Adaptive network for in-band signal separation
First Claim
Patent Images
1. A signal processor for separating a composite signal into at least one of its constituent signals, said processor comprising:
- means for dividing said composite signal into discrete sampled portions;
a plurality of neurons capable of receiving signals and producing an output signal said neurons including input neurons adapted to receive said sampled portions of said composite signal;
a plurality of synaptic connection means providing a weighted interconnection between selected ones of said neurons;
means for training said processor to produce an output that approximates at least one of said constituent signals, said training means including;
(a) means for presenting a composite input training signal to selected ones of said neurons;
(b) means for presenting a desired output, consisting of at least one of said constituent signals, to selected ones of said neurons; and
(c) means for changing the strength of said synaptic connection means to cause said signal processor to produce said desired output in response to said training signal; and
filter means for generating both low frequency and high frequency representations of said composite signal to present to said input neurons, wherein said low frequency representation includes a larger portion of said composite signal than said high frequency representation; and
wherein said desired output consists of high frequency representation of at least one of said constituent signals, whereby said processor is capable of receiving both a high frequency and a low frequency portion of said composite signal during training and during processing.
2 Assignments
0 Petitions
Accused Products
Abstract
An Adaptive Network For In-Band Signal Separation (26) and method for providing in-band separation of a composite signal (32) into its constituent signals (28), (30). The input to the network (26) is a series of sampled portions of the composite signal (32). The network (26) is trained with at least one of said composite signals (28) (30) using a neural network training paradigm by presenting one or more of the constituent signals (28) (30) to said network (28). The network (26) may be used to separate multiple speech signals from a composite signal from a single sensor such as a microphone.
35 Citations
13 Claims
-
1. A signal processor for separating a composite signal into at least one of its constituent signals, said processor comprising:
-
means for dividing said composite signal into discrete sampled portions; a plurality of neurons capable of receiving signals and producing an output signal said neurons including input neurons adapted to receive said sampled portions of said composite signal; a plurality of synaptic connection means providing a weighted interconnection between selected ones of said neurons; means for training said processor to produce an output that approximates at least one of said constituent signals, said training means including; (a) means for presenting a composite input training signal to selected ones of said neurons; (b) means for presenting a desired output, consisting of at least one of said constituent signals, to selected ones of said neurons; and (c) means for changing the strength of said synaptic connection means to cause said signal processor to produce said desired output in response to said training signal; and filter means for generating both low frequency and high frequency representations of said composite signal to present to said input neurons, wherein said low frequency representation includes a larger portion of said composite signal than said high frequency representation; and wherein said desired output consists of high frequency representation of at least one of said constituent signals, whereby said processor is capable of receiving both a high frequency and a low frequency portion of said composite signal during training and during processing. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10)
-
-
11. A neural network for separating a composite signal into at least one of its constituent signals, said network comprising:
-
means for dividing said composite signal into discrete sampled portions; a plurality of neurons capable of receiving signals and producing an output signal said neurons including input neurons adapted to receive said sampled portions of said composite signal; a plurality of synaptic connection means providing a weighted interconnection between selected ones of said neurons, said weighted connections being fixed and derived from a separate processor said separate processor including; means for dividing said composite signal into discrete sampled portion; a plurality of neurons capable of receiving signals and producing an output signal, said neurons including input neurons adapted to receive said sampled portions of said composite signal; a plurality of synaptic connection means providing a weighted interconnection between selected ones of said neurons; means for training said processor to produce an output that approximates at least one of said constituent signals, said training means including; (a) means for presenting a composite input training signal to said neurons; (b) means for presenting a desired output, consisting of at least one of said constituent signals, to selected ones of said neurons; and filter means for generating both low frequency and high frequency representations of said composite signal to present to said input neurons, wherein said low frequency representation includes a larger portion of said composite signal than said high frequency representation; and wherein said desired output consists of high frequency representation of at least one of said constituent signals, whereby said network is capable of receiving both a high frequency and a low frequency portion of said composite signal during training and during processing.
-
-
12. A method for separating a composite signal into at least one of its constituent signals said method comprising:
-
dividing said composite signals into discrete sample portions; receiving said sampled portions in a plurality of neurons, said neurons including input neurons adapted to receive said sampled portions of said composite signal; and said neurons including output neurons adapted to produce output signals; providing a weighted interconnection between selected ones of said neurons; training said plurality of interconnected neurons to produce an output that approximates at least one of said constituent signals, said training including the steps of; (a) presenting a composite input training signal to said neurons; (b) presenting a desired output, consisting of at least one of said constituent signals, to selected ones of said neurons; and (c) changing the strength of said synaptic connection means to cause said output neurons to produce said desired output in response to said training signal; and generating both low frequency and high frequency representations of said composite signal to present to said input neurons, wherein said low frequency representation includes a larger portion of said composite signal than said high frequency representation; and wherein said output neurons produce an output that consists of a high frequency representation of at least one of said constituent signals, whereby said input neurons are capable of receiving both a high frequency and a low frequency portion of said composite signal during training and during processing. - View Dependent Claims (13)
-
Specification