Tran1312/Automatic-Speech-Recognition

Vietnamese ASR Collection

A consolidated Vietnamese speech collection for Automatic Speech Recognition (ASR), hosted at Tran1312/Automatic-Speech-Recognition.

The collection integrates audio–transcript pairs from multiple Vietnamese speech datasets into a unified metadata and storage format. Audio samples are normalized to the filename convention:

sample_<id>.wav

and described using a common JSONL schema.

This repository should not be interpreted as a newly recorded speech corpus. The underlying recordings and transcripts originate from previously released Vietnamese speech datasets and have been reorganized into a common representation for large-scale ASR training.

Data Sources

The collection contains data originating from the following sources:

Source	Reference
Bud500	VietAI — dataset card
PhoAudiobook	dataset card, ACL 2025 paper
VieNeu-TTS-140h	dataset card
VLSP 2020	dataset mirror
InfoRe Technology Public Dataset №1	dataset mirror
InfoRe Technology Public Dataset №2	dataset mirror
Large-Scale Vietnamese Speech Corpus (LSVSC)	dataset mirror, paper
Vietnamese Multi-Dialect (ViMD)	dataset card, EMNLP 2024 paper
Additional Vietnamese speech data	Original provenance is still being verified

Bud500 was developed by VietAI.

PhoAudiobook and VieNeu-TTS-140h were originally released in speech-synthesis contexts. Their audio–transcript pairs can nevertheless be used as supervised speech recognition data.

InfoRe Dataset №1 primarily contains read speech, while InfoRe Dataset №2 contains audiobook speech.

ViMD contains Vietnamese speech collected from regional radio and television programs.

Some portions of the consolidated collection do not currently preserve enough source-level information to establish their original provenance with confidence. These components are therefore not assigned to a specific external dataset until their origin can be verified.

Similarities in sample count, sampling rate, transcript characteristics, or dataset organization alone are not treated as sufficient evidence of dataset identity.

Repository Structure

The public dataset is organized around metadata files and audio storage:

.
├── README.md
├── address.jsonl
├── merged_dataset_all_sorted.jsonl
├── combined_metadata.jsonl
└── <audio storage>

The JSONL files contain metadata and references to audio samples.

The actual audio data is stored separately in the repository's audio storage/chunks. The exact physical organization should be checked directly from the repository's Files and versions page because storage layout may evolve independently of the metadata schema.

Metadata Format

Each JSONL file contains one UTF-8 encoded JSON object per line.

merged_dataset_all_sorted.jsonl

This file contains metadata for individual audio samples.

Example:

{
  "file": "sample_310808926.wav",
  "text": "làng trúc yên",
  "duration": 1.0,
  "original_sample_rate": 16000
}

Schema:

Field	Type	Description
file	string	Normalized audio filename.
text	string	Transcript corresponding to the audio sample.
duration	number	Audio duration in seconds.
original_sample_rate	integer	Original sampling rate recorded in the metadata, in Hz.

The current version contains approximately 4.07 million metadata records.

The records are ordered by duration in the current metadata release.

original_sample_rate describes the source metadata and should not replace reading the actual sampling rate from the waveform when building an audio-processing pipeline.

combined_metadata.jsonl

This file describes samples constructed from one or more source audio files.

Example:

{
  "combined_id": "combined_0",
  "original_files": [
    "sample_123.wav",
    "sample_456.wav"
  ],
  "transcript": "...",
  "total_duration": 14.5
}

Schema:

Field	Type	Description
combined_id	string	Identifier of the combined sample.
original_files	list[string]	Ordered list of component audio filenames.
transcript	string	Transcript associated with the combined sample.
total_duration	number	Total duration of the combined sample in seconds.

The current metadata contains approximately 1.35 million combined records.

A combined record does not contain a pre-generated waveform inside the JSONL file.

Instead, original_files specifies the ordered sequence of audio files required to construct the sample.

The order must be preserved:

audio_1 → audio_2 → ... → audio_n

Before concatenation, applications should ensure that all waveforms use a compatible sampling rate, channel configuration, and numerical representation.

total_duration describes each individual combined sample. It should not be interpreted as a fixed duration for the entire dataset.

