When Noise Reduction Meets Spatial Listening: Behavioral Evaluation of AI-Enhanced Mandarin Speech

Introduction: AI speech enhancement may change cues used in spatial listening. We examined signal-to-noise ratio (SNR), spatial separation, and processing strategy in Mandarin word recognition.

Methods: Thirty-four Mandarin listeners completed three within-participant experiments using four-alternative sentence-final disyllabic-word recognition. Experiment 1 established a no-AI baseline in target speech, competing speech, and white noise: SNR (-12/-8 dB) by target-competing-talker separation (0/15/60 degrees). Experiment 2 crossed no AI versus DeepFilterNet atten6 with the same separations at -12 dB SNR and a 0 dB target-to-competing-talker ratio. Experiment 3 compared no AI, full-band AI, and hybrid low-frequency-preserving AI at 0/15-degree target-noise separation without competing speech (-16 dB SNR), examining processing strategy in a noise-only masking scene. AI processed each ear independently after spatialisation. The hybrid combined original low frequencies and AI-processed high frequencies around a 1500-Hz crossover. Accuracy was the proportion of completely correct target-word responses. Trial-level logistic mixed models accounted for participants, items, and target talker; pairwise contrasts were Holm-adjusted. Short-time objective intelligibility (STOI) was calculated for the stimuli.

Results: Spatial separation improved accuracy in all experiments (all p<0.001). Experiment 1 accuracy averaged 38.5%, 71.3%, and 75.1% at 0, 15, and 60 degrees, respectively. Experiment 2 showed no significant AI effect (AI/no AI: 58.6%/60.7%; p=0.856). In Experiment 3, low-frequency-preserving AI outperformed full-band AI (72.9% versus 61.0%; odds ratio=2.20, adjusted p=0.028); neither differed significantly from no AI (68.4%; adjusted p=0.365 and p=0.178, respectively). Despite higher accuracy, low-frequency-preserving AI had lower STOI than full-band AI (0.460 versus 0.483).

Conclusions: Processing strategy affected recognition, and STOI did not track the behavioural ordering of the two AI strategies. Human-centred enhancement evaluation should include spatialised behavioural testing.