0.8 Taking the speak n spell into production  (Page 74/5)

Introduction

We were building our first pre-production devices to use for qualification and early samples to important people. It was during the summer after the June CES. We had decided that the production line would be set up in Midland, Texas at the airport in the old Windecker plant at the end of the runway. After we had completed the pre-production build we put them all on burn-in and life test to begin their qualification process. It was at this point that the night guard at the plant discovered a bug. With the long tested Texas concept of "one riot, one ranger", he was alone in the building as production lines were only running during the day. Well into the night, the guard began to hear voices. His only conclusion was that someone had gotten into the building. Now, I don't know whether he had his revolver out ready to fire, but, at that time guards at Texas Instruments were authorized to carry a weapon. So I'll not embellish the story with that aspect. As he made his way through the production lines he could still hear the voices which seemed to be coming from the back of the facility (where we had our burn in racks). He finally made it back to where the voices were and discovered that the burn in racks were talking to each other. Actually, the Speak N Spell design had a bug of which we never fixed. While powered up during burn in, we had a way of doing this through the module port, the bug started one Speak N Spell talking. For some strange reason, others would reply followed by more until it sounded like a crowd of people talking. I guess it is important to explain that he was not told about a new talking product being manufactured at the plant. I never saw the report he wrote up and never saw any bullet holes in the burn in racks, but suspect he had a once in a lifetime story to tell about the night he heard voices.

Surprisingly, taking the Speak N Spell into production was as difficult as creating it. We had never introduced a product to a production line which spoke. Normal production line expectations immediately became significant issues. For example, production lines are noisy; production line workers aren’t tested for hearing; quality standards for speech, particularly synthetic speech were not known.

We had already experienced quality issues when we created the vocabularies. It started with picking a good professional speaker to speak the words. Our synthesis method was Linear Predictive Coding. We chose LPC-10 at an 8kHz sample rate. That meant the LPC filter had ten coefficients. We used all ten for a voiced sound and only four for an unvoiced sound. We chose 8kHz as the right sample rate as it was what the phone system used. The issue is that many of the unvoiced sounds are actually above 4kHz (the bandwidth of the signal is, by definition, less than half of the sample rate according to the Nyquist theorem). To compound the quality issue, I had chosen not to put an anti-aliasing filter on the output of the synthesis chip. Rather than the filter, I chose to carefully pick a cheap output speaker that had a two pole roll off at 3.3KHz. The decision saved us money at the sacrifice of speech quality.