Workshop

Workshop program and overall results

Workshop

The SPEAR Challenge Workshop was held online on Thursday 30th of March. Below you can find videos of the presentations and a summary of the results.


Programme

UK time (BST)

Session 1 (3:00 pm - 4:15 pm)

Session 2 (4:30 pm - 5:45 pm)


Results

Teams

Results are presented using the annoymised algorithm labels. The teams contributing each algorithm are given in the table below.

ID Team name Algorithm name Affiliation Description
alg_A - passthrough -  
alg_B - baseline -  
alg_C audifon aud-1 audifon GmbH & Co. KG paper
alg_D audifon aud-2 audifon GmbH & Co. KG paper
alg_E IEM iem University of Music and Performing Arts Graz paper
alg_F IIPHIS iip-1 Sogang University paper
alg_G IIPHIS iip-2 Sogang University paper
alg_H IIPHIS iip-3 Sogang University paper
alg_I IIPHIS iip-4 Sogang University paper
alg_J UIUC uiuc-1 Univeristy of Illinois paper
alg_K UIUC uiuc-2 Univeristy of Illinois paper
alg_L UIUC uiuc-3 Univeristy of Illinois paper
alg_M UIUC uiuc-4 Univeristy of Illinois paper
alg_N UIUC uiuc-5 Univeristy of Illinois paper
alg_O ICL icl-1 Imperial College London  
alg_P ICL icl-2 Imperial College London  


Metrics

Each segment of enhanced audio was compared to the direct path in-ear signals using a large number of objective metrics (see Evaluation metrics). Here we show the difference in metric between each algorithm and the unprocessed noisy mixture (alg_A - passthrough).

MBSTOI plot
STOI (L) plot
STOI (R) plot
ESTOI (L) plot
ESTOI (R) plot
PESQ (L) plot
PESQ (R) plot
PESQ_NB (L) plot
PESQ_NB (R) plot
SegSNR (L) plot
SegSNR (R) plot
fwSegSNR (L) plot
fwSegSNR (R) plot
SI_SDR (L) plot
SI_SDR (R) plot
SDR (L) plot
SDR (R) plot
ISR (L) plot
ISR (R) plot
SAR (L) plot
SAR (R) plot
HASPI (L) plot
HASPI (R) plot


Listening tests

Each pair of algorithms was compared by at least 10 participants. In each experiment 20 stimuli were presented.

Using a Bradley-Terry model as a hierarchical generalized model under a Bayesian framework to analyze the pair comparison data, the probability of the second algorithm in a specified ordered pair is dependent on the stimulus and the participant.

In the following figure, points represent median probability estimate of the 2nd algorithm being preferred; bars represent the 89% highest density credible interval surrounding the median estimate; asterisks indicate statistically significant probabilities (different from 0.5).

pairwise comparison plot

Considering all the pairwise data, the model predicts the number of times a particular algorithm would win in a fully balanced set of pairwise comparison trials. From this we obtain the rank order as shown below.

rank order plot