The Auditory Modeling Toolbox

Applies to version: 0.10.0

View the code

Go to function

TAAL2011 - The Short-time objective intelligibility measure (STOI)

Usage

d = taal2011(sigclean, sigproc, fs);

d = stoi(sigclean, sigproc, fs) returns the output of the Short-Time Objective Intelligibility (STOI) measure described in Taal et. al. (2010) & (2011), where sigclean and sigproc denote the clean and processed speech, respectively, with sample rate fs measured in Hz. The output d is expected to have a monotonic relation with the subjective speech-intelligibility, where a higher d denotes better intelligible speech. See Taal et. al. (2010) & (2011) for more details.

The model consists of the following stages:

  1. Removal of silent frames. Frames (of length 512) of the input signals that have an energy of 40 dB less than the most energetic frame are removed.
  2. Expansion of the signals into a Fourier filterbank with a Hanning window length of 25ms and 256 channels covering the the frequency range from 0 to 5 kHz. The energy of the bands are then summed into third-octaves
  3. The output d is computed by a correlation process. See the referenced papers for more details.

Examples:

The following example shows a simple comparison between the intelligibility of a noisy speech signal and the same signal after noise reduction using a simple soft thresholding (spectral subtraction):

% Get a clean and noisy test signal
[f,fs]=cocktailparty;
Ls=length(f);
f_noisy=f+0.05*pinknoise(Ls,1);

% Simple spectral subtraction to remove the noise
a=128; M=256; g=gabtight('hann',a,M);
c_noise   = dgtreal(f,g,a,M);
c_removed = thresh(c_noise,0.01);
f_removed = idgtreal(c_removed,g,a,M);
f_removed = f_removed(1:Ls);

% Compute the STOI of noisy vs. removed
d_noisy   = taal2011(f, f_noisy, fs)
d_removed = taal2011(f, f_removed, fs)

This code produces the following output:

d_noisy =

    1.0000


d_removed =

    0.9915

The original STOI model can be downloaded from http://msp.ewi.tudelft.nl/content/short-time-objective-intelligibility-measure This is a standalone version not depending on LTFAT and AMToolbox, and licensed under a different license, but the models are functionally equivalent.

References:

C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen. A Short-Time Objective Intelligibility Measure for Time-Frequency Weighted Noisy Speech. In Acoustics Speech and Signal Processing (ICASSP), pages 4214--4217. IEEE, 2010.

C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen. An Algorithm for Intelligibility Prediction of Time-Frequency Weighted Noisy Speech. IEEE Transactions on Audio, Speech and Language Processing, 19(7):2125--2136, 2011.