Anna Warzybok

Polish sentence matrix test for speech intelligibility measurement in noise

Abstract The purpose of this study was to develop the Polish sentence matrix test (PSMT) to measure intelligibility of speech presented against a background noise. The PSMT consists of five columns containing: 10 names, 10 verbs, 10 numerals, 10 adjectives, and 10 nouns. Since each word was available as a separate sound file, it was possible to generate different sentences by juxtaposing randomly selected words taken from respective columns. This approach allows 100 000 unique sentences of a fixed grammatical structure to be generated. The speech reception threshold (SRT), i.e. the signal-to-noise ratio (SNR) providing 50% speech intelligibility and S50, the slope of an intelligibility function at the SRT point, were shown to be −9.6 dB and 17.1 %/dB, respectively. Note that in this study dB is regarded as dB SNR, otherwise reference is given. PSMT was also evaluated using an adaptive 1-up/ 1-down staircase procedure in investigations with and without participation of an experimenter. No significant differences were shown for SRTs obtained in these investigations. Sumario El propósito de este estudio fue desarrollar la prueba matriz de frases en polaco (PSMT) para medir la inteligibilidad de lenguaje presentado contra ruido de fondo. El PSMT consiste de cinco columnas que contienen: 10 nombres, 10 verbos, 10 numerales, 10 adjetivos y 10 palabras. Dado que cada palabra estaba disponible como un archivo de sonido separado, fue posible generar diferentes frases por medio de la yuxtaposición aleatoria de palabras seleccionadas tomadas de las respectivas columnas. Este enfoque permite la generación de 100.000 frases singulares con una estructura gramatical fija. El umbral de recepción del lenguaje (SRT), p.e., la tasa de señal-ruido (SNR) brindando el 50% de inteligibilidad del lenguaje, y el S50, la pendiente de una función de inteligibilidad en el punto del SRT, fueron –9.6 dB y 17.1%/dB, respectivamente. Note que en este estudio los dB corresponden a dB SNR, a menos que se brinde otra referencia. También se evaluó el PSMT usando un procedimiento adaptativo en escalera de 1 hacia arriba y 1 hacia abajo, en investigaciones con y sin un experimentador. No se encontraron diferencias significativas para los SRT obtenidos en estas investigaciones.

The multilingual matrix test: Principles, applications, and comparison across languages: A review

Objective: A review of the development, evaluation, and application of the so-called ‘matrix sentence test’ for speech intelligibility testing in a multilingual society is provided. The format allows for repeated use with the same patient in her or his native language even if the experimenter does not understand the language. Design: Using a closed-set format, the syntactically fixed, semantically unpredictable sentences (e.g. ‘Peter bought eight white ships’) provide a vocabulary of 50 words (10 alternatives for each position in the sentence). The principles (i.e. construction, optimization, evaluation, and validation) for 14 different languages are reviewed. Studies of the influence of talker, language, noise, the training effect, open vs. closed conduct of the test, and the subjects’ language proficiency are reported and application examples are discussed. Results: The optimization principles result in a steep intelligibility function and a high homogeneity of the speech materials presented and test lists employed, yielding a high efficiency and excellent comparability across languages. The characteristics of speakers generally dominate the differences across languages. Conclusion: The matrix test format with the principles outlined here is recommended for producing efficient, reliable, and comparable speech reception thresholds across different languages.

Matrix sentence intelligibility prediction using an automatic speech recognition system.

Objective: The feasibility of predicting the outcome of the German matrix sentence test for different types of stationary background noise using an automatic speech recognition (ASR) system was studied. Design: Speech reception thresholds (SRT) of 50% intelligibility were predicted in seven noise conditions. The ASR system used Mel-frequency cepstral coefficients as a front-end and employed whole-word Hidden Markov models on the back-end side. The ASR system was trained and tested with noisy matrix sentences on a broad range of signal-to-noise ratios. Study sample: The ASR-based predictions were compared to data from the literature (Hochmuth et al, 2015) obtained with 10 native German listeners with normal hearing and predictions of the speech intelligibility index (SII). Results: The ASR-based predictions showed a high and significant correlation (R² = 0.95, p < 0.001) with the empirical data across different noise conditions, outperforming the SII-based predictions which showed no correlation with the empirical data (R² = 0.00, p = 0.987). Conclusions: The SRTs for the German matrix test for listeners with normal hearing in different stationary noise conditions could well be predicted based on the acoustical properties of the speech and noise signals. Minimum assumptions were made about human speech processing already incorporated in a reference-free ordinary ASR system.

