Intelligibility Test Data from HI and NH Listeners

Compiled by Yuxuan Wang


This folder contains test data from hearing-impaired (HI) and normal-hearing (NH) listeners on HINT utterances in two noises and algorithm-segregated noisy utterances. The test results are described in the following paper:


README

The data can be downloaded in one ZIP file (about 370 MB). The main idea behind the segregation algorithm is to estimate the ideal binary mask (IBM) using deep neural network classifiers.

Our experiment uses 260 Hearing In Noise Test (HINT) sentences, a speech shaped noise (SSN), and a 8-talker babble noise as the training and test material. The HINT sentences are labeled as 6001.wav through 6280.wav. The SSN and the babble noises are labeled as n5.wav and n6.wav, respectively. The clean utterances and the noises can be found in folder "./data/clean_utterance" and "./data/noise". All signals have a sampling frequency of 16 kHz. In our experiment, 100 sentences (6001~6070 and 6251~6280) are used for classifier training, and 160 sentences (6071~6230) are used for testing.

-- To avoid mixing with a fixed noise segment, we cut each noise into 10 short segments. Each mixture is created by first randomly permuting the noise segments to form a new version of the noise, which is then mixed with a clean utterance at a specified SNR (-8, -5, -2, or 0 dB). To enlarge the training set for improving generalization, each training utterance gets mixed multiple times (6 in our experiments). The training mixtures and labels (IBMs) can be found in "./data/training". The test mixtures and labels can be found in "./data/testing". If a file has an extension of ".dat", it means that this file is written as an ASCII file, which can be read by using MATLAB function "load(filename)".

-- We have included the separation results in "./results/objective_metrics.xlsx". Specifically, we list classification accuracy as well as two objective intelligibility metrics, the HIT minus the FA rate (HIT-FA) [1] and the Short-Time Intelligibility Measure (STOI) [2], for each stimulus. The classification accuracy and HIT-FA are obtained by comparing an estimated IBM with the IBM. The STOI results are obtained by using clean speech as the reference signal. The "Mixture STOI" column denotes the STOI results of unprocessed mixtures, whereas "Processed STOI" denotes those of processed by the algorithm.

-- The human intelligibility scores are listed in "./results/subject_scores.xlsx".

[1] G. Kim, Y. Lu, Y. Hu, and P. Loizou, "An algorithm that improves speech intelligibility in noise for normal-hearing listeners," Journal of the Acoustical Society of America, pp. 1486-1494, 2009.

[2] C. Taal, R. Hendriks, R. Heusdens, and J. Jensen, "An algorithm for intelligibility prediction of time-frequency weighted noisy speech," IEEE Trans. Audio, Speech, Lang. Process., pp. 2125-2136, 2011.