Abstract
In this article, we examine the audio material collected by crowdsourcing methods in the Czech section of the portal forvo.com, which is unique worldwide in terms of the number of languages, audio recordings and users. Our aim is to assess the possible representativeness (with respect to the standard norm of Czech and/or naturalness) of the recordings for phonetic research and also for the purposes of non-native lay users.
We selectively focus on perceptually conspicuous, mostly hypercorrect deviations from natural pronunciation (when the assimilation of voicing or place of articulation does not occur), but we also deal with phenomena that are non-orthoepic yet established in spontaneous speech (such as frequent phonetic reductions and devoicing). These deviations from natural speech behaviour can be attributed chiefly to the unnatural speech situation that lay speakers find themselves in when they are being recorded, and their adherence to the written form of the recorded entries.
Based on the findings, we evaluate the Czech part of the forvo.com portal as being rather unusable for ordinary phonetic research, and as being usable, though only with caution, to a non-native lay user, with no claim to full representativeness.