Development of the Russian matrix sentence test

Objective: To develop the Russian matrix sentence test for speech intelligibility measurements in noise. Design: Test development included recordings, optimization of speech material, and evaluation to investigate the equivalency of the test lists and training. For each of the 500 test items, the speech intelligibility function, speech reception threshold (SRT: signal-to-noise ratio, SNR, that provides 50% speech intelligibility), and slope was obtained. The speech material was homogenized by applying level corrections. In evaluation measurements, speech intelligibility was measured at two fixed SNRs to compare list-specific intelligibility functions. To investigate the training effect and establish reference data, speech intelligibility was measured adaptively. Study sample: Overall, 77 normal-hearing native Russian listeners. Results: The optimization procedure decreased the spread in SRTs across words from 2.8 to 0.6 dB. Evaluation measurements confirmed that the 16 test lists were equivalent, with a mean SRT of − 9.5 ± 0.2 dB and a slope of 13.8 ± 1.6%/dB. The reference SRT, − 8.8 ± 0.8 dB for the open-set and − 9.4 ± 0.8 dB for the closed-set format, increased slightly for noise levels above 75 dB SPL. Conclusions: The Russian matrix sentence test is suitable for accurate and reliable speech intelligibility measurements in noise.

Effects of spatial and temporal integration of a single early reflection on speech intelligibilitya)

In order to study the interaction between the intelligibility advantage in rooms due to the presence of early reflections and due to binaural unmasking, a series of speech reception threshold experiments was performed employing a single reflection of the frontal target speech source as a function of its delay ranging from 0 to 200 ms. The direction of the reflection and the spatial characteristic of the interfering noise (diotic, diffuse, or laterally localized) were varied in the experiments. For the frontal reflection, full temporal integration was observed for all three noise types up to a delay of at least 25 ms followed by gradual intelligibility decay at longer delays. At 200 ms delay the reflection introduced additional intelligibility deterioration. For short delays, intelligibility was not reduced when the reflection was spatially separated from the direct sound in the diffuse and lateral noise conditions. A release from the deterioration effect at 200 ms delay was found for all spatially separated reflections. The suppression of a detrimental reflection was symmetrical in diffuse noise, but azimuth-dependent in lateral noise. This indicates an interaction of spatial and temporal processing of speech reflections which challenges existing binaural speech intelligibility models.

A study on speech quality and speech intelligibility measures for quality assessment of single-channel dereverberation algorithms

This paper reports on the evaluation of several objective quality measures for predicting the quality of the dereverberated speech signals. The correlations between subjective quality assessment for single-channel dereverberation techniques and objective speech quality as well as speech intelligibility measures are analyzed and discussed. Six different single-channel dereverberation algorithms were included in the evaluation to account for different types of distortions. The subjective quality was assessed along the four attributes reverberant, colored, distorted and overall quality following the recommendations of ITU-T P.835. The objective measures included system-based, i.e. channel-based, as well as signal-based measures.

Subjective speech quality and speech intelligibility evaluation of single-channel dereverberation algorithms

In this contribution, six different single-channel dereverberation algorithms are evaluated subjectively in terms of speech intelligibility and speech quality. In order to study the influence of the dereverberation algorithms on speech intelligibility, speech reception thresholds in noise were measured for different reverberation times. The quality ratings were obtained following the ITU-T P.835 recommendations (with slight changes for adaptation to the problem of dere-verberation) and included assessment of the attributes: reverberant, colored, distorted, and overall quality. Most of the algorithms improved speech intelligibility for short as well as long reverberation times compared to the reverberant condition. The best performance in terms of speech intelligibility and quality was observed for the regularized spectral inverse approach with pre-echo removal. The overall quality of the processed signals was highly correlated with the attribute reverberant or/and distorted. To generalize the present outcomes, further studies are needed to account for the influence of the estimation errors.

