This paper explores the use of conversational speech question and answer systems in the challenging context of public spaces in slums. A major part of this work is a comparison of the source and speed of the given responses; that is, either machine-powered and instant or human-powered and delayed. We examine these dimensions via a two-stage, multi-sited deployment. We report on a pilot deployment that helped refine the system, and a second deployment involving the installation of nine of each type of system within a large Mumbai slum for a 40-day period, resulting in over 12,000 queries. We present the findings from a detailed analysis and comparison of the two question-answer corpora; discuss how these insights might help improve machine-powered smart speakers; and, highlight the potential benefits of multi-sited public speech installations within slum environments. CCS CONCEPTS• Information systems → Speech / audio search; • Humancentered computing → Field studies; Interaction paradigms; Sound-based input / output; Interaction techniques.

No abstract

No abstract

Swarachakra is an Abugida text input keyboard available in 12 Indian languages. We enhanced an accessible version of Swarachakra Marathi with speech input. However, speech input could be error-prone, and especially so for languages where speech recognition technologies are new. Such errors could either slow the user down due to the need for editing, or go unnoticed, leading to high uncorrected error rates. We therefore conducted a withinsubject empirical study to compare the user performance of keyboard-only input method with keyboard+speech input method with 11 novice visually impaired users. We found that keyboard+speech input was almost 11 times faster, reaching 182 characters per minute, and had a lower uncorrected error rate than the keyboard-only input, and in spite of having higher corrected error rates. Though we used a wide variety of phrases in our study, we observed that all phrases were faster on average with the keyboard+speech input method. To the best of our knowledge, ours is the first empirical study to evaluate the performance of speech enabled text input in Marathi for visually impaired people. This is the highest reported speed by visually impaired users in any Indian language.