Access the Sota Khmer ASR dataset

Free access, approved automatically. Please tell us who you are and how you will use it.

Log in or Sign Up to review the conditions and access this dataset content.

Sota Khmer ASR training data

By Ing Muyleang · collected, cleaned and curated by Ing Muyleang.

All the speech used to train Sota សោតៈ, a small Khmer + English speech recognition model, in one place. Real people's voices: do not use this data to impersonate anyone. Public samples: Sota-ASR-Khmer-Data-Preview · Website: https://muyleanging.github.io/sota-web/

1,593 h of Khmer + English speech, collected by Ing Muyleang · 947,571 clips · 16 kHz mono FLAC (PCM 16), clips ≤ 20 s.

Split Clips Hours Speakers
train 937,733 1,573.8 ≈18,195
dev 2,638 5.2 8
test (Sota test set) 6,513 12.0 16
fleurs_test (outside test) 655 2.4 –
ddd_test (outside test) 32 0.1 2

Test, dev, fleurs_test and ddd_test are never trained on: their speakers and sentences are removed from train.

Files

  • shards/asr-NNNNN.tar (49 × ~2 GB): the audio. Extract all into one folder; audio paths are relative to it.
  • train.jsonl, dev.jsonl, test.jsonl, fleurs_test.jsonl, ddd_test.jsonl: one row per clip. summary.json: the numbers above.

Download

pip install -U huggingface_hub soundfile
hf auth login          # first click "Get access" on this page (approved at once)
hf download muyleanging/Sota-ASR-Khmer-Data --repo-type dataset --local-dir data
cd data && for t in shards/*.tar; do tar -xf "$t"; done   # creates the audio folders

Only the lists, or only a few shards to try it first:

hf download muyleanging/Sota-ASR-Khmer-Data --repo-type dataset --local-dir data --include "*.jsonl" "summary.json"
hf download muyleanging/Sota-ASR-Khmer-Data --repo-type dataset --local-dir data --include "shards/asr-0000[0-1].tar"

Read it in Python:

import json, soundfile as sf
rows = [json.loads(l) for l in open("data/train.jsonl", encoding="utf-8")]
r = rows[0]
audio, sr = sf.read("data/" + r["audio"])   # 16 kHz mono
print(r["text"], r["duration"], sr)

Splits: train.jsonl, dev.jsonl, test.jsonl, fleurs_test.jsonl, ddd_test.jsonl. Each row's audio path is relative to the folder where you extracted the tars.

Columns

Column Meaning
audio path inside the shards (shard says which tar)
text the original transcript
target the training label: NFC, numbers spelled out, ឲ→ឱ, every character known to the Sota tokenizer
duration, speaker, source, kind (real/synthetic), latin (has English), style (v6 only)

Cleaning

Audio resampled to 16 kHz mono and peak-limited; silent/unreadable files removed. Labels with unspelled numbers or unknown characters removed; clips too fast for CTC removed; no training sentence appears in any test set.

License and citation

CC BY-SA 4.0 (because parts of the real speech are CC BY-SA). You may use, share and adapt it, also commercially, if you credit Ing Muyleang, cite the dataset below, keep the credits of the original sources (footnote), and share changes under the same license.

@misc{ing2026sotaasrdata,
  author    = {Ing, Muyleang},
  title     = {Sota Khmer ASR dataset: 1,593 hours of Khmer and English speech},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/muyleanging/Sota-ASR-Khmer-Data}
}

© 2026 Ing Muyleang.

Licenses: real speech from NXKMER TTS 450 (CC BY 4.0), OpenSLR 42 (CC BY-SA 4.0), DDD khmer-speech-dataset (CC BY-SA 4.0), LibriSpeech (CC BY 4.0), FLEURS (CC BY 4.0), km-speech-corpus, iAny, YODAS2. Some voices made with VoxCPM / VoxCPM2 (OpenBMB).

Downloads last month
-