Benchmarking Fine-Tuned Speech-to-Text Models for Lao: CTC, Transducer, and Generative Architectures on a Common Test Set
Lao is a low-resource tonal language written in an abugida without spaces between words, which complicates both the modeling and the evaluation of automatic speech recognition (ASR). This paper presents a controlled benchmark of four open speech-to-text models fine-tuned for Lao from four different pretrained families: the Omnilingual ASR CTC checkpoint omniASR-CTC-300M-v2, wav2vec 2.0 XLSR-53 with a CTC head, a NeMo FastConformer-Transducer initialized from a Thai model, and the generative Qwen3-ASR-0.6B. All models were trained on the same 112.4-hour corpus of 31,996 clips and scored on the same 1,583 held-out clips with a Lao-safe text normalizer that preserves tone marks and vowel signs. The omniASR model reaches the lowest character error rate (CER), 7.94%, followed by the FastConformer-Transducer at 8.44% with about one third of the parameters, XLSR-53 at 8.82%, and Qwen3-ASR at 11.47% at roughly ten times the per-clip latency. We also document four practical pitfalls: Lao combining marks missing from a base vocabulary, NFKC decomposition of Lao characters by a subword tokenizer, a widely used normalizer that silently deletes tone marks during scoring, and an interaction of regularizers that stalls CTC training. All four models are publicly released.
Authors
- Khieovongphachanh Vimontha
- Bounnhavong Chindavone
- Luangxaysana Khanthanou
- PHOUMMAVONG Phonepadith
Institutions
- National University of Laos (LA)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.5281/zenodo.24391427
- Primary Topic
- Speech Recognition and Synthesis
- Type
- article
- Field-Weighted Citation Impact
- 0.00