Phone-like acoustic models (AMs) used in large-vocabulary automatic speech recognition (ASR) systems are usually trained with speech collected from young adult speakers. Using such models, ASR performance may decrease by about 10% absolute when transcribing elderly speech. Ageing is known to alter speech production in ways that require ASR systems to be adapted, in particular at the level of acoustic modeling. In this study, we investigated automatic age estimation in order to select age-specific adapted AMs. A large corpus of read speech from European Portuguese speakers aged 60 or over was used. Age estimation (AE) based on i-vectors and support vector regression achieved mean error rates of about 4.2 and 4.5 years for males and females, respectively. Compared with a baseline ASR system with AMs trained using young adult speech and a WER of 13.9%, the selection of five-year-range adapted AMs, based on the estimated age of the speakers, led to a decrease in WER of about 9.3% relative (1.3% absolute). Comparable gains in ASR performance were observed when considering two larger age ranges (60-75 and 76-90) instead of six five-year ranges, suggesting that it would be sufficient to use the two large ranges only.

--

Automatic speech recognition,Elderly speech,Automatic age estimation,I-vector extraction

• Ciências da Computação e da Informação - Ciências Naturais
• Engenharia Eletrotécnica, Eletrónica e Informática - Engenharia e Tecnologia
• Línguas e Literaturas - Humanidades