Audio Address Index

address.jsonl provides the mapping between normalized audio filenames and their storage groups.

Example:

{
  "chunk": "chunk_0",
  "id": "sample_110000000.wav"
}

Schema:

Field	Type	Description
id	string	Normalized audio filename.
chunk	string	Storage group containing the corresponding audio file.

The current address index contains approximately 4.14 million entries distributed across 10 chunk identifiers.

The lookup relationship is:

metadata file
    ↓
audio filename
    ↓
address.jsonl
    ↓
chunk
    ↓
audio file

For a combined sample:

combined_metadata.original_files[i]
    ↓
address.id
    ↓
address.chunk
    ↓
audio file

chunk is a logical storage identifier. It should not be interpreted as a URL, archive extension, or complete physical path.

Downloading the Metadata

Install the Hugging Face client:

pip install huggingface_hub datasets

The metadata can be downloaded using:

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="Tran1312/Automatic-Speech-Recognition",
    repo_type="dataset",
    allow_patterns=[
        "*.jsonl",
        "README.md",
    ],
    local_dir="asr_dataset",
)

This example downloads the metadata only.

Audio files can be downloaded separately depending on the subset of the collection required by the application.

For large-scale training, downloading only the necessary audio storage groups can substantially reduce unnecessary disk usage and network transfer.

If authentication is required:

hf auth login

Reading Metadata

Metadata can be processed sequentially without loading the entire file into memory:

import json
from pathlib import Path

root = Path("asr_dataset")

with (root / "combined_metadata.jsonl").open(encoding="utf-8") as f:
    sample = json.loads(next(f))

print(sample["combined_id"])
print(sample["transcript"])
print(sample["original_files"])

The Hugging Face datasets library can also stream the JSONL metadata:

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files="asr_dataset/combined_metadata.jsonl",
    split="train",
    streaming=True,
)

sample = next(iter(dataset))

Here, split="train" is simply required by the generic JSON loading interface to expose the supplied file as a Dataset. It does not describe the original organization of the underlying source corpora.

Locating Audio Files

Given a sample from combined_metadata.jsonl, the corresponding storage chunks can be identified through address.jsonl.

needed = set(sample["original_files"])
locations = {}

with (root / "address.jsonl").open(encoding="utf-8") as f:
    for line in f:
        row = json.loads(line)

        if row["id"] in needed:
            locations[row["id"]] = row["chunk"]

            if len(locations) == len(needed):
                break

for filename in sample["original_files"]:
    print(locations[filename], filename)

This linear scan is suitable for inspecting individual samples.

For model training, repeatedly scanning millions of address records would be inefficient. A persistent lookup structure should instead be built once, for example using:

• an in-memory hash table;
• SQLite;
• LMDB;
• another indexed key-value store.

This changes audio lookup from repeated sequential scans to near-constant-time filename lookup.

Reconstructing Combined Audio

A combined record can be reconstructed conceptually as:

waveforms = []

for filename in sample["original_files"]:
    waveform, sample_rate = load_audio(filename)
    waveform = preprocess_audio(waveform, sample_rate)
    waveforms.append(waveform)

combined_waveform = concatenate(waveforms)

A practical implementation should:

1. Resolve each filename through address.jsonl.
2. Locate the corresponding audio file.
3. Read the actual waveform properties.
4. Convert incompatible sampling rates if necessary.
5. Normalize channel configuration when necessary.
6. Concatenate the waveforms in original_files order.
7. Compare the resulting duration against total_duration as a consistency check.

Dataset Scale

The current metadata contains approximately:

Component	Records
Individual-sample metadata	4,070,788
Combined-sample metadata	1,350,708
Audio-address entries	4,142,885
Storage groups referenced by address index	10

These counts describe different structures and therefore are not expected to match.

For example, one combined sample can reference multiple individual audio files, while the address database may contain audio files that are not referenced by a particular consolidated metadata file.

Provenance

The collection reorganizes data originating from multiple independent Vietnamese speech resources.

The normalization process changes identifiers and storage organization, which means that the current metadata does not always retain a direct mapping to the original record identifier from the source dataset.

