Predicting speech release from masking through spatial separation in distance ! Alexandre Chabot-Leclerc! Torsten Dau! ! Center for Applied Hearing Research! Technical University of Denmark! ! September 12, 2014! 1 

2 • CRM speech material (Bolia et al., 2000)! On-axis spatial release 
 from masking Westermann et al. (2012) • Speech masker,
 Speech-modulated SSN • Presented over headphones • Compensation of room coloration • SNR measured at the ears 0.5 2 5 10 Maskers distance [m] −14 −12 −10 −8 −6 −4 −2 0 SRT [dB at the ears] Dichotic speech 0.5 2 5 10 Maskers distance [m] −14 −12 −10 −8 −6 −4 −2 0 SRT [dB at the ears] Dichotic speech Diotic speech 0.5 2 5 10 Maskers distance [m] −14 −12 −10 −8 −6 −4 −2 0 SRT [dB at the ears] Dichotic speech Diotic speech Dichotic noise 

Types of masking 3 Energetic! masking! affects audibility Modulation! masking Informational! masking! affects object! formation 

Long-term audibility-based binaural models 4 Binaural speech! intelligibility model (BSIM)! (Beutelmann et al. (2010) JASA) Lavandier and Culling (2010)! (Jelfs et al. (2011)) Better ear Binaural! advantage Binaraul benefit + Band-pass filtering BRIR (left) BRIR (right) 

Long-term binaural models do not predict SRM 5 0.5 2 5 10 Maskers distance [m] −2 0 2 4 6 8 10 12 Spatial release from masking (dB) Data Data (diotic) 0.5 2 5 10 Maskers distance [m] −2 0 2 4 6 8 10 12 Spatial release from masking (dB) Data Data (diotic) Jelfs BSIM 

6 P S+N - P N P N Audio-domain filtering and Hilbert envelope Temporal modulation filtering SNR env Ideal observer Speech + Noise Noise Input SNR [dB] Predicted % Correct Integration across channel The sEPSM The speech-based envelope power spectrum model Jørgensen and Dau (2011) JASA PS+N PN PN 

The sEPSM predicts almost all the monaural SRM 7 0.5 2 5 10 Maskers distance [m] −2 0 2 4 6 8 10 12 Spatial release from masking (dB) Data Data (diotic) Jelfs BSIM 0.5 2 5 10 Maskers distance [m] −2 0 2 4 6 8 10 12 Spatial release from masking (dB) Data Data (diotic) Jelfs BSIM sEPSM 

8 Maskers @ 0.5 m Maskers @ 10 m sEPSM PS+N PN PN PS+N PN PN 

The multi-resolution sEPSM (mr-sEPSM) Short-term calculation of the SNR in the modulation domain Jørgensen and Dau (2013) JASA 9 

0.5 2 5 10 Maskers distance [m] −2 0 2 4 6 8 10 12 Spatial release from masking (dB) Data Data (diotic) Jelfs BSIM sEPSM The mr-sEPSM does not predict improved intelligibility 10 0.5 2 5 10 Maskers distance [m] −2 0 2 4 6 8 10 12 Spatial release from masking (dB) Data Data (diotic) Jelfs BSIM sEPSM mr-sEPSM 

0 2 10 5 10 0 Masker distance [m] Spatial release from masking [dB] Data The mr-sEPSM fails because of the smeared maskers 11 0 2 10 5 10 0 Masker distance [m] Spatial release from masking [dB] Data sEPSM (MM) 0 2 10 5 10 0 Masker distance [m] Spatial release from masking [dB] Data sEPSM (MM) ESII (EM) 0 2 10 5 10 0 Masker distance [m] Spatial release from masking [dB] Data sEPSM (MM) ESII (EM) mr-sEPSM Informational masking? More steady-state 

SRM due to separation in distance is dominated by informational masking •The dominant factor seems to be release from informational masking, due to easier segregation. •Release from long-term modulation masking accounts for a large portion of the SRM. •… but is counteracted by increased masking from the maskers becoming more steady-state. •Release from long-term energetic masking does not contribute to SRM. •Long-term binaural processing does not provide and SRM. 12 

Texte PAMBOX — http://pambox.org An auditory modeling toolbox in Python 

Thank you This research was supported in part by: •The National Science and Engineering Research Council of Canada (NSERC) •Phonak, and •The Technical University of Denmark 14