Browse Prior Art Database

Small Vocabulary, Discrete Word Recognizer Using a Microprocessor and Winograd Fourier Transform

IP.com Disclosure Number: IPCOM000088174D
Original Publication Date: 1977-Apr-01
Included in the Prior Art Database: 2005-Mar-04
Document File: 1 page(s) / 13K

Publishing Venue

IBM

Related People

Dixon, NR: AUTHOR [+3]

Abstract

There are many applications, such as command control, sortation and small set data entry, in which a small vocabulary, discrete-word recognizer could be used to interface a machine. A particular form for the discrete-word recognizer is described which makes use of the Winograd concept for computing the discrete Fourier transform with a small number of multiplications [*].

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 52% of the total text.

Page 1 of 1

Small Vocabulary, Discrete Word Recognizer Using a Microprocessor and Winograd Fourier Transform

There are many applications, such as command control, sortation and small set data entry, in which a small vocabulary, discrete-word recognizer could be used to interface a machine. A particular form for the discrete-word recognizer is described which makes use of the Winograd concept for computing the discrete Fourier transform with a small number of multiplications [*].

The figure shows a block diagram of the discrete-word recognizer for telephone quality speech which uses a single 16-bit, commercially available microprocessor 1 for its engine.

The analog component consists of a microphone amplifier 2 and a low-pass filter 3 which prevents aliasing error from occurring when sampling and converting via an analog-to-digital converter 4 to a digital PCN representation. For telephone speech the filter 3 cuts off at or below 4 kHz. The sampling race is normally 8 kHz. In real time, Che speech in PCM form is accepted by the microprocessor 1. When some number of samples, N, are accumulated, a Winograd Fourier Transform Algorithm (WFTA) is initiated in processing portion 5 to calculate the N-point spectrum in real time, as data is taken. Typical values here are N=16, where the WFTa would require 10 real multiplications and 60 real additions to calculate 8 spectral outputs each 2 milliseconds.

After the WFTA is taken, still in real time, some spectral processing is performed in microprocessor portion 5. Overall amplitude and some spectral outputs made up from linear combinations of N/2 WFTA outputs are calculated and stored in a small temporary store 6 of a random-access me...