Extended high frequencies improve phoneme recognition: Evidence from automatic speech recognition in spatial speech mixtures

Extended high frequencies (EHFs; above 8 kHz) improve speech perception in noise, but the underlying mechanisms remain unclear. Debate continues over whether the benefit arises from EHFs providing direct cues to phonemes or more indirectly reflects the listener's cochlear health. To examine whether and how EHFs contribute to phoneme recognition-which is difficult to test in humans given the wide variability in EHF thresholds-this study leveraged an acoustic automatic speech recognition (ASR) model. English speech from the VCTK corpus was resynthesized to create spatial audio where target speech was masked by an interfering talker separated by 20°, 45°, 80°, or 120° azimuth at target-to-masker ratios (TMRs) from +3 to -12 dB. A convolutional neural network bi-directional long short-term memory model was trained to decode target phonemes from cochleagrams of broadband or low-pass filtered (e.g., 8 kHz cutoff) speech. In masked conditions, EHFs improved phoneme recognition across all spatial separations, particularly at TMRs ≤ -9 dB. The improvement was not found in quiet. Removing EHFs disproportionately increased phoneme error rates for consonants, consistent with consonants' spectral concentration at higher frequencies. These findings indicate that EHFs contribute directly to phoneme recognition in adverse conditions, supporting their inclusion in clinical audiometry and ASR system development.