Dynamic digital filter using neural networks

• US 5,532,950 A
• Filed: 04/25/1994
• Issued: 07/02/1996
• Est. Priority Date: 04/25/1994
• Status: Expired due to Term

• Current Assignee: Truewave, L.L.C.
• Original Assignee: Wadia Digital Corporation
• Inventors: Hustig, Charles H., Najafi, Hossein L., Kinne, James, Moses, Donald W.
• Primary Examiner(s): Malzahn, David H.
• Application Number:
  US08/232,733
• Time in Patent Office: 799 Days
• Field of Search: 364/724.19, 364/724.2, 364/724.02, 364/724.01, 395/27
• US Class Current: 708/322
• CPC Class Codes: H03H 17/02 Frequency selective networks digital computers for complex mathematical operations G06F17/10

Abstract

The present invention provides an apparatus for decoding and classifying a digital audio input signal and for reconstructing the digital audio input signal, so that when the reconstructed signal is converted to an analog signal by a digital to analog converter ("DAC"), the analog signal can drive a preamplifier, power amplifier or speakers directly. In particular, the present invention proposes a digital filter than can be adapted to have appropriate filtering characteristics based on the signal being filtered. The invention uses a neural network to adjust coefficients of a digital filter, depending on whether the digital audio input signal is more periodic or more aperiodic. If the digital audio input signal is more periodic, the coefficients will configure the digital filter so that the filter has the characteristics of an analog brickwall filter. Whereas if the digital audio input signal is more aperiodic, the coefficients produced by the neural network will configure the digital filter to have more characteristics of an interpolation filter. The neural network is trained to recognize certain periodic and aperiodic signals and to produce digital filter parameters, preferably polynomial coefficients, correspondingly. The coefficients are selected to respond to the pure or blended periodic and aperiodic features of certain archetypal input signals.

Claims

1. An apparatus for reconstructing an audio input signal, comprising:

means for receiving a plurality of digital audio input signals provided by a digital signal source representing samples of the audio input signal;

a trained neural network, which is a result of a set of training signals with corresponding filter coefficients previously taught to the neural network, responsive to the digital audio input signals to dynamically provide digital filter coefficients wherein the coefficients are produced based on the digital audio input signals received, unaccompanied by any training signals;

a digital filter for producing a digital audio output signal for reconstruction of the audio input signal, said filter receiving the digital audio input signals and the coefficients provided by the neural network, wherein said coefficients dynamically configure the digital filter to produce a filtered digital audio output signal.

2. The apparatus for reconstructing an audio input signal as recited in claim 1, wherein said means for receiving comprises a pre-processor means for outputting processed digital audio input signals to the neural network.

3. The apparatus for reconstructing an audio input signal as recited in claim 1, wherein said neural network produces one of two sets of coefficients, whereby when a digital audio input signal is received by the neural network, the neural network classifies the input signal in one of two categories and responsive to the category outputs to the digital filter a corresponding one of the two sets of coefficients.

4. An apparatus for reconstructing an audio input signal, comprising:

means for receiving a plurality of digital audio input signals provided by a digital signal source representing samples of the audio signal;

a neural network responsive to the digital audio input signals to dynamically provide a first weighting signal which indicates the extent to which the digital audio input signals correspond to a first predetermined signal characteristic and a second weighting signal which indicates the extent to which the digital audio input signals correspond to a second predetermined signal characteristic;

a first digital filter configured for smoothing a first class of digital audio input signals, wherein said first digital filter receives the digital audio input signals and produces a first interpolated audio signal at its output;

a second digital filter designed for smoothing a second class of digital audio input signals, wherein said second class is complementary to the first class of digital audio input signals and said second digital filter receives the digital audio input signals and produces a second interpolated audio signal at its output;

a first multiplier receiving the first interpolated audio signal and the first weighting signal from the neural network, whereby the first multiplier will output a first multiplied signal which is the product of the first interpolated audio signal received at its input and the first weighting signal;

a second multiplier receiving the second interpolated audio signal and the second weighting signal from the neural network, whereby the second multiplier will output a second multiplied signal which is the product of the second interpolated audio signal received at its input and the second weighing signal;

a summer which receives the first and second multiplied signals and combines the first multiplied signal and the second multiplied signal to output a filtered composite digital signal.

5. The apparatus for reconstructing an audio input signal as recited in claim 4, wherein the apparatus further comprises a pre-processor means for outputting processed digital audio input signals to the neural network.

6. The apparatus for reconstruction of an audio input signal as recited in claim 5, wherein the first digital filter is configured for filtering periodic signals.

7. The apparatus for reconstruction of an audio input signal as recited in claim 6, wherein the second digital filter is configured for filtering aperiodic signals.

