Modulation-Domain Speech Filtering For Noise Reduction

University of California System: University of California, Berkeley
posted on 12/19/2011

• request more info 
•   
•  LICENSE ONLINE

There are problems with discrimination speech from background noise in applications such as cell phones, hearing aids, telecommunications, automatic speech recognition, which limit the use of these important technologies. In response to this challenge, investigators at University of California at Berkeley have developed a speech filtering algorithm that improves intelligibility of speech in noisy backgrounds.  This speech filtering algorithm is applied as a first stage of processing, before transmission, analysis, or listening.  The speech filtering algorithm can be located in its entirety on a dedicated DSP or FPGA.   These systems allow the complexity needed for the computations of the gain stage, but still function fast enough for real-time denoising. The investigators have implemented the speech filtering algorithm using a computer to process signals offline (post-hoc). The inspiration for our model comes from our research on the statistics of animal vocalizations (including human speech) and on the responses of auditory neurons to these sounds. They have shown that neurons in avian primary and secondary auditory areas are tuned to represent the statistical structure of vocalizations. These areas are analogous to cortical auditory areas in humans.  The preferred embodiment uses many relatively narrow bands in the analysis filter-as low as 50 Hz bandwidth and spacing. Another embodiment uses fewer, wider bands in the analysis filter, perhaps several thousand Hz in width.

The most significant advantage of X is its ability to exploit important, basic statistical information about all human speech in a simple fashion. Some of these properties are included implicitly in advanced ASR systems. Separating these simple, physically-driven features from more complicated verbal and syntactic rules brings this power to bear in simpler, lower-power systems like cell phones. Additional complexity is also the primary drawback of this system: it requires more computational power, and because it uses a longer history, it could introduce delays that are unacceptable for low- like hearing aids and cochlear implants.

• cell phones
• hearing aids
• telecommunications
• automatic speech recognition

• low complexity
• fast, real-time denoising
• narrow bandwidths to 50 Hz, or
• few, wide bandwidth to 10K Hz