Optimizing the Quality of Synthetically Generated Pseudowords for the Task of Minimal-Pair Distinction

Training the distinction of vowel lengths or learning to differentiate between voiced and voiceless plosive sounds in form of minimal pair differentiation is one of the treatments fostering phonological awareness for people with reading and/or writing disabilities. While text-to-speech systems can automatically generate minimal pairs (e.g., bin and pin), the quality of the pronunciation of pseudowords is not always optimal. We present a novel approach for using text-to-speech tools to artificially generate the pronunciation of German pseudowords, which is evaluated in a crowdsourcing task of the discrimination of minimal pairs. While the input for generating audio files for real words is provided as plaintext, the audio files for pseudowords are generated from the SAMPA transcription, a computer-readable phonetic alphabet, of their real-word counterparts. The task of selecting the correct word from a minimal pair of a pseudoword and its lexical counterpart was completed equally successfully when a pseudoword was generated by our method or pronounced by a human (χ2(1) = 2.43, p = .119).