How to Control and Utilize Crowd‐Collected Speech
Ian McGraw, Joseph H. Polifroni · 2013
This chapter describes several efforts at collecting audio data using mechanical turk (MTurk). The authors address two overarching characteristics of speech resources that researchers may be particularly concerned with: quality and variety. The first experiments, initially presented in McGraw et al. , cover the simple collection of read speech with Amazon Mechanical Turk (MTurk), using the standard mechanisms. The final set of experiments in the chapter describes an attempt to move beyond paid crowdsourcing and collect data from a large, preexisting user base. In this final set of experiments, the authors place speech data collection in the context of games with a purpose (GWAP). In doing so, the authors examine the implications of placing a speech recognizer in the loop in the game scenario. The chapter ends with a discussion of how various automated mechanisms for verification of transcription quality affected word error rate (WER).