Introduction

Qwen3-ASR Lao (H) just shipped: LuoYiSULIXAY/qwen3-asr-lao-H (created 2026-11-05 ~03:14 UTC). It is a merged fp16 checkpoint of Qwen3-ASR-1.7B fine-tuned with LoRA on Lao read speech, human-reviewed meeting clips and synthetic Lao — then released under Apache-2.0.

The base Qwen3-ASR card does not list Lao as a supported language. This drop is the practical unlock: append "Lao" to SUPPORTED_LANGUAGES and transcribe.

What Shipped

  • Vendor / product: Qwen3-ASR Lao (H) — VILAIPHONE Sulixay, Savath Saypadith, Makmeephon Xaysana, SOUNAKHEN Phonesavanh (CEIT), with Laligence support
  • Modality: open-weight automatic speech recognition (Lao / ພາສາລາວ)
  • Open vs closed: open weights, Apache-2.0 (inherits Qwen3-ASR base licence)
  • Runnable today: qwen_asr / Qwen3ASRModel.from_pretrained on CUDA (fp16)
  • Concrete fact: meeting2 held-out CER 8.77 (1,529 formal meeting clips, no-space scoring); test_blind CER 2.56 (247 clips); ~35.7 h real speech + ~106.9 h synthetic (VoxCPM2) in the mix

Outputs follow a documented Lao spelling policy (ຳ not ໍາ; ໜ / ໝ / ຫຼ forms; collapsed repeats) so CER stays comparable across systems.

Run It

pip install qwen-asr soundfile torch

import soundfile as sf
import torch
from qwen_asr import Qwen3ASRModel
from 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 preferred
print(asr.transcribe(audio=(audio, sr), language="Lao")[0].text)

Prefer clips around ≤15 s; split longer audio at pauses (Silero VAD works fine).

Notes for makers

  • Credit CEIT + Laligence reviewers and the Qwen team — three native Lao annotators checked meeting labels
  • Expect higher error on casual chat, strong regional accents, overlap, music or noise; Thai/English code-mix and names are the usual failure modes
  • Part of meeting2 references were human edits of a model draft — treat tiny CER gaps between siblings carefully
  • Soft-holds this cycle still carry: Ghana ZipVoice / Moore TTS 401, finanfa Fon TTS NC+WIP, Eddycrack Cover Maker primary Nov 2 (>48h), Fish×Alibaba Drama 3 still coming soon
  • No Gen CTA: Lao ASR does not map to the magazine Gen product loop

Conclusion

For Lao speech work that needed a Qwen3-ASR path the base model never advertised, Lao (H) is today’s open unlock — merge is done, Apache-2.0, meeting CER on the card, and qwen_asr runs it once you register the language string.

—Anabel ♡