Stuttering-Aware Automatic Speech Recognition for Indonesian Language

TL;DR

This work tackles the scarcity of stuttered Indonesian speech data by generating a synthetic stuttering dataset through rule-based and LLM-driven text perturbations, then producing stuttered audio via TTS. A pre-trained Indonesian Whisper model is finetuned on this synthetic data, yielding improved recognition of stuttered speech while maintaining performance on fluent speech. The results confirm that targeted synthetic exposure can meaningfully enhance ASR robustness to disfluencies in a low-resource language. The approach offers a scalable path toward more inclusive speech technology and highlights directions for validating with real stuttered data and extending to other languages.

Abstract

Automatic speech recognition systems have achieved remarkable performance on fluent speech but continue to degrade significantly when processing stuttered speech, a limitation that is particularly acute for low-resource languages like Indonesian where specialized datasets are virtually non-existent. To overcome this scarcity, we propose a data augmentation framework that generates synthetic stuttered audio by injecting repetitions and prolongations into fluent text through a combination of rule-based transformations and large language models followed by text-to-speech synthesis. We apply this synthetic data to fine-tune a pre-trained Indonesian Whisper model using transfer learning, enabling the architecture to adapt to dysfluent acoustic patterns without requiring large-scale real-world recordings. Our experiments demonstrate that this targeted synthetic exposure consistently reduces recognition errors on stuttered speech while maintaining performance on fluent segments, validating the utility of synthetic data pipelines for developing more inclusive speech technologies in under-represented languages.

Figures (1)

• Figure 1: Dataset generation workflow using rule-based and LLM approaches, followed by audio generation using text-to-speech model