17

This clip from BBC Archive allegedly shows a computer that can do voice recognition, speaker recognition and speech synthesis with minimal delay in 1984.

I find that hard to believe, they must have someone hiding in the background and playing prerecorded clips, don't they?

6
  • 2
    One of the early applications for voice activated devices was one worn on the arm by a fork lift operator. Allowing a forklift operator to use both hands for operating the fork lift was enormously valuable. Other appplications followed, as the technology became cheaper. Feb 2 at 12:40
  • As for speech synthesis in 1984. The original 128 kByte Macintosh supported MacInTalk in 1984 (en.wikipedia.org/wiki/PlainTalk#Original_MacinTalk). It had a distinct computer-ish voice, but it was quite good at speaking (and you could change inflection, etc. with superfluous punctuation
    – Flydog57
    Feb 3 at 21:07
  • In 1961 Max Mathews and John Kelly made a computer synthesized version of 'Daisy Bell'. Sung voice + accompaniment. This later got a nod in 2001 (the movie, not the year) when it is sung by HAL as he is dying. Also in Futurama when Bender dates the Planet Express Ship.
    – dmedine
    Feb 4 at 2:00
  • this ~800 Byte TTS is from 1985 for 48K ZX ...
    – Spektre
    Feb 5 at 14:44
  • @Spektre And is just horrible. I had a speech synthesizer on my Atari ST as well and it was so bad it would not be suited for the purpose of the show.
    – AndreKR
    Feb 5 at 15:10

4 Answers 4

15

Yes, there was a real Votan system; at the time of the broadcast (1984), it was apparently a pretty new product. On the Internet Archive there is an episode of The Computer Chronicles (link plucked from the Twitter thread) in which the system is demonstrated in a little more depth (from 18:22), after which there is a mildly technical interview with the CEO of the device’s manufacturing company.

Given that this was most probably done as an advertisement, the CEO does waffle a bit, but there is still some valuable technical information to be gleaned from it: for speech production, apparently the system keeps a number of pre-recorded clips, stored in compressed form (we may guess, some form of ADPCM?), which it simply plays back on demand, one after another. As for the recognition part, this seems to be done as speaker-dependent recognition, which is why the system recognises not only what is being said, but also who is speaking; on the other hand, this requires that the system be calibrated to each speaker’s voice separately. This process of calibration is probably what the BBC host referred to as being ‘introduced to the computer’.

As it happens, The 8-Bit Guy reviewed of a speech processing hardware-software package from 1987, with a segment demonstrating menu navigation with single-word spoken commands. The capabilities of the Votan system seem roughly consistent with that level of technology.

The way Votan was presented does superficially make it seem more advanced than it actually was. Still, it seems no less honest a demonstration than the Truthbot 2000.

20

There were definitely issues of scale of hardware available, but in the fall of 1977, I think it was, I got my Commodore Pet 32kB., the B version had the full-sized ("business") keyboard. I don't recall if it was sometime in '78 or '79 that I found in Byte magazine an ad from Cognivox to add voice recognition to my Pet! Amazing! And, it was cheap enough, I bought it!

You had to train it, and it would only recognize ONE person's voice. And, the vocabulary was limited due to resource limitations .... Put this on a Vax 11/780 and WOW, now you have a large vocabulary. Only a little more work and it could recognize anyone's voice so long as their dialect wasn't too bad. And, indeed, I joined Digital Equipment Corp in 1983 and stayed with them until 1989, and sometime around 1986, I'd guess, Digital added this to their software lineup. And, some of my customers bought it!

Its primary use was on phone systems. I was one of the guys who supported Bellcore (my friend and colleague Angel V. was the primary on the account). Bellcore was the research arm of AT&T before the breakup but was also the group that handled the "800" business. And we had several arms of Bellcore's business that we supported. And one of the things they used was this voice recognition... I'm 100% sure that continued until I left in '89. Unfortunately, at the moment, I don't recall DEC's name for the product.

...After posting "DECtalk" popped into my head, but that's a definite maybe as I know they had several packages in this category and one of them was ONLY the system doing speech synthesis...

7
  • 7
    DECTalk was a speech synthesis device, considered by many the gold standard in voice quality at the time. It supported several voices with software-adjustable pitch and speed and was eventually available for license in software form as well. However I don’t believe that DEC used the DECTalk trademark for any speech recognition product. For reference, according to Wikipedia Stephen Hawking used a DECTalk. Feb 2 at 6:15
  • 1
    DECtalk provided the voice for Stephen Hawking. Feb 2 at 12:41
  • So they probably trained every single sentence for both the man and the boy, had a button on the microphone to signal end of speech, and had pre-recorded output including the individual names. Yeah, ok, that makes it more believable.
    – AndreKR
    Feb 2 at 20:02
  • And indeed the difficulty with speech recognition is not doing it, but doing it robustly. A system that recognizes a limited vocabulary that does not deviate too much from a particular dialect and uses a slow, careful speaking style? Sure. Recognizing what people say from noisy call center audio, when they are joining their words together naturally and speaking in a variety of dialects? State of the art speech recognition can still be wrong 10-20% of the time or more, even in 2021, when a human would detect 98% of the words or more.
    – Obie 2.0
    Feb 3 at 21:55
  • 1
    @AndreKR ...I'm not at all sure what you're saying.
    – Richard T
    Feb 4 at 5:14
15

By the mid 1980s speech recognition was already a non-issue. Experimental system were able to handle vocabularies of 20,000 Words or more. For situations with confined vocabulary like shown in the clip every average computer with acceptable performance could do it.