Known sources are documented where sufficient evidence is available.

For components whose provenance cannot yet be established reliably, the source is explicitly left unverified rather than inferred from indirect properties such as:

• number of samples;
• sampling rate;
• transcript similarity;
• filename patterns;
• directory structure.

This distinction is important for maintaining accurate dataset documentation and reproducibility.

Usage and Licensing

Rights to the original audio recordings, transcripts, and source datasets remain with their respective authors and organizations.

This repository reorganizes material from multiple datasets and therefore does not automatically establish a single license applicable to every underlying component.

Users should verify the original license, usage conditions, redistribution restrictions, and citation requirements for each source used in their experiments.

In particular, some source datasets impose restrictions on redistribution or limit usage to research and educational purposes.

Citation does not replace compliance with those conditions.

Citation

The consolidated repository can be cited as:

@misc{tran1312_asr_collection,
  author       = {{Tran1312}},
  title        = {Automatic-Speech-Recognition: Vietnamese ASR Collection},
  howpublished = {Hugging Face dataset repository},
  url          = {https://huggingface.co/datasets/Tran1312/Automatic-Speech-Recognition},
  note         = {Accessed 2026-10-06}
}

This citation refers to the consolidated repository itself.

For research publications, the original datasets actually used in an experiment should also be cited.

Source References

@misc{Bud500,
  author = {Anh Pham and Khanh Linh Tran and Linh Nguyen and Thanh Duy Cao and Phuc Phan and Duong A. Nguyen},
  title = {Bud500: A Comprehensive Vietnamese ASR Dataset},
  year = {2024},
  url = {https://github.com/quocanh34/Bud500}
}

@article{tran2024lsvsc,
  author = {Linh Thi Thuc Tran and Han-Gyu Kim and Hoang Minh La and Su Van Pham},
  title = {Automatic Speech Recognition of Vietnamese for a New Large-Scale Corpus},
  journal = {Electronics},
  year = {2024},
  volume = {13},
  number = {5},
  pages = {977},
  doi = {10.3390/electronics13050977},
  url = {https://www.mdpi.com/2079-9292/13/5/977}
}

@inproceedings{vu2025zeroshottexttospeechvietnamese,
  author = {Thi Vu and Linh The Nguyen and Dat Quoc Nguyen},
  title = {Zero-Shot Text-to-Speech for Vietnamese},
  booktitle = {Proceedings of ACL},
  year = {2025},
  url = {https://arxiv.org/abs/2506.01322}
}

@inproceedings{dinh-etal-2024-multi,
  author = {Dinh, Nguyen and Dang, Thanh and Thanh Nguyen, Luan and Nguyen, Kiet},
  title = {Multi-Dialect Vietnamese: Task, Dataset, Baseline Models and Challenges},
  booktitle = {Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  publisher = {Association for Computational Linguistics},
  pages = {7476--7498},
  url = {https://aclanthology.org/2024.emnlp-main.426}
}

References

1. Bud500 / VietAI. Bud500: A Comprehensive Vietnamese ASR Dataset (2024). Dataset.
2. Thi Vu, Linh The Nguyen, Dat Quoc Nguyen. Zero-Shot Text-to-Speech for Vietnamese. ACL, 2025. Paper; PhoAudiobook.
3. pnnbao-ump. VieNeu-TTS-140h. Dataset.
4. VLSP 2020. VLSP 2020 – VinAI – ASR Challenge Dataset. Dataset mirror.
5. InfoRe Technology. Public Dataset №1 (25 hours). Dataset.
6. InfoRe Technology. Public Dataset №2 (Audiobooks). Dataset.
7. Linh Thi Thuc Tran, Han-Gyu Kim, Hoang Minh La, Su Van Pham. Automatic Speech Recognition of Vietnamese for a New Large-Scale Corpus. Electronics, 2024. Paper; dataset.
8. Nguyen Dinh, Thanh Dang, Luan Thanh Nguyen, Kiet Nguyen. Multi-Dialect Vietnamese: Task, Dataset, Baseline Models and Challenges. EMNLP 2024. Paper; dataset.