Overcoming biases in state-of-the-art automatic speech recognition for young children with speech disorders

State-of-the-art models like Whisper and GPT-4o face significant challenges in recognizing and processing child speech, particularly disordered speech, due to the limited availability of annotated child speech data and inherent demographic biases. Our research aims to bridge this gap by adapting these models for more accurate classification and recognition of disordered vs. non-disordered child speech. Using the SEED corpus, we leveraged advanced data augmentation, transfer learning, and parameter-efficient fine-tuning, achieving significant Word Error Rate (WER) reductions: from 57.5% to 10.3% for disordered speech and from 33.6% to 6.04% for non-disordered speech. Our novel approach also addresses demographic biases, ensuring equitable ASR performance across diverse groups. By overcoming current methodological constraints and emphasizing fairness and precision, our research not only advances ASR capabilities but also establishes a benchmark for inclusive technological advancement. These developments are crucial for practical applications in speech therapy and education, showcasing our commitment to impactful research in this domain. Future plans include enhancing model robustness by incorporating self-supervised learning techniques and exploring domain adaptation to ensure adaptability across varied speech environments. This comprehensive approach represents a significant step forward in signal processing, providing robust solutions to longstanding challenges in child speech recognition.