Search

Extraction of a target sound source amidst multiple interfering sound sources is difficult when there are fewer sensors than sources, as is the case for human listeners in the classic cocktail-party situation. This study compares the signal extraction performance of five algorithms using recordings of speech sources made with three different two-microphone arrays in three rooms of varying reverberation time. Test signals, consisting of two to five speech sources, were constructed for each room and array. The signals were processed with each algorithm, and the signal extraction performance was quantified by calculating the signal-to-noise ratio of the output. A frequency-domain minimum-variance distortionless-response beamformer outperformed the time-domain based Frost beamformer and generalized sidelobe canceler for all tests with two or more interfering sound sources, and performed comparably or better than the time-domain algorithms for tests with one interfering sound source. The frequency-domain minimum-variance algorithm offered performance comparable to that of the Peissig–Kollmeier binaural frequency-domain algorithm, but with much less distortion of the target signal. Comparisons were also made to a simple beamformer. In addition, computer simulations illustrate that, when processing speech signals, the chosen implementation of the frequency-domain minimum-variance technique adapts more quickly and accurately than time-domain techniques.

The reverberation time (RT) is an important parameter for characterizing the quality of an auditory space. Sounds in reverberant environments are subject to coloration. This affects speech intelligibility and sound localization. Many state-of-the-art audio signal processing algorithms, for example in hearing-aids and telephony, are expected to have the ability to characterize the listening environment, and turn on an appropriate processing strategy accordingly. Thus, a method for characterization of room RT based on passively received microphone signals represents an important enabling technology. Current RT estimators, such as Schroeder’s method, depend on a controlled sound source, and thus cannot produce an online, blind RT estimate. Here, a method for estimating RT without prior knowledge of sound sources or room geometry is presented. The diffusive tail of reverberation was modeled as an exponentially damped Gaussian white noise process. The time-constant of the decay, which provided a measure of the RT, was estimated using a maximum-likelihood procedure. The estimates were obtained continuously, and an order-statistics filter was used to extract the most likely RT from the accumulated estimates. The procedure was illustrated for connected speech. Results obtained for simulated and real room data are in good agreement with the real RT values.

This paper describes algorithms for signal extraction for use as a front-end of telecommunication devices, speech recognition systems, as well as hearing aids that operate in noisy environments. The development was based on some independent, hypothesized theories of the computational mechanics of biological systems in which directional hearing is enabled mainly by binaural processing of interaural directional cues. Our system uses two microphones as input devices and a signal processing method based on the two input channels. The signal processing procedure comprises two major stages: (i) source localization, and (ii) cancellation of noise sources based on knowledge of the locations of all sound sources. The source localization, detailed in our previous paper [Liu et al., J. Acoust. Soc. Am. 108, 1888 (2000)], was based on a well-recognized biological architecture comprising a dual delay-line and a coincidence detection mechanism. This paper focuses on description of the noise cancellation stage. We designed a simple subtraction method which, when strategically employed over the dual delay-line structure in the broadband manner, can effectively cancel multiple interfering sound sources and consequently enhance the desired signal. We obtained an 8–10 dB enhancement for the desired speech in the situations of four talkers in the anechoic acoustic test (or 7–10 dB enhancement in the situations of six talkers in the computer simulation) when all the sounds were equally intense and temporally aligned.

This paper presents a two-microphone technique for localization of multiple sound sources. Its fundamental structure is adopted from a binaural signal-processing scheme employed in biological systems for the localization of sources using interaural time differences (ITD). The two input signals are transformed to the frequency domain and analyzed for coincidences along left/right-channel delay-line pairs. The coincidence information is enhanced by a nonlinear operation followed by a temporal integration. The azimuths of the sound sources are estimated by integrating the coincidence locations across the broadband of frequencies in speech signals (the "direct" method). Further improvement is achieved by using a novel "stencil" filter pattern recognition procedure. This includes coincidences due to phase delays of greater than 2pi, which are generally regarded as ambiguous information. It is demonstrated that the stencil method can greatly enhance localization of lateral sources over the direct method. Also discussed and analyzed are two limitations involved in both methods, namely missed and artifactual sound sources. Anechoic chamber tests as well as computer simulation experiments showed that the signal-processing system generally worked well in detecting the spatial azimuths of four or six simultaneously competing sound sources.

