Measuring Speech Recognition With a Matrix Test Using Synthetic Speech.

Only 48 young normal-hearing students aged 18 to 25 years were tested, so it does not show how heterogeneous participant groups would perform. The study used one German matrix test and one female text-to-speech voice from one system, so other languages, voices, or text-to-speech systems may not give the same result. The text-to-speech system was chosen using subjective ratings in quiet, while the speech recognition tests were done in noise, so the best synthetic voice for noisy listening is not proven. An expert listener manually selected the most natural-sounding synthetic word combinations, and this hand sorting took approximately 4 hours. The synthetic version required a worse threshold than natural speech, and recognition scores were significantly lower for synthetic speech at all three tested signal-to-noise ratios. Training did not fully stabilize during the first session; plateau thresholds still differed from measurement-phase thresholds by 0.9 dB for synthetic speech and 0.7 dB for natural speech. Effort reduction was limited to producing the speech material, because the speech material still needs evaluation for equivalent intelligibility of test lists.

Funding was provided by the European Regional Development Fund and the Niedersächsisches Vorab of the Lower Saxony Ministry for Science and Culture.

Do not conclude that synthetic speech can replace natural speech in clinical speech audiometry. This was tested only in 48 young normal-hearing students, with one German matrix test and one female text-to-speech voice, and the synthetic version required a 0.5 dB SNR worse threshold than natural speech.