DNU-ViSpeech

Publication status: The manuscript describing DNU-ViSpeech is currently under review and has not yet been accepted for publication. This repository provides the dataset and associated benchmark results for research and reproducibility. The paper citation will be added when publicly available.

Public release: the dataset owner authorized public release on 2026-09-05 after attesting that all applicable human-review gates were completed. See PUBLIC_RELEASE_APPROVAL.md for the approval record.

Dataset Description

DNU-ViSpeech is a Vietnamese scripted/read-speech corpus for automatic speech recognition (ASR). This release contains 31,070 utterances from 26 anonymized speakers, totaling 75.570 hours. Audio is WAV, 16 kHz, mono, uncompressed PCM 16-bit.

The primary protocol is DNU-ASR-v1.1. Its evaluation split combines unseen-speaker and recording-configuration shift; those two factors cannot be isolated with the current data.

Motivation

The corpus supports research on Vietnamese ASR adaptation, especially whether compact models can approach larger zero-shot systems after domain fine-tuning on read speech.

Dataset Structure

The repository follows the Hugging Face AudioFolder layout. Each split has its own metadata.csv, and file_name is relative to that CSV.

data/
  train/metadata.csv
  train/audio/speaker001/...
  validation/metadata.csv
  validation/audio/speaker019/...
  test/metadata.csv
  test/audio/speaker023/...

Load it with:

from datasets import load_dataset

dataset = load_dataset("TrangLe2108/DNU-ViSpeech")

Normalization and scoring

The exact DNU-ASR-v1.1 normalization and scoring implementation used for the reported WER, CER, and speaker-macro WER is included in this repository:

• scripts/lib_text_normalize.py: v1.1 Vietnamese text normalizer.
• scripts/17_score_asr_v11.py: WER, CER, speaker-macro WER, duration-bucket metrics, and bootstrap WER confidence intervals.
• tests/test_normalize_v11.py and tests/test_scoring_v11.py: unit tests for both implementations.
• protocols/asr_v1.1/test_compound_shift.jsonl: locked normalized references for the 3,345-item test split.

Install the scoring dependencies and run the tests:

python -m pip install jiwer==4.0.0 numpy==1.26.4 pytest
python -m pytest -q tests/test_normalize_v11.py tests/test_scoring_v11.py

Score a prediction CSV containing experiment_id, audio_id, and hypothesis_normalized columns:

python scripts/17_score_asr_v11.py \
  --manifest protocols/asr_v1.1/test_compound_shift.jsonl \
  --predictions path/to/predictions.csv \
  --model-label "My ASR model" \
  --setting zero-shot

Generate hypothesis_normalized from each raw model hypothesis with normalize_text_v11 in scripts/lib_text_normalize.py. References and hypotheses must use the same v1.1 normalizer.

Released benchmark results

The canonical DNU-ASR-v1.1 outputs reported in Table III of the manuscript currently under review are released for all six settings. Each setting contains exactly 3,345 test utterances:

• predictions/asr_v1.1/test_compound_shift/: normalized system hypotheses keyed by anonymized utterance ID, plus sanitized run manifests. Raw hypotheses are deliberately omitted.
• results/asr_v1.1/scored/: per-utterance WER and CER scoring outputs.
• results/asr_v1.1/: canonical aggregate, per-speaker, duration-bucket, paired-comparison, and efficiency tables.
• results/asr_v1.1/release_manifest.json: file hashes, row counts, protocol linkage, and reported metrics for the six settings.

Model	Setting	WER (%)	CER (%)	Speaker-macro WER (%)
Whisper-large-v3	Zero-shot	7.15	3.84	5.83
PhoWhisper-large	Zero-shot	4.83	2.74	3.86
PhoWhisper-small	Zero-shot	6.60	3.60	5.25
PhoWhisper-small	Fine-tuned	4.89	2.77	4.10
Parakeet-CTC-0.6B	Zero-shot	6.69	3.66	5.41
Parakeet-CTC-0.6B	Fine-tuned	4.93	2.53	4.11

The released CSVs preserve the canonical normalized text and metric values, projected to the minimum fields needed to reproduce Table III of the manuscript currently under review. Scored text fields are explicitly named reference_normalized and hypothesis_normalized. Run manifests remove private machine paths while retaining checkpoint hashes and inference provenance. Raw hypotheses and fine-tuned checkpoints are not distributed. Diagnostic, control, and superseded runs are intentionally excluded from the canonical tables.

Data Fields

• audio: created by AudioFolder from file_name.
• file_name: relative path to the WAV file inside the split.
• audio_id: stable anonymized sample identifier. One retained legacy identifier, speaker012_0842.orig, does not follow the usual speakerNNN_NNNN suffix convention.
• speaker_id: anonymized identifier from speaker001 to speaker026.
• text: human-reviewed spoken/reference transcript used for ASR.
• text_normalized: v1.1-normalized ASR reference.
• prompt_text: text the participant was asked to read.
• prompt_text_normalized: normalized prompt text.
• reference_source: prompt_confirmed or human_corrected.
• verification_level: none, reviewed, or verified under the v1.1 protocol.
• review_status: accepted or corrected.
• qc_status: corpus QC status.
• resampled_from_raw: whether the source recording required resampling before release.
• duration: metadata duration in seconds.
• sample_rate: 16,000 Hz for every file.
• num_channels: 1 for every file.
• checksum_sha256: SHA-256 of the released WAV.
• split: train, validation, or test.

Sensitive demographic attributes and internal reviewer notes present in the project metadata are deliberately omitted from this Hugging Face release.

