Robust brain inspired prediction of speech recognition accuracy across many words, voices, and natural background sounds

Although humans generally excel at recognizing speech in noisy natural environments, performance varies widely due to differences in the background acoustics and individual talker attributes such as voice quality, intonation, and pitch. The underlying neural and acoustic factors of this perceptual variability remain elusive, making it an unresolved challenge to predict when speech recognition will succeed or fail. Here, we characterize human speech recognition across a wide range of natural conditions using speech spoken by 326 diverse male and female talkers of multiple ages embedded in 33 unique background noises. Across this vast soundscape, speech recognition is strongly driven by the spectrum and modulation statistics of the noise. Yet, articulatory speech features - including fundamental and formant frequencies - exert categorically distinct modulatory effects on recognition accuracy across the talkers age and gender, and specific words. This perceptual variability is captured by a low-dimensional statistical representation of sounds based on computations that emulate intermediate stages of sound processing and feature extraction in auditory midbrain. A model utilizing these sound summary statistics accurately predicts individual listeners single-trial recognition behavior, accounting for performance variance across individual voices, words, and backgrounds. Thus, speech-in-noise perception across diverse natural conditions relies largely on a simple set of spectrotemporal summary statistics known to be represented by central neural populations. Significance StatementAlthough humans can easily maintain conversations amidst environmental noise and voice variations, this capacity is severely degraded by hearing impairment and challenges modern speech recognition technologies. This study evaluates human speech recognition across a broad spectrum of environmental noises and diverse speaker demographics. We show that a brain-inspired auditory representation of sounds extracts critical statistical sound features driving human behavior, precisely predicting variations in perceptual accuracy across distinct voices, words, and environmental backgrounds.