Modeling the masking of formant transitions in noise

Formant transitions are critical for identifying the place of articulation for consonants. If these transitions are masked by background noise, perceptual confusions can occur. To better understand the masking of formant transitions, masking thresholds were measured for tone glides and single-formant trajectories of varying frequency extent (0-3 ERBs), duration (10, 30 and 100 ms), and center frequency (.5, 1.5, 3.5 kHz). Results show that thresholds are independent of frequency extent and only depend on the duration and center-frequency of the transition. A novel, time-frequency detection model, fit to previous noise-in-noise masking experiments (JASA 101, 2789-802 (1997)), is proposed which can predict these data. Although there have been several studies on the noise-masking of stationary signals such as tones (e.g. Garner and Miller, 1947; Plomp and Bouman, 1959), few studies have measured the masking of non-stationary stimuli. Collins and Cullen (1978) measured the masked thresholds of both rising and falling tone glides with frequency extents of 200 to 700 Hz and 1200 to 1700 Hz, and durations between 10 and 120 ms. Tone thresholds were about 4 dB lower than thresholds for glides and between durations of 10 and 35 ms, rising glides were more detectable than falling glides. Nabelek (1978) measured glide thresholds over a wider range of frequency extents and durations and only found substantial differences between glide and tone thresholds at the largest frequency extents and shortest durations. None of these studies, however, has specifically measured the masking of formant transitions. It is not clear whether the thresholds for these multiple-harmonic stimuli are similar to those of glides. Further, no model has been developed to predict the noise-masking of any type of non-stationary stimuli. With this in mind, masking experiments were conducted using glides and formant-transitions of varying center-frequency, duration, and frequency extent. A novel, time-frequency detection model was developed that can predict the masked thresholds of these non-stationary stimuli. Traditional models of simultaneous masking (e.g. Fletcher, 1940, Patterson, 1976) have focused on the masking of long-duration, narrowband stimuli. In these models, the signal and noise are filtered through the " optimal " auditory filter centered around the signal's center frequency and if the filtered SNR is greater than a certain threshold, then the sound is heard. However for glides and formant transitions, the " optimal " filter is constantly changing, and thus, there may need to be a mechanism that combines information across multiple filter outputs. One …