PROGRAMMABLE VLSI FOR SPEECH PROCESSING
There are two problems which come for designing the cost effective systems for commercial speech/signal processing applications. The First one is, the systems used here require high performance arithmetic processors. There is a Standard solution to this problem i.e. to have involved hardware implementations which are generally very expensive, require large number of components. The second designing problem is for the analog interface. This area has relatively little prominence, yet in many cases the analog subsystem is a major contributor to both system cost and potential system performance.
The Peripheral Speech Processor was born out of a perceived need for a development vehicle that could demonstrate a wide range of laboratory speech processing techniques on a common, highly scalable hardware/software platform. This production tool could then be used as a means of prototyping products. With a single board signal processor, the resulting device architecture will accommodate a variety of speech and signal processing applications. It includes analogue filtering, A/D and D/A conversion, signal processing, and data buffering, among other functions needed to support such applications. A single signal processing integrated circuit is at the heart of the PSP. A bidirectional 16 bit data bus connects this processor to the rest of the PSP system.
The implementation of a full high-performance speaker based word recognizer using a PSP is a more impressive example. Figure 3 depicts a device that can recognize isolated words as well as related expressions. It has a 32-second vocabulary capacity, or around 40 utterances. The processor load during recognition is around 50% when using a more traditional vocabulary of 20 words with a total template period of 12 seconds. The basic recognition technique is to “scan1” incoming speech data to extract a collection of distinguishing features, compute a similarity measure or “distance” between the spoken utterance and a previously stored reference pattern, and perform dynamic programming time alignment, and use the resulting error measure to make a recognition decision.
The PSP word recognition system also has the ability to create and modify vocabulary. The development of VLSI technology has changed our ability to design realistic, low-cost signal processing systems dramatically. A broad range of algorithms can be implemented through software modification using a fixed hardware baseline. Cycles of system creation and, as a result, system costs are reduced dramatically. The era of commercial signal processing has arrived.
