Noise statistics drive background- and channel-specific interference in auditory midbrain representations of speech.

Although humans and animals excel at interpreting target sounds in competing noise, sound recognition abilities can vary widely due to statistical characteristics of the interfering background. Yet, how the brain leverages statistical sound information to segregate sounds in noise remains unclear. Here, using many natural auditory textures as background stimuli, we test how the encoding of speech is altered by natural noises in the inferior colliculus (IC) of unanesthetized rabbits. We identify foreground- and background-driven neural response components to sound mixtures and find that the background statistics alter the neural representation of speech in a frequency- and modulation-specific manner. Neural encoding signal-to-noise ratios (SNRs) vary extensively across background categories, frequencies, and modulations, and these influence the encoding of speech independently of the acoustic SNR. This variability is driven by both the background spectrum and modulation statistics, which distort the speech representation and show distinct interference effects in the modulation ranges for rhythm and pitch. Thus, spectrum and modulation statistics critical to speech recognition in noise are reflected in IC neural population activity. These neural response statistics and resulting distortions likely determine the coding fidelity for speech in downstream cortical regions and provide a neural basis for differences in speech intelligibility in real-world noises.