Data Splits

Split	Speakers	Utterances	Hours
train	18 (speaker001–speaker018)	22,616	55.936
validation	4 (speaker019–speaker022)	5,109	12.545
test	4 (speaker023–speaker026)	3,345	7.088
Total	26	31,070	75.570

The split is speaker-disjoint, checksum-disjoint, and normalized-reference-text-disjoint. The original project label dev is exposed as the conventional Hugging Face split name validation; no samples were reassigned.

Collection Process

Participants received sentence lists and self-recorded Vietnamese read speech, usually over multiple sessions. Each sentence was stored as a separate WAV file. The target was approximately three hours per speaker, but three test speakers remain below that target.

The split assignment was not random or stratified. Speakers 001–022 were recorded in a native 16 kHz configuration, while most files from speakers 023–026 required resampling from 44.1/48 kHz. Consequently, speaker identity, split, and recording configuration are confounded in the test condition.

Speaker Information

Only anonymized speaker IDs are published here. The release contains 26 speakers. Direct identity mappings and demographic fields are not included.

Transcription Procedure

Transcripts originated as pre-existing prompts. Quality control checked whether each audio file matched its prompt. ASR-assisted screening was used to prioritize human review; it was not used to silently replace ground truth. The v1.1 schema separates prompt text from the spoken/reference transcript and records human corrections.

The source project reports a full ASR re-audit of all 31,070 files, followed by correction of 60 transcript rows and re-locking of the protocol. Verification coverage still varies: many training rows passed automated checks without utterance-level listening, whereas the test rows are marked verified or reviewed under v1.1.

Train/Validation/Test Protocol

• Train only on train.
• Use validation for checkpoint and hyperparameter selection.
• Use test only for final evaluation.
• Apply the same v1.1 text normalizer to references and hypotheses.
• Do not interpret the test set as a pure channel-robustness or pure unseen-speaker benchmark.

Ethics and Consent

The project documentation states that all 26 included participants provided written consent for public release of their audio and transcripts. Signed forms and the identity mapping are retained outside this repository because they contain direct identifiers. The dataset owner attested on 2026-09-05 that the applicable consent-record review was completed before public release.

Voice recordings are biometric data and can enable re-identification. Anonymized IDs do not eliminate that risk. Participant withdrawal is handled by removing the corresponding speaker from the next release and recording the change; historical frozen snapshots are not silently rewritten.

Personal and Sensitive Information

No email address, telephone number, government-ID-like number, or known self-identification recording was found in the active release metadata by automated checks. Automated scanning cannot prove the absence of every possible identifier.

Audio itself remains sensitive biometric data. To reduce linkage risk, this Hugging Face package excludes gender, birth year, region, recording device, inferred recording environment, and internal reviewer notes.

Intended Uses

• Vietnamese read-speech ASR training and evaluation.
• Compact-model adaptation and accuracy/efficiency research.
• Evaluation on anonymized, speaker-disjoint splits under the documented compound shift.
• Reproducible comparison with the locked DNU-ASR-v1.1 protocol.

Out-of-Scope Uses

• Natural or spontaneous conversation evaluation.
• Claims that 26 speakers represent Vietnamese speech broadly.
• Fairness conclusions across demographic groups.
• Pure channel-robustness or pure unseen-speaker claims.
• Identifying, profiling, surveilling, or authenticating participants outside the documented research consent.
• Commercial use, which is prohibited by the license.

Biases and Limitations

• The corpus has only 26 speakers and limited demographic and regional diversity.
• It contains read speech rather than spontaneous speech.
• The test split is smaller than train and has uneven duration by speaker.
• Recording configuration is confounded with the split boundary.
• Script difficulty differs substantially across speakers; supplementary speaker026 material is notably harder.
• Prompt text was assembled from multiple sources, but per-sentence provenance was not recorded and cannot be reconstructed from released metadata.
• Verification depth varies across the corpus.
• One legacy sample ID contains .orig and does not match the documented usual ID pattern.

Licensing

The project provides the package under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). The source license states that it applies only to material the project owns. Because per-sentence prompt provenance was not recorded, this remains a documented limitation. The dataset owner attested on 2026-09-05 that the applicable rights/provenance and license review was completed and accepted for this release.

Benchmark Information

Baseline results under DNU-ASR-v1.1 for the 3,345-utterance compound-shift test split are:

Model	Setting	Parameters	WER	CER	Speaker-macro WER
Whisper large-v3	zero-shot	1.55B	7.15%	3.84%	5.83%
PhoWhisper-large	zero-shot	1.55B	4.83%	2.74%	3.86%
PhoWhisper-small	zero-shot	244M	6.60%	3.60%	5.25%
PhoWhisper-small	fine-tuned	244M	4.89%	2.77%	4.10%
Parakeet-CTC-0.6B Vietnamese	zero-shot	609M	6.69%	3.66%	5.41%
Parakeet-CTC-0.6B Vietnamese	fine-tuned	609M	4.93%	2.53%	4.11%

These results use the corrected v1.1 references. The fine-tuned systems were retrained after the transcript audit. Consult the benchmark repository reports for run-level provenance and known stale historical artifacts.

Citation

The project citation identifies one dataset author. The dataset owner confirmed the authorship record for this release on 2026-09-05.

@dataset{le2026dnuvispeech,
  author    = {Lê Thị Thùy Trang},
  title     = {DNU-ViSpeech},
  year      = {2026},
  version   = {1.1},
  publisher = {Dai Nam University}
}

See CITATION.cff for the machine-readable citation.