System (10) is disclosed including an acoustic sensor array (20) coupled to processing subsystem (30). System (10) processes inputs from array (20) to extract a desired acoustic signal through the suppression of interfering signals. The extraction/suppression is performed by modifying the array (20) inputs in the frequency domain with weights selected to minimize variance of the resulting output signal while maintaining unity gain of signals received in the direction of the desired acoustic signal. System (10) may be utilized in hearing , cochlear implants, speech recognition, voice input devices, surveillance devices, hands-free telephony devices, remote telepresence or teleconferencing, wireless acoustic sensor arrays, and other applications.

A hearing system (21) has a first module (30) and a second module (30). The first module (30) acoustically couples to skin of a user of the hearing system (21) and includes a sensor (32) to detect acoustic signals. The second module (30) includes a hearing stimulator arrangement to stimulate hearing of the user in response to the acoustic signals detected with the first module (30). The first module (30) and second module (30) are each operable to bidirectionally communicate through an ultrasonic communication link comprising at least a portion of the body of the user.

A new algorithm was developed to emulate the ‘‘cocktail party effect’’ in hearing aids [Liu et al., ASA Symposium (1997)]. Speech signals are received binaurally; the interferers and the desired source are localized; and the signal from the strongest interferer is cancelled by frequency and time selective beamforming. Here we evaluate the algorithm’s performance in real conditions including reverberation. Binaural recordings of speech uttered simultaneously by three different speakers around the receivers were made in three different rooms (an anechoic room, a mildly and a highly reverberant room). Noise cancellation was evaluated by the intelligibility weighted SNR (IWSNR) gain between processed and unprocessed signals [Link and Buckley, J. Acoust. Soc. Am. 91]. Algorithm performance is severely degraded as the reverberation time increases: the IWSNR gain falls from 8–9 dB in the anechoic room to 1–2 dB in the highly reverberant room. Also, knowing of the exact location of the sources does not add significant improvement. The results suggest that the room impulse response be deconvolved from the received signals to suppress the multiple reflections of sound. We suggest a new strategy combining noise cancellation with blind identification of the room impulse response. [Research supported by the Beckman Institute.]

A desired acoustic signal is extracted from a noisy environment by generating a signal representative of the desired signal with processor (30). Processor (30) receives aural signals from two sensors (22, 24) each at a different location. The two inputs to processor (30) are converted from analog to digital format and then submitted to a discrete Fourier transform process to generate discrete spectral signal representations. The spectral signals are delayed to provide a number of intermediate signals, each corresponding to a different spatial location relative to the two sensors. Locations of the noise source and the desired source, and the spectral content of the desired signal are determined fron the intermediate signal corresponding to the noise source locations. Inverse transformation of the selected intermediate signal followed by digital to analog conversion provides an output signal representative of the desired signal with output device (90). Techniques to localize multiple acoustic sources are also disclosed. Further, a technique to enhance noise reduction from multiple sources based on two-sensor reception is described.

A desired acoustic signal is extracted from a noisy environment by generating a signal representative of the desired signal with processor ( 30 ). Processor ( 30 ) receives aural signals from two sensors ( 22, 24 ) each at a different location. The two inputs to processor ( 30 ) are converted from analog to digital format and then submitted to a discrete Fourier transform process to generate discrete spectral signal representations. The spectral signals are delayed to provide a number of intermediate signals, each corresponding to a different spatial location relative to the two sensors. Locations of the noise source and the desired source, and the spectral content of the desired signal are determined from the intermediate signal corresponding to the noise source locations. Inverse transformation of the selected intermediate signal followed by digital to analog conversion provides an output signal representative of the desired signal with output device ( 90 ). Techniques to localize multiple acoustic sources are also disclosed. Further, a technique to enhance noise reduction from multiple sources based on two-sensor reception is described.

