Perceptual compensation for effects of reverberation in speech identification: a computer model based on auditory efferent processing

Amy V. Beeston, Guy J. BrownInterspeech 2010

TL;DR

This paper presents a computer model that mimics how humans identify speech in reverberant environments. By simulating auditory efferent processing, the model shows how the brain might dynamically adjust its sensitivity to compensate for the blurring effects of echoes, offering a novel explanation for perceptual constancy in speech.

What Problem Does This Paper Address?

Human speech perception is remarkably robust, allowing us to understand speech even in noisy or echo-filled rooms. This robustness, known as perceptual constancy, means our brains can compensate for acoustic distortions like reverberation. While previous research has shown that listeners can adjust their perception based on the acoustic environment, the underlying neural mechanisms are not fully understood. This paper investigates whether a low-level auditory mechanism, specifically efferent processing that controls the dynamic range of auditory nerve responses, can explain this perceptual compensation for reverberation in speech identification.

Methods

The authors developed a computer model that extends a previous auditory model by incorporating an efferent feedback loop. The core of the model simulates the auditory nerve's response to sound. This involves:

  • Afferent Pathway: The acoustic signal is processed through a bank of dual-resonance nonlinear (DRNL) cochlear filters, mimicking the cochlea's frequency analysis. This is followed by a model of inner hair cell transduction, which converts the filtered signal into a simulated auditory nerve firing rate.
  • Efferent Pathway: A feedback loop monitors the "mean-to-peak ratio" (MPR) of the simulated auditory nerve response. A higher MPR indicates more reverberation. Based on this MPR, an efferent signal is generated to attenuate the input to the nonlinear pathway of the DRNL filters. This attenuation effectively controls the dynamic range of the auditory nerve's response.
  • Speech Identification: The model uses a template-matching approach. Spectro-temporal excitation patterns (STEPs) derived from the simulated auditory nerve responses are compared to pre-defined templates for the words "sir" and "stir" using a mean squared error (MSE) metric.

The model was tuned on specific "near-near" and "far-far" conditions from existing human listener data and then tested on other conditions without further tuning.

Datasets

The paper evaluates the model against human listener data from experiments conducted by Watkins [1]. These experiments involved presenting listeners with a continuum of speech stimuli between "sir" and "stir" under various reverberation conditions (e.g., "near-far," "far-far"). The paper does not introduce new datasets but uses existing perceptual data for validation.

Key Insights and Ideas

  • Efferent Control as a Mechanism for Constancy: The model demonstrates that efferent suppression, by dynamically adjusting the auditory system's dynamic range, can effectively compensate for reverberation without needing to explicitly measure or process "reverberation tails." This suggests a simpler, low-level explanation for perceptual constancy.
  • MPR as a Reverberation Cue: The mean-to-peak ratio (MPR) of the auditory nerve response serves as a robust indicator of reverberation. As reverberation increases, the MPR increases, signaling the need for efferent attenuation.
  • Dynamic Range Adjustment: Efferent attenuation shifts the rate-level curves of auditory nerve fibers, effectively reducing sensitivity to lower-level sounds and making the system less susceptible to the "filling-in" effect of reverberant energy. This allows subtle acoustic cues, like the /t/ in "stir," to become more apparent.
  • Time-Reversed Reverberation Abolishes Compensation: A crucial finding is that when reverberation is time-reversed (meaning the echoes arrive before the direct sound), the perceptual compensation effect is abolished, both in human listeners and the model. This supports the idea that the model's MPR-based mechanism, which is sensitive to the temporal characteristics of the signal, is key.

Key Results

The computer model qualitatively matched human listener data across various conditions, including:

  • Perceptual Constancy: The model successfully replicated the shift in category boundary observed in human listeners when the context reverberation matched the test word reverberation.
  • Effect of Time-Reversed Reverberation: The model showed that time-reversing the context reverberation abolished the perceptual compensation effect, aligning with human data. This was achieved by the model's MPR metric reflecting a reduced MPR in such conditions, leading to less efferent attenuation.

Limitations and Future Work

The authors acknowledge several limitations:

  • Monaural Processing: The current model is monaural, while human hearing is typically binaural, which might play a role in real-world reverberant environments.
  • Uniform Efferent Attenuation: Efferent attenuation is applied equally across all frequency regions. Recent research suggests within-frequency-channel compensation might be more accurate. The authors suggest extending the model to implement frequency-specific efferent feedback.
  • Simplified Hair Cell Model: The model uses a simplified hair cell transduction mechanism that does not account for adaptation. The authors note that efferent suppression might also play a role in "release from adaptation," which could be important for speech intelligibility in noise.

Future work includes developing a reverberation-robust ASR front-end based on these human-like perceptual mechanisms and evaluating the model on more naturalistic speech material and a wider range of consonant confusions.

Summary updated

BibTeX
@inproceedings{beeston2010perceptual,
  author = {Amy V. Beeston and Guy J. Brown},
  title = {{Perceptual compensation for effects of reverberation in speech identification: a computer model based on auditory efferent processing}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure