Variants on low-latency convolution-smooth updating of...

Details Variants on low-latency convolution-smooth updating of... Variants on low-latency convolution-smooth updating of...

1999-03-30

2002-07-16

Ngo, Ohuong Dinh (Department: 2124)

Electrical computers: arithmetic processing and calculating

Electrical digital calculating computer

Particular function performed

Reexamination Certificate

active

06421697

ABSTRACT:

FIELD OF THE INVENTION
The present invention relates to low latency convolution of digital signals and, in particular, describes a highly efficient form of convolution suitable for use on a processor architecture.
BACKGROUND OF THE INVENTION
The convolution of two signals is an extremely important process in many digital signal processing applications such as audio signal processing, radar signal analysis, video signal processing etc. The convolution process is commonly carried out in the Fourier Domain as it requires reduced computational resources.
Turning now to
FIG. 1
, there is illustrated the standard Fourier Domain “overlap and save” convolution process for convolving an audio signal. The process is more fully set out in the standard text such as “Digital Signal Processing”, by John Proakis and Dimitis Manolakis, McMillan Publishing Company, 1992.
In the traditional overlap and save method as illustrated
20
in
FIG. 1
, the input signal
21
is digitized and divided into N sample blocks
22
with N normally being a power of 2. Similarly, an impulse response of length N
23
is determined normally by taking desired environmental measurements, and padded to length 2N by zeros. A first mapping
24
is applied to the 2N blocks of the impulse response
23
so as to form N complex numbers having real and imaginary coefficients. The FFT is then applied to produce N Frequency coefficients. The step
24
can be carried out once before processing begins and the corresponding frequency domain coefficients
25
stored for later use.
Next, blocks of length 2N of the input audio are taken and again a fast fourier transform is applied so as to determine corresponding frequency domain data
28
corresponding to the 2N real input values. Next, the two sets of data are element-by-element multiplied
30
so as to product frequency domain data
31
. An inverse fourier transform is then applied to produce 2N real values with the first N
34
being discarded and the second N
35
becoming the output values
36
for the output audio. The process illustrated in
FIG. 1
is well known as a standard frequency domain convolution process.
International PCT Application No. PCT/AU93/00330 entitled “Digital Filter Having High Accuracy and Efficiency”, assigned to Lake DSP Pty Ltd, discloses a convolution filter which extends the aforementioned process so as to provide for an extremely low latency and the ability to perform long convolutions.
Unfortunately, it is often the case, particularly in headtracked audio signal processing applications, that the impulse response coefficients
23
must be rapidly altered whilst an audio input signal is being convolved. Unfortunately, when such alterations are carried out, the output audio signal
36
often contains distracting audible ‘clicks’ which are highly undesirable.
SUMMARY OF THE INVENTION
It is an object of the present invention to provide for a convolution process suitable for utilization in producing cross fades.
In accordance with a first aspect of the present invention, there is provided a method of performing a convolution on an input data stream, the method comprising the steps of: (a) transforming an initial series of input data values into the frequency domain to form initial frequency domain data values; (b) multiplying the frequency domain data values with a set of coefficients so as to form convolved frequency data; (c) transforming the convolved frequency data into a spatial domain so as to form an initial series of output data values; (d) performing steps (a) to (c) on a subsequent series of input data values of the input data stream so as to form a subsequent series of output data values; the subsequent series of input data values overlapping the initial series of input data values; and (e) cross fading overlapping portions of the initial and subsequent output data values.
The initial series of input data values and the subsequent series of input data values data are preferably convolved with different sets of coefficients. The sets of coefficients, in one embodiment, represent impulse response functions.
The cross fading can comprise utilizing a fade up interpolation at an initial end of the output data values and a fade down at a final end of the output data values.


REFERENCES:
patent: 5270953 (1993-12-01), White
patent: 5606575 (1997-02-01), Williams

Affiliated with

Also associated with

Rating

Variants on low-latency convolution-smooth updating of... does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Variants on low-latency convolution-smooth updating of..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Variants on low-latency convolution-smooth updating of... will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-2884795