How Much Data is Enough ? Fine-Tuning Large Pre-Trained ASR Models for Low-Resource Burmese Language

Abstract

Automatic Speech Recognition (ASR) systems achieve near-human accuracy for high-resource languages such as English and Mandarin, but performance lags for low-resource and non-alphabetic languages due to data scarcity and linguistic complexity. Burmese, a tonal, syllable-based language, poses additional ASR challenges due to its complex phonology, script structure, and limited transcribed speech data. This study investigates how fine-tuning data volume affects Burmese ASR performance across a controlled range of training sizes. Three pre-trained ASR models: Whisper, W2v-BERT 2.0, and Wav2Vec2-XLS-R were fine-tuned on a combined Burmese corpus drawn from OpenSLR-80, Google FLEURS, and UCSY News. Four training conditions (6, 8, 10, and 12 hours) were used to assess data size vs. Word Error Rate (WER). W2v-BERT 2.0 performed best, achieving 32.70% WER at 12 h and crossing the 40% reference point with just 10 h. We use WER < 40% throughout as a fixed reference point for comparing the data requirements of different architectures, not as a claim of deployment-grade accuracy. These findings demonstrate that Burmese ASR can be developed with limited training data.