Mismatched Crowdsourcing based Language Perception for Under-resourced Languages

Mismatched crowdsourcing is a technique for acquiring automatic speech recognizer training data in under-resourced languages by decoding the transcriptions of workers who don’t know the target language using a noisy-channel model of cross-language speech perception. All previous mismatched crowdsourcing studies have used English transcribers; this study is the first to recruit transcribers with a different native language, in this case, Mandarin Chinese. Using these data we are able to compute statistical models of cross-language perception of the tones and phonemes from transcribers based on phone distinctive features and tone features. By analyzing the phonetic and tonal variation mappings and coverages compared with the dictionary of the target language, we evaluate the different native languages’ effect on the transcribers’ performances.