Looking close reveals that this demonstration needs only a hand full of words to be recognized: Fan and Light, On and Off, Helloand Nothing. This and the recognition of each speaker, due being trained to each one's way of speaking.

A parser would work much like known from classic text adventures. The take torch/light torch type of games. They did allow rather sophisticated input as well. One could enter Please turn the light on. Game wise this was not different from typing light on - which I guess most of us did - as all the filling words are simply ignored. Same here.

And like with an Adventure game, someone demonstrating the 'astonishing' and 'intelligent' reaction of the game by use use of well worded sentences, making the conversation look more natural and sophisticated. Sales-Hokus-Pokus :)

The math behind speech recognition isn't hard at all, even an 8 bit computer can do it. The real task was to find the 'right' word to the solution computed, which means a large (*1) amount of RAM and Disk space needs to be at hand.

Such simple systems are in wide use since the mid 1980s - everyone of us has already hit those voice systems asking to say yes or no to a series of question or speak a number to select functions.

What differs such simple system from more modern is to guess meaning and use this in turn to rework what has been 'heard'. It takes a huge amount of data to move past detecting a few words within a confined setting. Really 'understand' what's requires and turn that into actions. There is a reason Google, with their large treasure of language sniplets, carries a leading role in today's kind of speech recognition.

Then again, the answers given in that clip are prerecorded phrases. Not because generation was complicated - Remember Speak&Spell? - but because it sounded awful due the high complexity how human speech combines words and carries over sounds. Speech output had a hype phase around that time. There were not only hardware synthesizers for next to all home computers - but as well software based solutions.

The use of pre-recorded phrases where standard in speech systems way into the 2000s. The building blocks used are clearly noticable by pauses between sentences and before the names. Different reaction to different speakers is as well easy, as the system was obvious (as mentioned by the host) trained specific to the host and the boy.

So yes, action shown was state of the art in 1984.

An April 1987 IEEE Spectrum article lists 17 commercial availabe/used systems, ranging from voice dialling to dictation, recognizing vocabularies between 13 and 20,000 words.


*1 - Large at the time. So make that a megabyte or two of RAM and 20 MiB or so in Disk space :))

7
  • 3
    The use of pre-recorded phrases is standard in speech systems way into the 2020s. I hear station announcements done this way every day (the others are from a real live human). One particularly obvious cue is "to", which always ends up a homophone to "two" even when a native speaker would say /tə/; the rhythm around "to" is also more context-dependent in natural speech than splicing pre-recorded phrases together. There's also a slight unnatural emphasis on numbers (e.g. a change from platform 2A to 2B with emphasis on the numeral not the change)
    – Chris H
    Feb 2 at 12:37
  • I heard train station announcements on Monday that were totally messed up. Robotic sounding voice - but I know some people like that! What gave it away was the mixed up messages about one train vs. another, repeated. No way a real person would have done that - but someone punching buttons into an automated announcement system would explain it. Feb 2 at 14:09
  • 1
    @manassehkatz-Moving2Codidact I've been hearing a spate of bizarre errors recently: repetition of many instances of "one" for example. I'm pretty sure the messages here (UK) are auto-generated on the fly from information that's also presented as text given how consistently they match and comments from staff - station staff don't even have easy access to suppress automated messages except by making another announcement over the top, which mutes the automated one.
    – Chris H
    Feb 2 at 16:40
  • 1
    No, the sentences had not to be trained that hard. It's more about speed an pitch for each speaker. Also, it's just a few words. in this example only Fan, Light, on, off and Hello. Everything else is just noise that will be ignored. Think of the parser like one used for text adventures, where one could enter 'Please make the light on' the same way as 'light on' as all the filling words are ignored. Same here. Of course when demonstrating the game one would use whole sentences as it makes the conversation look more natural - much like here.
    – Raffzahn
    Feb 2 at 20:16
  • 3
    I still get cheap amusement out of the way map applications (Google, Apple) mispronounce the names of local streets. Actually, in my house, we've developed a bad habit of referring to the streets using the mispronounced names, so I guess the speech recognition is doing the training and we're the subjects. Feb 2 at 21:22
3

The first experiments in speech recognition go all the way back to the 1950s. https://en.wikipedia.org/wiki/Speech_recognition#Pre-1970 But, it wasn't commercialized until the late 1980s, and it didn't get to be really practical until two to three decades after that. (For some definition of "really practical".)

1
  • 8
    Sorry, this is false, "[voice / speech recognition] wasn't commercialized until the late 1980s". Cognivox did it in the late '70s.
    – Richard T
    Feb 2 at 2:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.