The reverberation time (RT) is an important parameter for characterizing the quality of an auditory space. Sounds in reverberant environments are subject to coloration. This affects speech intelligibility and sound localization. State‐of‐the‐art signal processing algorithms for hearing aids are expected to have the ability to evaluate the characteristics of the listening environment and turn on an appropriate processing strategy accordingly. Thus, a method for the characterization of room RT based on passively received microphone signals represents an important enabling technology. Current RT estimators, such as Schroeder’s method or regression, depend on a controlled sound source, and thus cannot produce an online, blind RT estimate. Here, we describe a method for estimating RT without prior knowledge of sound sources or room geometry. The diffusive tail of reverberation was modeled as an exponentially damped Gaussian white noise process. The time constant of the decay, which provided a measure of the RT, was estimated using a maximum‐likelihood procedure. The estimates were obtained continuously, and an order‐statistics filter was used to extract the most likely RT from the accumulated estimates. The procedure was illustrated for connected speech. Results obtained for simulated and real room data are in good agreement with the real RT values.

This study examines the robustness of the frequency‐banded minimum‐variance beamformer (FBMVB) when processing signals from in‐the‐ear microphones or signals that include head‐related transfer function (HRTF) effects. In previous work using simulated signals from a free‐field array [Lockwood et al., J. Acoust. Soc. Am. 106, 2278 (1999)], the two‐channel FBMVB proved to be highly effective in extracting the signal of a target sound from a known direction amidst multiple interfering sound sources from other unknown directions. When the signals were modified to include the effects of a measured HRTF [G. Gardner and K. D. Martin, J. Acoust. Soc. Am. 97, 3907–3908 (1995)], the performance of the FBMVB improved as measured by the SNR gain and intelligibility‐weighted SNR gain. The FBMVB also outperformed conventional algorithms in both the free‐field and HRTF tests. Additionally, analytical results show that under certain conditions the FBMVB may extract a target source and perfectly cancel multiple interfering sound sources despite differences in the frequency‐dependent intersensor gains for each source. Consequently, the FBMVB may extract signals from the on‐axis direction without knowledge of the HRTF, as long as the HRTF in the on‐axis direction is identical for both ears. [Work supported by NIH‐NIDCD Grant R21DC04840.]

The two‐sensor frequency‐banded minimum‐variance beamformer (FBMVB) [Lockwood et al., J. Acoust. Soc. Am. 106, 2278 (1999)], shown to extract a target speech signal from multiple interfering speech sources, is extended to three and four sensors. The extended FBMVB retains the real‐time computational advantages of the two‐sensor FBMVB. Results from a battery of simulated and real acoustic environment tests suggest that the extended FBMVB offers better extraction of the target speech than the two‐sensor FBMVB. Measuring the quality of extraction as signal‐to‐noise ratio gain and intelligibility‐weighted signal‐to‐noise ratio gain, the extended FBMVB demonstrates performance improvements of 2–3 dB over the already high performance of the two‐sensor FBMVB for most tests. In general, results suggest that the extended FBMVB is more robust with respect to crowded interferer settings and room reverberation.

