Phoneme-group specific octave-band weights in predicting speech intelligibility

Abstract

In an earlier study we derived robust frequency-weighting functions for prediction of the intelligibility of short nonsense words. These frequency-weighting functions are applied for prediction of intelligibility such as with the speech transmission index (STI). Six independent experiments revealed essentially similar frequency-weighting functions for the prediction of the nonsense word scores with respect to signal-to-noise ratio and gender [Speech Communication 28 (1999) 109]. Although the frequency weightings do not vary significantly for signal-to-noise ratio or gender, other studies have shown that using different types of speech material (i.e., nonsense words, phonetically balanced words and connected discourse) resulted in quite different frequency-weighting functions. This may be related to the distribution of specific phonemes in the test material. In order to obtain a more generic description of the frequency weighting, four relevant groups of phonemes were identified. In situations with reduced intelligibility, a small confusion rate of the phonemes between the groups and a high confusion rate of the phonemes within each group was observed. For each group a specific frequency-weighting function and a good prediction of the phoneme group scores could be obtained. It was shown that from these (weighted) phoneme group scores, word scores could be predicted with a prediction accuracy of ca. 4% (this corresponds to a signal-to-noise ratio of about 1 dB). Hence, this method provides a more generic way to predict intelligibility scores for different types of speech material.

Introduction

Octave-band weighting functions represent the contribution of each octave band to the intelligibility of a speech signal. In an earlier study it was found that these weighting functions are robust for signal-to-noise ratio and gender of the speaker (Steeneken, 1992; Steeneken and Houtgast, 1999). However, experiments based on different types of speech material showed quite different frequency-weighting functions (French and Steinberg, 1947, Steeneken and Houtgast, 1980, Steeneken and Houtgast, 1999, Pavlovic, 1987, Studebaker et al., 1987, Duggirala et al., 1988). This may be related to the specific distribution of phonemes in the test material, since the frequency-weighting functions do vary significantly according to phonetic content. In Fig. 1 typical weighting functions are given that are derived from two standards on the objective prediction of speech intelligibility (speech transmission index, STI, described by IEC 60286-16, 1998; speech intelligibility index, SII, by ANSI S3.5, 1997) and for consonants and vowels from a study by Steeneken (1992). Fig. 1 shows a large difference between the curves for consonants and for vowels. The weighting function for the vowels has a maximum for the contributions in the 0.5 and 2 kHz octave band. Consonants and equally balanced CVC words (words of the type consonant–vowel–consonant with an equally balanced phoneme distribution) cover a wider frequency range (125 Hz–8 kHz). Obviously the octave-band contributions depend on the type of speech considered.

This is in agreement with the differences found for the effect of various types of distortions on the intelligibility of vowels and consonants. This is illustrated in Fig. 2, showing a scatter diagram of the initial consonant score versus vowel scores for male speech in 78 transmission conditions with various combinations of bandwidth and signal-to-noise ratio.

For diagnostic assessment of speech communication systems it is of interest to consider not only the overall performance derived from a specific intelligibility test (i.e., related to the speech material) but also to identify the performance for specific phonemes or groups of phonemes (Miller and Nicely, 1955). For example, the standard for the SII recommends six groups of frequency-weighting factors for prediction of different subjective intelligibility measures.

From an experiment with CVC-word tests we obtained confusions among consonants and among vowels for many different transmission conditions. For the consonants, a clustering of three groups of consonants with many intra-group confusions was found (fricatives, plosives, and vowel-like consonants). The phonemes within each of these groups show a quite similar response for various types of degradation, and confusions are mainly between phonemes within each group (Steeneken, 1992). This is given in Table 1 for 17 representative Dutch initial consonants obtained for male speech and 26 different combinations of band-pass limiting.

In the table the phonemes (SAMPA notation, 1987) that show many mutual confusions are grouped together. This results in a clustering of the plosives (p, t, k, b, d), the fricatives (f, s, v, z, x), and the vowel-like consonants (m, n, l, R, w, j, h). Some confusions are found between phonemes belonging to different clusters: f, s → p, t; v, z → w, j, h, and b → w.

A similar representation for 15 Dutch vowels derived from the same set of transmission conditions did not show a systematic clustering. Hence, for the determination of (Dutch) phoneme-specific octave-band weights, in total four clusters of phonemes are likely to be considered: fricatives, plosives, vowel-like consonants and vowels. These clusters of frequently used phonemes (>2% for Dutch language) consist of 17 initial consonants and 15 vowels. For reasons of simplicity the final consonants were not considered separately as these consonants (11) are mainly a sub-set of the initial consonants.