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

Area of Science:

• Auditory Neuroscience
• Speech Processing
• Machine Learning

Background:

• Extended high frequencies (EHFs) above 8 kHz enhance speech perception in noise, but their precise role is debated.
• It is unclear if EHFs offer direct phonemic cues or indirectly reflect cochlear health.

Purpose of the Study:

• To investigate the direct contribution of EHFs to phoneme recognition in challenging listening conditions.
• To overcome human variability in EHF thresholds by utilizing an acoustic automatic speech recognition (ASR) model.

Main Methods:

• An acoustic ASR model (CNN-BiLSTM) was trained on speech from the VCTK corpus, resynthesized with spatial audio and masker interference at various spatial separations and target-to-masker ratios (TMRs).
• The model decoded phonemes from cochleagrams of both broadband and low-pass filtered (8 kHz cutoff) speech to assess the impact of EHFs.

Main Results:

• EHFs significantly improved phoneme recognition in masked conditions across all spatial separations, particularly at TMRs of -9 dB and below.
• No significant improvement from EHFs was observed in quiet listening conditions.
• Removing EHFs disproportionately increased phoneme error rates for consonants, aligning with their higher spectral concentration.

Conclusions:

• EHFs provide direct acoustic cues that enhance phoneme recognition, particularly in speech-masked scenarios.
• These findings support the integration of EHFs into clinical audiology assessments and the development of advanced ASR systems.