Supra-Segmental Changes in Speech Production as a Result of Spectral Feedback Degradation: Comparison with Lombard Speech

Abstract

Perturbations to acoustic speech feedback have been typically localized to specific phonetic characteristics, for example, fundamental frequency (F0) or the first two formants (F1/F2), or affect all aspects of the speech signal equally, for example, via the addition of background noise. This paper examines the consequences of a more selective global perturbation: real-time cochlear implant (CI) simulation of acoustic speech feedback. Specifically, we examine the potential similarity between speakers' response to noise vocoding and the characteristics of Lombard speech. An acoustic analysis of supra-segmental characteristics in speaking rate, F0 production, and voice amplitude revealed changes that paralleled the Lombard effect in some domains but not others. Two studies of speech intelligibility complemented the acoustic analysis, finding that intelligibility significantly decreased as a result of CI simulation of speaker feedback. Together, the results point to differences in speakers' responses to these two superficially similar feedback manipulations. In both cases we see a complex, multi-faceted behavior on the part of talkers. We argue that more instances of global perturbation and broader response assessment are needed to determine whether such complexity is present in other feedback manipulations or if it represents a relatively rare exception to the typical compensatory feedback response.

Keywords: Lombard speech; Speech motor planning; acoustic feedback; cochlear implant; intelligibility.

Figure 1.

Signal transformation used to simulate cochlear implant acoustic processing. Natural sound (upper left) is spectrally analyzed (upper right) to determine the signal strength within particular bandwidths of frequencies. Envelopes for each of these frequency “channels” are then applied to broadband noise (lower right), shaping it into a spectrally-degraded version of the original signal (lower right, spectrum; lower left, spectrogram).

Figure 2.

Mean durations for sentences produced by speakers in the experimental (n = 9; solid line) and control (n = 3; dashed line) groups (color online).

Figure 3.

Average F0 of sentences. F0 changed significantly across Epochs in the experimental group (solid line), but not in the control (dashed line) (color online).

Figure 4.

Average sentence F0 variability (standard deviation) produced by female speakers in the experimental (n = 7; solid line) and control (n = 3; dashed line) groups. No change across epochs was observed in the control group, while experimental talkers’ F₀ variability was significantly lower in Epoch 2 than in Epochs 1 or 3 (color online).

Figure 5.

Mean voice amplitudes for sentences. Amplitudes changed significantly across Epochs in the experimental group (solid line), but not in the control (dashed line) (color online).

Figure 6.

Rates at which listeners chose speech tokens produced with normal acoustic feedback as “easier to understand” under acoustic transformation than speech from Epoch 2. Responses to experimental talkers in black (solid line) and to controls in gray (dash–dot line). Only judgments of experimental talkers’ speech differed significantly from chance (50%, dashed line).