Precedence Effect: Listening experiment data and statistical analyses spreadsheets

Excel spreadsheet with the data collected from a subjective, quantitative speech in noise test (SINT) conducted in the Listening Room at the University of Salford in March 2020. 

The listening experiment tested how the psychoacoustic phenomenon of the precedence effect can be utilised with augmented loudspeaker arrays in an object-based audio paradigm to improve speech intelligibility in the home environment. A practical application of this research will be in the implementation of media device orchestration, i.e. the creation of low-cost, ad-hoc loud speaker arrays using commonly found devices, such as mobile phones, laptop computers, tablets, smart speakers, and so on, to spatialise audio in the home.

This speech-in-noise test was conducted under controlled conditions. With audio reproduced by one of three different arrays of loudspeakers in a given trial, subjects listened to spoken sentences played simultaneously with noise. They were tasked with correctly identifying target words. Correct word scores collated and converted to word recognition percentages act as a quantifiable proxy for speech intelligibility. After confirming that they fulfilled the criterion for use, data were statistically analysed using 2-way RMANOVA.

The three configurations of loudspeaker arrays were:

* L1R1_base (a two-loudspeaker control condition): 

a stereo pair of front left and front right loudspeakers at -/+30 degrees azimuth and 2m distance from the listener position; speech + noise reproduced by both loudspeakers.

* L1R1C2 (three loudspeakers): 

L1R1_base + an additional (AUX) loudspeaker in the true front centre position (0 degrees azimuth and 1.7m distance from listener position) reproducing just speech.

* L1R1R2 (three loudspeakers): 

L1R1_base + an AUX loudspeaker in the right-hand position (+90 degrees azimuth and 1.7m distance from listener position) reproducing just speech.

For the array configurations with the three loudspeakers, the precedence effect was initiated by applying a 10 ms delay to the speech signal reproduced by the AUX loudspeaker, such that the sound source (first arrivals) would still be perceived as being from the phantom centre between the L1 and R1 loudspeakers, but with a boost to the speech signal. The relevant equalisation (EQ) was applied to the speech signal for the C2 and R2 AUX loudspeakers though to maintain the same perceived comb filtering effects for all three loudspeaker array configurations. 

Analysis of the results is provided in the PhD thesis by P. Demonte.

-----------------------------------------------------------------------

Spreadsheet pages:

* Read Me - provides a more in-depth explanation of the independent variables tested

* Raw data - as collected in the speech-in-noise test. The columns denote: subject number; trial number; audio files playing from each loudspeaker in a trial; loudspeaker array configuration; masking noise type; Harvard speech corpus list and sentence number; spoken sentence played; the five target words in each sentence; the sentence as heard and noted by the subject; correct word score applied (out of a total of 5 per trial); correct word ratio.

* CWR_all - correct word percentages collated for each subject for each combination of independent variables, and the corresponding studentized residuals as a quality check for outliers.

* NormalDistTest - criteria for normal distribution (Shapiro-Wilk test)

* 2-way RMANOVA_16subjects - Mauchley's test of Sphericity, and Tests of Wtihin-Subjects Effects (2-way RMANOVA)

* SimpleMainEffects - analysis of the conditional effects

* Participants_MainTest - anonymised data collated from the subjects via a short pre-screening questionnaire: age; gender, handedness (left or right); confirmation of subjects as native English speakers, and whether or not they are bi-/multilingual in case of outliers.