An Italian matrix sentence test for the evaluation of speech intelligibility in noise.

Objective: Development of an Italian matrix sentence test for the assessment of speech intelligibility in noise. Design: The development of the test included the selection, recording, optimization with level adjustment, and evaluation of speech material. The training effect was assessed adaptively during the evaluation measurements with six lists of 20 sentences, using open- and closed-set response formats. Reference data were established for normal-hearing listeners with adaptive measurements. Equivalence of the test lists was investigated using the open-set response format at three signal-to-noise ratios (SNRs). Study sample: A total of 55 normal-hearing Italian mother-tongue listeners. Results: The evaluation measurements at fixed SNRs resulted in a mean speech reception threshold (SRT) of − 7.3 ± 0.2 dB SNR and slope of 13.3 ± 1.2 %/dB. The major training effect of 1.5 dB was observed for the first two consecutive measurements. Mean SRTs of − 6.7 ± 0.7 dB SNR and − 7.4 ± 0.7 dB SNR were found from the third to the sixth adaptive measurement for open- and closed-set test response formats, respectively. Conclusions: A good agreement has been found between the SRTs and slope and those of other matrix tests. Since sentences are difficult to memorize, the Italian matrix test is suitable for repeated measurements.

Comparing Binaural Pre-processing Strategies III Speech Intelligibility of Normal-Hearing and Hearing-Impaired Listeners

A comprehensive evaluation of eight signal pre-processing strategies, including directional microphones, coherence filters, single-channel noise reduction, binaural beamformers, and their combinations, was undertaken with normal-hearing (NH) and hearing-impaired (HI) listeners. Speech reception thresholds (SRTs) were measured in three noise scenarios (multitalker babble, cafeteria noise, and single competing talker). Predictions of three common instrumental measures were compared with the general perceptual benefit caused by the algorithms. The individual SRTs measured without pre-processing and individual benefits were objectively estimated using the binaural speech intelligibility model. Ten listeners with NH and 12 HI listeners participated. The participants varied in age and pure-tone threshold levels. Although HI listeners required a better signal-to-noise ratio to obtain 50% intelligibility than listeners with NH, no differences in SRT benefit from the different algorithms were found between the two groups. With the exception of single-channel noise reduction, all algorithms showed an improvement in SRT of between 2.1 dB (in cafeteria noise) and 4.8 dB (in single competing talker condition). Model predictions with binaural speech intelligibility model explained 83% of the measured variance of the individual SRTs in the no pre-processing condition. Regarding the benefit from the algorithms, the instrumental measures were not able to predict the perceptual data in all tested noise conditions. The comparable benefit observed for both groups suggests a possible application of noise reduction schemes for listeners with different hearing status. Although the model can predict the individual SRTs without pre-processing, further development is necessary to predict the benefits obtained from the algorithms at an individual level.

How much does language proficiency by non-native listeners influence speech audiometric tests in noise?

Objective: The current study investigates the extent to which the linguistic complexity of three commonly employed speech recognition tests and second language proficiency influence speech recognition thresholds (SRTs) in noise in non-native listeners. Design: SRTs were measured for non-natives and natives using three German speech recognition tests: the digit triplet test (DTT), the Oldenburg sentence test (OLSA), and the Göttingen sentence test (GÖSA). Study sample: Sixty-four non-native and eight native listeners participated. Results: Non-natives can show native-like SRTs in noise only for the linguistically easy speech material (DTT). Furthermore, the limitation of phonemic-acoustical cues in digit triplets affects speech recognition to the same extent in non-natives and natives. For more complex and less familiar speech materials, non-natives, ranging from basic to advanced proficiency in German, require on average 3-dB better signal-to-noise ratio for the OLSA and 6-dB for the GÖSA to obtain 50% speech recognition compared to native listeners. Conclusions: In clinical audiology, SRT measurements with a closed-set speech test (i.e. DTT for screening or OLSA test for clinical purposes) should be used with non-native listeners rather than open-set speech tests (such as the GÖSA or HINT), especially if a closed-set version in the patients own native language is available.