Whither Speech Recognition? (1969) [pdf]

This quote by William James from 1899 struck me as important:

> How little we actually hear, when we listen to speech, we realize when we go to a foreign theatre - for there what troubles us is not so much that we cannot understand what the actors say as that we cannot hear their words.

As a non-native speaker of English, I started watching English movies with subtitles when I was a teenager. This had an interesting effect: after a few years of doing this, I am now used to knowing each word that is spoken in a movie exactly, on its own - after all, it is clearly printed on the screen.

I now get nervous watching movies in my native language (German) without subtitles, simply because I am not able to extract each word precisely. Somehow I trained myself to expect an exact "acoustic" understanding from movies, as opposed to a "semantic" understanding. It is incredibly how the human brain is able to extract the meaning of a spoken sentence by context, facial expressions and gesture, even if we only understand half of the sentence acoustically.

This note from 1969 implies that speech-recognition researchers were hardly more than charlatans. http://www.dragon-medical-transcription.com/history_speech_r... says that the founders of Dragon Systems (iirc the first successful speech-recognition company) started in 1970. (Though they didn't start the company until 1982.)

So in retrospect I'd guess the level of funding at the time was closer to right than this critique, even if most of the work was flimflam. (The author wrote a popular book about information theory which I liked, so this is disappointing.)

Speech recognition is great (when it works) for when you need hands-free control of a machine. The big issue with it, though, is that no-one seems to want to publish any reference as to exactly what commands you can use. It's not a natural language interface (and I'd argue that we don't yet have the technical capability for a real natural language interface) so you're left with the equivalent of a command line without any way to discover ccommands. Touchscreen gesture interfaces have the same problem, implementers are too busy trying to maintain the illusion that it's "intuitive" to actually explain the secret handshakes you're meant to use with them.

They claimed 95% correctness 50 years ago.

When am I going to wake up and be able to dictate into my phone and make corrections with my voice? We must be close.

I don’t mind the mistakes made when dictating but having to pull up the keyboard takes away from the “magic“.

We can do “sound pressure wave to phoneme stream” pretty darn well, and in way that generalizes to anything you have a phonetic mapping for. (cf DeepSpeech, etc)

Going from text stream of phonemes (“mmaaaayyynaammeezzzbahhhb”) to text stream of sequence of words (“my name is bob”)is vastly more limited.

One of them is speech recognition, the other is language modeling.

language comprehension? From experience, it’s cheaper and faster to get results by making a new human :)