8. A method for reconstructing an audio input signal comprising the steps of:

training a neural network on a set of training signals to produce filter coefficients corresponding to predetermined characteristics of the training signals;

receiving a plurality of digital audio input signals provided by a digital signal source;

providing the plurality of digital sample audio input signals to the trained neural network and a digital filter;

providing to the digital filter, filter coefficients from the trained neural network based on the digital audio input signals received, unaccompanied by any training signals; and

filtering the digital audio input signal with the digital filter, said digital filter being dynamically configured by the filter coefficients to correspond to the digital audio input signal.

9. The method for reconstructing an audio input signal as recited in claim 8, further comprising the step of preprocessing the digital audio input signal in the processing means to output processed digital audio input signals.

10. An adaptable digital filter comprising:

means for receiving a plurality of digital input signals provided by a digital signal source;

a trained neural network, which is a result of a set of training signals with corresponding at least one predetermined digital filter parameter previously taught to the neural network, responsive to said digital input signals to provide at least one digital filter parameter based on the digital input signals received, unaccompanied by any training signals;

a digital filter which dynamically receives the digital input signals and the at least one digital filter parameter from the neural network, and the at least one digital filter parameter provided by the neural network dynamically configures the digital filter.

11. A method for dynamically configuring a digital filter comprising the steps of:

training a neural network on a set of training signals with at least one corresponding predetermined digital filter parameter;

providing digital input signals to the trained neural network;

dynamically developing at least one digital filter parameter in the trained neural network based on the digital input signals received, unaccompanied by any training signals; and

dynamically transferring said at least one digital filter parameter to the digital filter, the at least one digital filter parameter configures the digital filter in response to the digital input signals.

12. An apparatus for filtering digital signals comprising:

means for receiving a plurality of varying digital input signals in time-ordered sequence;

a digital filter communicating with the means for receiving, said filter having at least one filter parameter that can be selectively set to two or more values, each value associated with different filter characteristics;

a signal pattern recognizer communicating with the digital filter and the means for receiving, said recognizer being responsive to said plurality of input signals to selectively set the at least one filter parameter to one of said two or more values, and said recognizer being trained to selectively set the at least one filter parameter based on the input signals received, unaccompanied by any training signals, at a value associated with optimal filter characteristics for a particular time-ordered sequence of varying digital input signals.

13. The apparatus for filtering digital signals as recited in claim 12, wherein the plurality of input signals comprises a time-ordered sequence of digital samples representing analog material.

14. The apparatus for filtering digital signals as recited in claim 13, wherein:

the time-ordered sequence of signals comprises a mix of both periodic and non-periodic elements;

the at least one filter parameter can be set to a plurality of different values, corresponding to equiripple filter characteristics, monotonic filter characteristics, and blends of equiripple and monotonic filter characteristics; and

the recognizer is responsive to the mix of periodic and non-periodic elements to selectively set the digital filter parameter to a value that corresponds to optimal filter characteristics for a particular mix of both periodic and non-periodic elements.

15. The apparatus for filtering digital signals as recited in claim 13, wherein the at least one filter parameter can be set to two different values, a first value corresponding to equiripple filter characteristics and a second value corresponding to monotonic filter characteristics.

16. The apparatus for filtering digital signals as recited in claim 15, wherein the recognizer is responsive to a time-ordered sequence of input signals that is periodic to selectively set the digital filter parameter to the first value and is responsive to a time-ordered sequence of signals that is non-periodic to selectively set the digital filter parameter to the second value.

17. The apparatus of claim 14 wherein the plurality of input signals comprises a time-ordered sequence of digital samples representing analog audio material.

18. The apparatus for filtering digital signals as recited in claim 12, wherein:

the recognizer selectively sets as the at least one filter parameter, a first weighting signal and a second weighting signal, each of said signals being associated with the filter characteristics of one of an equiripple filter and a monotonic filter, respectively; and

the digital filter comprises;

two digital filters, a first having equiripple filter characteristics and producing a first filter output and a second having monotonic filter characteristics and producing a second filter output;

a first multiplier receiving said first weighting signal and said first output and outputting a first multiplied signal;

a second multiplier receiving said second weighting signal and said second output and outputting a second multiplied signal; and

a summer receiving said first and second multiplied signals and outputting the sum of both signals.

19. The apparatus of claim 12 wherein the recognizer is a neural network.

20. The apparatus of claim 12 wherein the recognizer is a spectral flatness analyzer.

21. A method for filtering sample digital signals comprising the steps of:

training a neural network on a set of training signals with at least one corresponding filter parameter for each training signal;

providing digital signals to said trained neural network and a digital filter;

providing to the digital filter, the at least one filter parameter from the trained neural network, said at least one filter parameter based on the digital signal provided to the neural network unaccompanied by any training signal; and

filtering the digital signal with the digital filter, said digital filter being dynamically configured by the at least one filter parameter to correspond to the digital input signal.