Effects of single-band syllabic amplitude compression on temporal speech information in nonsense syllables and in sentences.

The effects of single-band amplitude compression on the use by subjects with normal hearing of temporal speech information were assessed using speech stimuli that had been processed to remove most spectral information before being compressed. The resulting signal-related-noise (SRN) stimuli isolated the effects of compression on the temporal information in speech by making it impossible for subjects to identify stimulus items on the basis of spectral speech information. Subjects with normal hearing listened to /aCa/ SRN disyllables that had been subjected to single-band compression at various combinations of compression ratio (CR) and time constants (TC). Performance was reduced only in the most severe compression condition (CR = 8; TC = 50), and then only slightly. Additional testing showed that subjects could use both periodicity and compression-overshoot artifactual information--in addition to envelope information--to identify the compressed /aCa/ stimuli. When a list of 10 context-controlled sentences was converted to SRN and compressed at CR = 8 and TC = 50, the ability of subjects with normal hearing to identify the sentences was significantly affected. Results established that (a) subjects with normal hearing differ widely in their abilities to use temporal information for speech identification, even after training; (b) subjects can learn to use both temporal envelope and periodicity information for identification if disyllables, even though; (c) subjects with normal hearing need envelope but not periodicity information to identify SRN sentences in a closed set. These results suggest that single-band compression at CR = 8 and TC = 50 would be undesirable for persons with limited ability to resolve speech spectral information. It is currently not known how less severe compression conditions would affect envelope information in sentences.

Robust adaptive microphone array processing for hearing aids: realistic speech enhancement.

The problem of combining the outputs of an array of microphones as a single input for a hearing aid is investigated. Emphasis is placed on the conservative prediction of realistically achievable performance gains provided by the array over a single microphone. Performance improvement is measured as a change in the speech reception threshold (SRT) between single microphone and multimicrophone conditions. Consistent with previous work, predictions of this change in SRT using intelligibility averaged gain, [symbol: see text] are shown to be good. Consequently, this measure is used, along with changes in signal-to-noise ratios (SNRs), to evaluate array performance. The results presented include the effects of acoustic headshadow, small room reverberation, microphone placement uncertainty, and desired speaker location uncertainty. It is in this context that realistic predictions of speech enhancement provided by robust adaptive microphone array processors are discussed. Performance improvements are demonstrated relative to the "best" single microphone in the array for three types of spatial filters: Fixed, robust block processed, and robust adaptive. The performance of the robust block processed arrays is shown to be attainable with adaptive implementations. One fundamental criterion employed in robust beamformer design directly limits the amount of cancellation of the desired signal that can occur.

A Revised Speech Intelligibility Rating (RSIR) test: listeners with normal hearing.

Two experiments were conducted to examine the intelligibility of 72 passages of connected discourse prepared by Cox and McDaniel in their development of the Speech Intelligibility Rating (SIR) test. Intelligibility was assessed with a method-of-adjustment (MOA) procedure in which listeners adjusted the level of a multi-talker babble until they could just understand 50% of a passage; the measure of intelligibility was the signal-to-babble ratio, dB S/B. The objective was to develop a Revised Speech Intelligibility Rating (RSIR) test that would comprise a large number of equivalent passages that produce reliable intelligibility measures. In experiment 1, the S/B ratio was based on the overall root-mean-square (rms) levels of speech and babble, as represented by the average level of frequent peaks observed on a VU meter. Across all 72 passages, mean intelligibility was -1.43 dB S/B, and the measure of intelligibility for 42 passages was within +/- 0.5 dB of the overall mean for all 72 passages. In experiment 2, the S/B ratio was based on long-term rms levels of speech and babble measured in 16 one-third-octave bands, with center frequencies from 160 to 5000 Hz. In an effort to achieve greater equivalence in intelligibility among passages, the overall rms level of each passage was attenuated by the difference between SB16-band for an individual passage and S/B16-band for a reference passage. Mean intelligibility across all 72 passages was -8.06 dB, and the measure of intelligibility was within +/- 0.5 dB of the overall mean for 64 of the 72 passages. For those 64 passages, the 95% critical difference for five MOAs was 0.72 dB, which corresponds to an estimated percentage critical difference of 10.8%.

Frequency-importance functions for words in high- and low-context sentences.

The relative importance and absolute contributions of various spectral regions to speech intelligibility under conditions of either neutral or predictable sentential context were examined. Specifically, the frequency-importance functions for a set of monosyllabic words embedded in a highly predictive sentence context versus a sentence with little predictive information were developed using Articulation Index (AI) methods. Forty-two young normal-hearing adults heard sentences presented at signal-to-noise ratios from -8 to +14 dB in a noise shaped to conform to the peak spectrum of the speech. Results indicated only slight differences in 1/3-octave importance functions due to differences in semantic context, although the crossovers differed by a constant 180 Hz. Methodological and theoretical aspects of parameter estimation in the AI model are discussed. The results suggest that semantic context, as defined by these conditions, may alter frequency-importance relationships in addition to the dynamic range over which intelligibility rises.