0.2 Demonstrating the speak n spell  (Page 10/4)

Introduction

During the duration of the program we had several demonstrations to various groups. The first demonstration was to our Senior Vice President in December of 1976. It’s purpose was to set the stage for the full funding of the program. The demonstration was created using the IDEA funding previously discussed.

The second demonstrations was pivotal to the program as it was to the Board of Directors early in 1978 prior to having the actual circuits out of the wafer fab. By this time we were very confident that we had a working solution and that it could be a greater success than we had expected.

The third demonstration was to a group of children. This was the first time we let the target end users play with the product.

There were two other demonstrations that I will cover in later chapters that were important but not particularly pivotal to the program.

To our senior vp

It is extremely difficult to describe a product that does not yet exist. Even if it is as basic as a toy to help children learn to spell. As I have said earlier, once we received the IDEA funding we began the task of demonstrating not only the product idea but also the synthetic speech quality.

Richard, having the speech theory well understood, began putting together the demonstration using the TI Speech Research Lab’s computer. It was a 980 minicomputer from Texas Instruments with an array processor attached to it. Even with all of this performance at hand both the analysis and synthesis had to be done in non-real time. This fact should have discouraged us as the project was to do the synthesis on an IC in real time rather than non-real time on a computer the size of a room. But it didn’t seemed to phase us at all.

Several discussions were had on various aspects of the synthetic speech algorithm we would use. We concluded that LPC would be the best choice. We also chose not to do text to speech as it sounded too artificial for our vision of the product. We chose to use an eight kilohertz sample rate as it gave a good tradeoff between data rate and quality. The number of poles chosen was 10. At the time the thought was you needed two coefficients per formant and two more for the general speech envelope. Another thought was you needed the sampling rate plus two coefficients for optimal quality. In either case we chose 10 coefficients, giving us LPC-10 for the product. For the demonstration Richard stored digitized speech segments on a hard drive (50Mbytes on a free standing machine about the size of a file cabinet) in order to eliminate any time lag in the demonstration.

The voice used for the demonstration was Richard’s boss, George Doddington. Richard processed the raw data, creating the LPC-10 coding. He also wrote a quick control algorithm. So with the mini-computer, a keyboard and a monitor, it was ready for demonstration.