On the Difficulties of Automatic Speech Recognition for Kindergarten-Aged Children

Abstract

Automatic speech recognition (ASR) systems for children have lagged behind in performance when compared to adult ASR. The exact problems and evaluation methods for child ASR have not yet been fully investigated. Recent work from the robotics community suggests that ASR for kindergarten speech is especially difficult, even though this age group may benefit most from voice-based educational and diagnostic tools. Our study focused on ASR performance for specific grade levels (K-10) using a word identification task. Grade-specific ASR systems were evaluated, with particular attention placed on the evaluation of kindergarten-aged children (5-6 years old). Experiments included investigation of grade-specific interactions with triphone models using feature space maximum likelihood linear regression (fMLLR), vocal tract length normalization (VTLN), and subglottal resonance (SGR) normalization. Our results indicate that kindergarten ASR performs dramatically worse than even 1st grade ASR, likely due to large speech variability at that age. As such, ASR systems may require targeted evaluations on kindergarten speech rather than being evaluated under the guise of "child ASR." Additionally, results show that systems trained in matched conditions on kindergarten speech may be less suitable than mismatched-grade training with 1st grade speech. Finally, we analyzed the phonetic errors made by the kindergarten ASR.

A Frequency Normalization Technique for Kindergarten Speech Recognition Inspired by the Role of fo in Vowel Perception

“…All ASR systems were trained with the Kaldi Speech Recognition Toolkit [38]. This setup is based on our previous investigation of child ASR [7]. Each system was tested using all 1,654 word utterances from the kindergarten speech data.…”

“…A number of past studies has explored the performance of child ASR systems [6,7,8]. Notably, a recent study by Kennedy et al explored the performance of child ASR in robotics applications [6].…”

“…The results revealed insufficient performance (15%-20% error rates) for even basic digit recognition tasks. Another study by Yeung and Alwan discovered the impact that a single year age difference can have on child ASR performance using deep neural network (DNN) hidden Markov model (HMM) ASR systems [7]. Those results highlighted the importance of targeting a specific age group (especially the kindergarten age group) when training and testing ASR systems, rather than grouping several age groups together.…”

“…For example, formant frequencies and the fundamental frequency (fo) are known to depend on the age of the child [9,10,11]. As the mismatch in age between training and testing data becomes larger, ASR performance degrades rapidly [7].…”

“…In this study, we propose an fo-based normalization procedure inspired by the tonotopic distances. Kindergarten is chosen as the target age group due to its relevance for educational and human-robot interface (HRI) applications [6], as well as its poor ASR performance [7]. We show that the tonotopic distances can be reformulated and generalized to motivate an fo-based frequency normalization approach for the entire speech spectrum.…”