An overview of high-resource automatic speech recognition methods and their empirical evaluation in low-resource environments

Abstract

Deep learning methods for Automatic Speech Recognition (ASR) often rely on large-scale training datasets, which are typically unavailable in low-resource environments (LREs). This lack of sufficient and representative training data poses a significant challenge for applying ASR systems in specific domains categorized as LREs. In this paper, we provide a comprehensive overview and empirical analysis of state-of-the-art deep learning techniques for ASR, which are primarily designed for high-resource environments (HREs). Our aim is to explore their potential effectiveness in LRE settings. We focus on identifying key factors that influence the adaptation of HRE models to LRE tasks. To this end, we survey advanced deep learning models and conduct a comparative evaluation of their performance in LRE contexts. Additionally, we propose that pre-training ASR models on HRE datasets, followed by domain-specific fine-tuning on LRE data, can significantly enhance performance in data-scarce settings. Using LibriSpeech and WSJ as our HRE datasets, we evaluate these models on two LRE datasets: UASpeech for dysarthria speech and iCUBE, our novel human–robot interaction dataset. Our systematic experiments, involving varying dataset sizes for pre-training, demonstrate the efficacy of combining pre-training and fine-tuning strategies to improve recognition accuracy in LREs. • This paper evaluates state-of-the-art ASR models trained on high-resource data for LREs. • We demonstrate that deeper model structures are not efficient for low-resource ASR tasks. • Training data from another domain cannot improve ASR accuracy in low-resource settings. • Pre-training on large datasets and fine-tuning with in-domain data improves LRE ASR tasks.