Deep learning model transcribes italian children’s speech into phonemes
Deep Learning Techniques for Phoneme Recognition in Italian Children' s Speech
Machine Learning
Summary
Speech therapists often struggle to write down exactly what children say, especially when children speak with accents, tones, or speech mistakes. The authors built a computer system called Broca that learned to recognize sounds in Italian children’s speech by first listening to a lot of adult speech and then adapting to a small amount of child speech. This system can turn spoken words into phonemes, the basic units of sound, with good accuracy even using very little child speech data. This could help therapists diagnose speech issues more easily and quickly.
What this means in practice
- •For speech therapists: Automatically transcribe children’s speech into phonemes to assist in diagnosing speech impairments using minimal child-specific training data.
- •For clinical software developers: Create efficient speech assessment tools for pediatric use that handle phonetic variability with low amounts of training data.$Commercial implications: Enables development of speech diagnostic software products for children with speech disorders using limited specialized data.
Authors
Nicola Barbaro, Cristina Gena, Francesco Petriglia, Andrea Meirone, Alessandro Mazzei, Arianna Viotti
Abstract
Speech therapists often face difficulties diagnosing impairments due to the lack of efficient tools for transcribing speech into the International Phonetic Alphabet (IPA). This work addresses this challenge with Broca, a Conformer-based deep learning system pretrained on 8 days of adult speech and fine-tuned on a 165-minute dataset of Italian child speech collected through a range of standardized diagnostic tests for children aged 3.5-6.5. Broca was optimized to handle phonetic variability in children's speech, including tone, accent, and speech errors, and achieved a state-of-the-art weighted Phoneme Error Rate of 13.36% on Italian speech. Remarkably, this performance was obtained using less than three hours of child-specific data, underscoring the model's efficiency and robustness in low-resource clinical settings. This work demonstrates that accurate, vocabulary-independent speech-to-IPA transcription can be achieved with minimal data, paving the way for more accessible, data-efficient tools to support speech assessment and diagnosis.