Phoneme Compression: processing of the speech signal and effects on speech intelligibility in hearing-Impaired listeners
André Goedegebure · RePub (Erasmus University Rotterdam) · 2005
Hearing-aid users often continue to have problems with poor speech understanding in difficult acoustical conditions. Another generally accounted problem is that certain sounds become too loud whereas other sounds are still not audible. Dynamic range compression is a signal processing technique that may be used in hearing aids to compensate for these remaining disabilities. Its main function is to provide sufficient amplification at low input levels without overloading the auditory system at high input levels. The time constants define the time needed by the compressor to realize a change in amplification. When using relatively large time constants, the compressor only reduces differences in overall level. This type of compression is known as Automatic Gain Control (AGC) or Automatic Volume Control (AVC).With short time constants the compressor also reduces the dynamic range of a fast-fluctuating signal like speech. This last type of system is therefore often called a syllabic or a phoneme compressor. The main goal of using phoneme compression is to optimize speech intelligibility by improving the detection of weak speech cues. We developed a phoneme compression system to improve the perception of high-frequency speech cues in hearing impaired listeners. The basic mechanism is a continuously changing balancing between low- and high-frequency amplification, steered by the input level of each speech part. As a consequence, the system should provide a relatively high amount of amplification to weak high-frequency speech cues. A specific configuration was developed to additionally reduce the negative effect of low-frequency amplification on the detection of high-frequency cues. This type of configuration is called anti-USOM processing as it is meant to compensate for “Upward-Spread-Of-Masking” (USOM) of high-frequency information by low-frequency signal parts. The main goal of the present thesis was to evaluate the effect of the different compression configurations on speech intelligibility in a group of hearing-impaired listeners with moderate-to-severe perceptive high-frequency losses (chapters 4 to 7). Additionally, we have investigated the effect of various types of compression on amplitude-modulated signals and speech(-like) signals (chapters 2 and 3). The acoustical measurements in chapters 2 and 3 provided a good insight in the effect of compression on modulating signals like speech. Speech can be considered as a stream of sounds with a continuously varying spectrum. These spectral differences lead to fluctuations of the envelope of the signal within individual frequency bands. The modulation depth is a measure for the amount of fluctuations. Phoneme compression will normally reduce the amount of fluctuations, resulting in a smaller modulation depth. By comparing the modulation depth in a signal before and after compression, the effective amount of compression can be obtained. This method was applied using an amplitude-modulated signal (chapter 2) and using speech(-like) signals (chapter 3). Another method compared the average level distributions of speech with and without compression (chapter 2). The results show that relatively short time constants were needed to affect the range of modulations that are relevant to speech intelligibility. Furthermore, an effective reduction of intensity differences within separate frequency channels was only possible if the compression was applied within independent frequency channels as well. Interestingly, the results were not only influenced by the compressor settings but also by the acoustical properties of the test signal. Intensity differences were reduced more effectively for speech in a stationary background noise compared to speech only. Chapters 4 and 6 describe the effects of different types of phoneme compression on speech intelligibility in hearing-impaired listeners. Phoneme scores were obtained in conditions with and without background noise. We evaluated the difference in performance between phoneme compression and a linear reference condition near comfortable presentation levels. This implies that the results could not be influenced by differences in overall level between the various conditions. The results described in chapter 4 show that hearing-impaired listeners may benefit from our type of phoneme compression in conditions without background noise. Consonant perception was improved by phoneme compression whereas the anti-USOM processing had an additional positive effect on vowel perception. Unfortunately, no such positive effects were found in conditions with background noise. Even substantially negative effects were found with the anti-USOM configuration that gave the best performance in quiet. The use of a more moderate type of anti-USOM in chapter 6 also resulted in a negative effect on phoneme recognition in background noise. No benefit was found for other types of phoneme compression in background noise. The use of a compression ratio of 4 resulted even in negative effects (chapter 4). This means that the performance in background noise gets poorer with an increasing amount of phoneme compression. The temporal behaviour of the background noise did not influence the results (chapter 6). We hoped to find positive effects from phoneme compression in a fluctuating background noise, but no such improvement was found. The results of chapter 5 can be used to understand the measured effects of compression on speech intelligibility. Two methods were used to analyse the perceptual confusions of chapter 4. INDSCAL was used to identify and visualise the most relevant differences in phoneme perception. However the interpretation of these differences was not always easy because the perceptual dimensions could be related to several perceptual features. Therefore, SINFA was used as a second method. The advantage of using this method was that the various effects could be separated for the different predefined articulatory features. In quiet, positive effects were found on the perception of features containing mainly high-frequency information. This is according to our original goal to improve the identification of high-frequency cues by phoneme compression. However, the perception of high-frequency cues appeared to be highly deteriorated at critical background noise conditions. As a consequence, the features containing low-frequency information had become of major importance. The use of anti-USOM processing removed low-frequency information that appeared to be relevant for the perception of low-frequency cues. Additionally, we evaluated three phoneme compression conditions in a small field study using an experimental body-worn hearing aid (chapter 7). The phoneme compression configurations were embedded in a slow-acting non-linear system to compensate for differences in overall level. The listeners used the system for a period of six weeks next to the own hearing aids. The performance with the various compression programs was measured every week. The main question was if the performance could be influenced by a frequent use of the system. In general, the results were similar to that in previous experiments. The performance in background noise tended to be poorer than the performance in quiet. Interestingly, the overall recognition score with phoneme compression improved over time. However, a large part of this improvement was also found for the reference condition. The tendency for a small additional improvement with phoneme compression may be attributed to acclimatization to the speech processing. The experiences of the hearing-impaired listeners with the phoneme compression programs differed between listeners and depended of the difference in performance with the own hearing aids. Generally they had no problems with the sound of the new programs.