
    name: Mozilla Common Voice (Vietnamese)
    hours: null
    task: [asr]
    speech_type: read
    domain: general
    sample_rate: 48000
    labels: human
    license: CC0-1.0
    commercial_ok: true
    access: open
    url: https://commonvoice.mozilla.org/vi/datasets
    hf: mozilla-foundation/common_voice_17_0
    paper: https://arxiv.org/abs/1912.06670
    year: 2019
    verified: true
    notes: >
      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.

  # ---------------------------------------------------------------------------
  # Domain-specific
  # ---------------------------------------------------------------------------
  - id: vietmed
    name: VietMed
    hours: 16
    hours_unlabeled: 2200
    task: [asr]
    speech_type: conversational
    domain: medical
    sample_rate: null
    labels: human
    license: unknown
    commercial_ok: null
    access: open
    url: https://huggingface.co/datasets/leduckhai/VietMed
    hf: leduckhai/VietMed
    paper: https://arxiv.org/abs/2404.05659
    year: 2024
    verified: true
    notes: >
      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".

  - id: vlsp2020-asr
    name: VLSP 2020 ASR (VinBigData-VLSP2020)
    hours: 100
    task: [asr]
    speech_type: mixed
    domain: news, general
    sample_rate: null
    labels: human
    license: research-only
    commercial_ok: false
    access: registration
    url: https://vlsp.org.vn/vlsp2020/eval/asr
    hf: null
    paper: null
    year: 2020
    verified: true
    notes: >
      ~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.

  - id: vlsp2021-asr
    name: VLSP 2021 ASR
    hours: 280
    hours_unlabeled: 400
    task: [asr]
    speech_type: mixed
    domain: general
    sample_rate: null
    labels: human
    license: research-only
    commercial_ok: false
    access: registration
    url: https://vlsp.org.vn/vlsp2021
    hf: null
    paper: null
    year: 2021
    verified: true
    notes: >
      ~280h labeled general-domain data plus ~400h unlabeled audio for the
      semi-supervised track. Registration required.

  - id: vlsp2023-asr
    name: VLSP 2023 ASR
    hours: null
    task: [asr]
    speech_type: mixed
    domain: general
    sample_rate: null
    labels: mixed
    license: research-only
    commercial_ok: false
    access: registration
    url: https://vlsp.org.vn/vlsp2023
    hf: null
    paper: null
    year: 2023
    verified: false
    notes: >
      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.

  - id: phost
    name: PhoST (English-Vietnamese Speech Translation)
    hours: 508
    task: [s2tt, asr]
    speech_type: mixed
    domain: TED talks, general
    sample_rate: null
    labels: human
    license: unknown
    commercial_ok: null
    access: open
    url: https://github.com/VinAIResearch/PhoST
    hf: null
    paper: https://arxiv.org/abs/2208.12813
    year: 2022
    verified: true
    notes: >
      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.

  - id: yodas-vi
    name: YODAS / YODAS2 (Vietnamese subset)
    hours: null
    task: [asr]
    speech_type: mixed
    domain: general
    sample_rate: 24000
    labels: mixed
    license: CC-BY-3.0
    commercial_ok: null
    access: open
    url: https://huggingface.co/datasets/espnet/yodas
    hf: espnet/yodas
    paper: https://arxiv.org/abs/2406.00899
    year: 2024
    verified: false
    notes: >
      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.


  - id: vlsp2025-asr-ser
    name: VLSP 2025 ASR/SER
    hours: null
    task: [asr, ser]
    speech_type: mixed
    domain: general, emotion
    sample_rate: null
    labels: mixed
    license: research-only
    commercial_ok: false
    access: registration
    url: https://vlsp.org.vn/vlsp2025
    hf: null
    paper: https://aclanthology.org/2025.vlsp-1.1.pdf
    year: 2025
    verified: true
    notes: >
      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.

  # ---------------------------------------------------------------------------
  # Dialect / speaker / paralinguistics
  # ---------------------------------------------------------------------------
  - id: vimd
    name: ViMD (Multi-dialect Vietnamese)
    hours: 100
    task: [asr, dialect-id]
    speech_type: broadcast
    domain: general
    sample_rate: null
    labels: human
    license: unknown
    commercial_ok: null
    access: open
    url: https://arxiv.org/abs/2410.03458
    hf: null
    paper: https://arxiv.org/abs/2410.03458
    year: 2024
    verified: false
    notes: >
      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.

  - id: vietnam-celeb
    name: Vietnam-Celeb
    hours: null
    task: [speaker-id, diarization]
    speech_type: spontaneous
    domain: entertainment
    sample_rate: null
    labels: human
    license: unknown
    commercial_ok: null
    access: open
    url: https://github.com/hkab/vietnam-celeb
    hf: null
    paper: null
    year: 2023
    verified: false
    notes: >
      Speaker-recognition corpus in the VoxCeleb tradition (Pham et al., 2023).
      The default choice for Vietnamese speaker verification / diarization work.

  - id: visec
    name: ViSEC
    hours: null
    task: [ser]
    speech_type: mixed
    domain: emotion
    sample_rate: null
    labels: human
    license: unknown
    commercial_ok: null
    access: open
    url: https://vlsp.org.vn/vlsp2025
    hf: null
    paper: https://aclanthology.org/2025.vlsp-1.1.pdf
    year: 2025
    verified: false
    notes: >
      ~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).

  # ---------------------------------------------------------------------------
  # TTS-oriented
  # ---------------------------------------------------------------------------
  - id: viet-tts-dataset
    name: viet-tts-dataset (ntt123)
    hours: 35.9
    task: [tts]
    speech_type: synthesized
    domain: literature
    sample_rate: null
    labels: human
    license: research-only
    commercial_ok: false
    access: open
    url: https://huggingface.co/datasets/ntt123/viet-tts-dataset
    hf: ntt123/viet-tts-dataset
    paper: null
    year: 2020
    verified: true
    notes: >
      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.