The goal of this study was to assess and compare the capacities of several array‐processing algorithms to extract on‐axis signals in the presence of one or more off‐axis interferers. Performance comparisons are made between the localization/extraction (L/E) algorithm [Liu et al., J. Acoust. Soc. Am. 108, 1888–1905 (2000)], the frequency‐banded minimum‐variance beamformer (FBMVB) [Lockwood et al., J. Acoust. Soc. Am. 106, 2278 (1999)], the Frost beamformer, the Griffiths–Jim generalized sidelobe canceller (GSC), and the Peissig–Kollmeier (P‐K) algorithm [Kollmeier et al., J. Rehab. Res. Dev. 30, 82–94 (1993)]. The two metrics used to evaluate algorithm performance were the signal‐to‐noise ratio gain (SNRG) and the intelligibility‐weighted signal‐to‐noise ratio gain (IWSNRG). The signals used simulated an omnidirectional microphone array in free field and were composed of sentence‐length speech by male and female speakers against a background of one to four interfering sound sources. Overall, the IWSNRGs for the L/E and FBMVB algorithms were similar to those for the Frost, GSC, and P‐K algorithms when there was a single interferer, but 3–4 dB higher when there were two or more interferers. This study shows that FBMVB and L/E algorithms hold promise for hearing aids to be used in noisy environments. [Work supported by NIH‐NIDCD Grant R21DC04840.]

The effectiveness of a newly developed signal‐processing algorithm to extract speech in the presence of multiple interferers was evaluated through intelligibility ratings from normal‐ and hearing‐impaired listeners. Sentences were presented with competing speech and multitalker babble (MTB) at four different SNRs to groups of individuals with sensorineural hearing loss and normal hearing. The sentences and competition had been processed through a real‐time, frequency‐banded minimum‐variance beamforming (FBMVB) algorithm in multitalker environments (Lockwood et al., 1999). Competing messages and MTB were simulated at +22 and +45 azimuths relative to the target source. Each target sentence with competing messages was submitted to the FBMVB algorithm to yield a set of 24 ‘‘processed’’ and 24 ‘‘unprocessed’’ sentences per SNR condition. Listeners rated the intelligibility (0%–100%) of the target sentences and repeated the last word of each sentence. Results from the ratings and the word recognition scores show improvement for every listener when the ensemble speech and competition were processed through the FBMVB algorithm. Results from these experiments indicate the advantage of the algorithm in selective extraction of speech in complex environments. Continued evaluation of the algorithm under real‐room conditions is in progress. [Work supported by R21DC04840 ONR N00014‐99‐1‐0612.]

A real‐time dual‐microphone‐based signal‐processing system has been developed that suppresses off‐axis interference from multiple sound sources. The system employs optimal filtering methods to attenuate the amplitude of all signals other than that from a desired receive direction. The algorithm was implemented on a Texas Instruments TMS320C62 fixed‐point digital signal‐processor evaluation board installed in a PC. Input/output was accomplished via a 16‐bit codec controlled by optimized assembly language routines. To simplify algorithm coding, the main algorithm was written in C and executes using approximately 55 processor time, allowing for future additions to enhance performance. It has a delay of 15 ms. The system can significantly improve the intelligibility of speech in noisy environments, as evidenced in off‐line tests by a marked intelligibility weighted SNR gain [Link and Buckley, J. Acoust. Soc. Am. 91, 1662 (1992)]. Initial real‐time testing with actual binaural signals supports the conclusion that this performance gain will be maintained. The results show that this system may be used as a main building block for hearing aids to be used in noisy environments. [Research supported by the Beckman Institute.]

A biologically inspired localization and speech‐enhancement technique, using two microphones, was recently developed for extracting and characterizing a sound in the presence of multiple interfering sources [Liu et al., The Second Biennial Hearing Aid R&D Conference (1997)]. In the previous study, a ‘‘stencil’’ filter was created in the frequency versus intermicrophone phase delay space which significantly enhanced the acuity of sound localization. The stencil followed the multiple curves which represent phase coherence at multiple time differences. However, source artifacts resulted from the curved lines that pose a problem for identifying real versus artifactual sources and for obtaining accurate estimates of sound azimuths. These problems are more severe when the number of sources is large and when some of the sources are less intense than the others. The current study further investigated the role of the curved lines in localization from a weighted stencil filter using different weighting methods. Fur...