A Danish nonsense word corpus for phoneme recognition measurements

Signal processing algorithms intended to improve speech intelligibility are common, e.g. in hearing aids. Due to the spectral and temporal variations between speech segments, the effect of such processing is likely to vary across the signal. To facilitate the analysis of these varying effects, a Danish speech corpus for measuring the perception of individual phonemes was developed.

More than 1150 nonsense words were created according to a common template and audio-visually recorded with two male and two female talkers. Carrier sentences were also recorded. The audio recordings of all words were presented to ten normal-hearing listeners to ensure that phoneme recognition scores under optimal conditions were close to 100%.

Accepted words were compiled into test lists with different characteristics. These lists were evaluated with seven normal-hearing listeners to determine norm data and to investigate memory effects. The speech recognition thresholds (SRTs) of the test lists varied between −7 to −0.9 dB signal-to-noise ratio (SNR) depending on list type and talker.

The recognition score sensitivity to the SNR was 5.2 to 8.9%/dB. Memory effects were small and not significant. The presented speech materials seem well suited for measuring phoneme recognition scores and thus for assessing signal processing effects that vary across speech segments.