3 ms·

Since Windows Vista, at least. I remember being quite surprised by how well it worked - I could dictate and there were very few mistakes. I was particularly ta

Since Windows Vista, at least. I remember being quite surprised by how well it worked - I could dictate and there were very few mistakes. I was particularly taken with how it understood commands. It was amazing to me that it could understand "open Firefox" and "open mIRC", since I figured it must have had to guess how arbitrary program names were pronounced. But yeah, here we are with pocket-sized devices supposedly more powerful than the computers of the time, which have to send off everything to The Cloud for approximately the same quality recognition. Progress!

TeMPOraL 6y ago
You're right. I wrote my first program using MS Speech API around 2006/2007. It was even better than that! One thing it had, that doesn't seem easily available to current cloud systems, was an easy API for restricted grammar. You'd write a piece of XML that described what kind of things could follow which other kind of things, to build e.g. a graph of commands. The other big thing it had was training. You could repeatedly read text to the system, and it'll improve the recognition of your voice. If at any point the system had a problem understanding you, you could just pop into Control Panel, type in what you tried to say, and do a training run on it, and from then on, it would understand you just fine. Using this, I built a simple music control program that offered playback control, volume control and playlist selection. I defined the words to use and described the grammar in XML, and had my program control a WinAMP instance (good ol' days when people cared about interoperability - all I had to do was to send WM_USER messages to the WinAMP window...). I soldered a 3m cable to a $0.1 microphone I bought in electronics supply store, glued it to the wardrobe, and had a Star Trek like music control. I printed out all the common combinations of spoken commands, and trained the Speech API backend with it for an hour, just reading the commands and changing my location, and the type and loudness of music that played in the background during training. I.e. same text, done {near microphone, from bed, from doorway} x {quiet, medium, loud} x {pop, classical, radio}, IIRC. After doing this, the system worked flawlessly - I could listen to very loud music on the loudspeakers, and it had no problem understanding me from any point in the room. This was much better experience that I can achieve today with my overpriced smartphone and Google's advanced cloud AI. I really miss offline-first, low-latency, end-user-trainable, transparent, bullshit- and surveillance-free speech recognition.

TeMPOraL 6y ago

You're right. I wrote my first program using MS Speech API around 2006/2007. It was even better than that! One thing it had, that doesn't seem easily available to current cloud systems, was an easy API for restricted grammar. You'd write a piece of XML that described what kind of things could follow which other kind of things, to build e.g. a graph of commands. The other big thing it had was training. You could repeatedly read text to the system, and it'll improve the recognition of your voice. If at any point the system had a problem understanding you, you could just pop into Control Panel, type in what you tried to say, and do a training run on it, and from then on, it would understand you just fine. Using this, I built a simple music control program that offered playback control, volume control and playlist selection. I defined the words to use and described the grammar in XML, and had my program control a WinAMP instance (good ol' days when people cared about interoperability - all I had to do was to send WM_USER messages to the WinAMP window...). I soldered a 3m cable to a $0.1 microphone I bought in electronics supply store, glued it to the wardrobe, and had a Star Trek like music control. I printed out all the common combinations of spoken commands, and trained the Speech API backend with it for an hour, just reading the commands and changing my location, and the type and loudness of music that played in the background during training. I.e. same text, done {near microphone, from bed, from doorway} x {quiet, medium, loud} x {pop, classical, radio}, IIRC. After doing this, the system worked flawlessly - I could listen to very loud music on the loudspeakers, and it had no problem understanding me from any point in the room. This was much better experience that I can achieve today with my overpriced smartphone and Google's advanced cloud AI. I really miss offline-first, low-latency, end-user-trainable, transparent, bullshit- and surveillance-free speech recognition.