Intervention? Ari Septayuda Arnold Nicholas Fadhlan Putra Harry Andiko Pratama Erkki Svante Nyfors Discussion The Challenges of Feedback
The provision of meaningful and consistent feedback to children is
a key priority. Discussion The Challenges of Feedback (Cont.)
Analyses even from a decade ago identified that ASR technology is
not able to provide corrective analytic feedback and can only provide evaluative feedback. Corrective= analyze and semantically interpret degrees of deviation between elicited speech and a given speech target. Evaluatif= indicating whether a given speech target was reached or not. For an ASR module operating in a speech intervention context, if it is unable to produce a lexicographic candidate for a given speech input, then the system must decide whether this error is an instance of speech error due to poor speech or a recognition error due to poor system performance. Discussion The Challenges of Feedback (Cont.)
Researchers have determined that systems rely on abstract
visualizations as feedback do not seem to work well for children. Neri et al. have identified a major problem with providing comparable waveforms, a popular form of feedback (e.g., Speech Viewer II), to the user. Although showing target and input waveforms in alignment can be motivating for the user (i.e., to try to emulate the target waveform by modifying their pronunciation), it does not necessarily lead to behavior modification (i.e., correction of articulation). Discussion Shifting from Analysis to Elicitation Motivation
A key observation here is that incorporating ASR can make the
computer system responsive to speech even if it does not provide detailed feedback. In the context of speech intervention, even rudimentary feedback can be of value, since it can motivate children to try multiple repetitions of words and phrases. This approach has been used in a remarkable study by Mitra et al., in the context of accent reduction, which found that even rudimentary feedback was helpful. Discussion Shifting from Analysis to Elicitation Motivation (Cont.)
Scientific researchers in this domain may quickly conclude that
the need for high-quality corrective analytic feedback clearly motivates the need for further work into automated speech analysis. And such work is ongoing. For instance, efforts in the area of acoustic training, which entails to the process of recording representative speech samples from a user to create or to augment an acoustic database. Discussion Shifting from Analysis to Elicitation Motivation (Cont.)
A point that needs mention is that the findings discussed here is
based on an assumption about the lexical unit being short and the language having a relative low ratio of morphemes to words (e.g., as in English); we expect the results to generalize to other moderately analytic languages, but may not generalize more broadly, for example to synthetic languages (e.g., Greek).