Credits
- Developed by: VILAIPHONE Sulixay, Savath Saypadith, Makmeephon Xaysana, SOUNAKHEN Phonesavanh — CEIT
- In collaboration with / supported by: Laligence
- Reviewers: 3 Lao native-speaker annotators
- Base model: Qwen3-ASR-1.7B by the Qwen team, Alibaba (Apache-2.0)
Usage
The checkpoint loads with the qwen_asr package from the Qwen3-ASR project.
The base model does not list Lao as a supported language, so add it before transcribing.
import soundfile as sfimport torchfrom qwen_asr import Qwen3ASRModelfrom qwen_asr.inference.utils import SUPPORTED_LANGUAGES if "Lao" not in SUPPORTED_LANGUAGES: SUPPORTED_LANGUAGES.append("Lao") asr = Qwen3ASRModel.from_pretrained( "LuoYiSULIXAY/qwen3-asr-lao-H", dtype=torch.float16, device_map="cuda:0", max_inference_batch_size=1, max_new_tokens=256,)audio, sr = sf.read("example.wav", dtype="float32") # mono; 16 kHz recommendedprint(asr.transcribe(audio=(audio, sr), language="Lao")[0].text)
Clips of up to about 15 seconds work best. For long recordings, split at pauses first (for example with Silero VAD).
Evaluation
CER is computed without spaces, after both reference and output are normalized to the spelling rules below.
Table with columns: Test set, Clips, Domain, CER (no spaces)| Test set | Clips | Domain | CER (no spaces) |
|---|
| meeting2 | 1,529 | formal Lao meetings (3 videos) | 8.77 |
| test_blind | 247 | mixed; transcribed by a reviewer who saw no model output | 2.56 |
On meeting2 this model is 1.15 CER points better than the same recipe without the 5 h of reviewed meeting speech
(95% bootstrap CI 0.87–1.43). The test_blind clips were selected with an earlier model, so that set favours it and the
differences there are not significant.
Spelling rules
Outputs and references follow one Lao spelling policy: ຳ (not ໍາ); ໜ, ໝ, ຫຼ (not ຫນ, ຫມ, ຫລ, except abbreviations such as
ທຫລ); repeated marks collapsed; ໆ written out. Text that uses other conventions will show extra "errors" that are only
spelling differences.
Limitations
- Trained mostly on read speech and formal meetings. Expect higher error on casual conversation, strong regional
accents, overlapping speakers, music or noise.
- Lao text mixed with Thai or English words, personal and place names, and numbers are the most common errors.
- Part of the meeting2 references were made by reviewers who edited a model's draft. Such labels can lean toward the
model that produced the draft, so treat small differences between models on meeting2 with care.
- The model can occasionally repeat a phrase on long or noisy input.
Intended use
Transcribing Lao speech for research, accessibility, captioning drafts and data collection, with human review where
accuracy matters. Do not use it to identify or monitor people, or as the only record in legal, medical or other
high-stakes settings.
Acknowledgements
We thank CEIT and Laligence for compute, data collection and review support, the Lao reviewers who checked the labels,
and the Qwen team for releasing Qwen3-ASR.
Citation
A paper describing the data pipeline is in preparation; this section will be updated with its citation.
@misc{qwen3_asr_lao_h, title = {Qwen3-ASR Lao (H): a Lao speech recognition model}, author = {VILAIPHONE, Sulixay and Saypadith, Savath and Xaysana, Makmeephon and SOUNAKHEN, Phonesavanh}, note = {CEIT, in collaboration with Laligence}, year = {2026}, url = {https://huggingface.co/LuoYiSULIXAY/qwen3-asr-lao-H}}