ASR Error Correction in Low-Resource Burmese with Alignment-Enhanced Transformers using Phonetic Features

2025-11-26

AI Summary

This study addresses the lack of systematic research on automatic speech recognition (ASR) error correction for low-resource Burmese. We propose the first Transformer-based correction framework that jointly models phonemes and alignment information. Our approach innovatively incorporates International Phonetic Alphabet (IPA) representations and forced-alignment features, while fusing outputs from multiple ASR backends to enable joint word- and character-level optimization within a sequence-to-sequence architecture. Experiments show that our method significantly reduces the average word error rate (WER) from 51.56% to 39.82% on unaugmented data, and improves the chrF++ score to 0.627 (+0.0406), substantially outperforming baselines. It maintains robust performance even with data augmentation (WER = 43.59%). This work establishes a transferable, phoneme-aware modeling paradigm for ASR error correction in low-resource languages.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Learning & Optimization for NLP
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

This paper investigates sequence-to-sequence Transformer models for automatic speech recognition (ASR) error correction in low-resource Burmese, focusing on different feature integration strategies including IPA and alignment information. To our knowledge, this is the first study addressing ASR error correction specifically for Burmese. We evaluate five ASR backbones and show that our ASR Error Correction (AEC) approaches consistently improve word- and character-level accuracy over baseline outputs. The proposed AEC model, combining IPA and alignment features, reduced the average WER of ASR models from 51.56 to 39.82 before augmentation (and 51.56 to 43.59 after augmentation) and improving chrF++ scores from 0.5864 to 0.627, demonstrating consistent gains over the baseline ASR outputs without AEC. Our results highlight the robustness of AEC and the importance of feature design for improving ASR outputs in low-resource settings.

Problem

• Correcting ASR errors in low-resource Burmese language
• Integrating phonetic and alignment features in Transformers
• Improving word- and character-level accuracy of ASR outputs

Innovation

• Transformers with IPA phonetic features for Burmese ASR
• Alignment-enhanced models for low-resource speech correction
• Feature integration strategy reducing word error rates