3 ms·

(disclaimer: I work on speech recognition for mobile devices) The experiment setup is a little weird here. The participants were given a set of pre-created phr

by gok 10y ago

(disclaimer: I work on speech recognition for mobile devices) The experiment setup is a little weird here. The participants were given a set of pre-created phrases to type/speech. At least for English, the phrase set [1] contains utterances like "circumstances are unacceptable" …which contains rare but in-vocabulary words. That makes it hard for keyboard input (hard for humans to spell long words, hard for touch input to predict unlikely words) but very easy for speech recognition (no other word sounds like "circumstances"). And that test set is so old that it's very likely that the speech recognizer from the experiment (or any state of the art speech recognizer) has already been trained on those sentences. The utterances being pre-selected is also unfortunate. When users are given a sentence to speak ahead of time, they tend not to hesitate or stutter. They also speak faster than when they're trying to think of something to say on the fly, which is more typical of text input on mobile devices. All that being said, it's certainly true that you can often input text very quickly with speech recognition, and it's getting better every day. :) [1] http://www.yorku.ca/mack/PhraseSets.zip http://www.yorku.ca/mack/PhraseSets.zip

kentlyons 10y ago
(similar disclaimer: I've done a bunch of text input work for mobile) That phrase set was explicitly designed for text entry (typing) experiments. While not optimal, it does allow for more direct comparison to a large body of previous studies using the same phrase set (and similar procedures). Having said that, keyboard input methods that provide suggestions/corrections at multi-character or word level features probably also benefit from longer words since the recognizer has more signal to work with. Revisiting the assumptions that went into that initial phrase set (character at a time input) in light of modern text input techniques might be a good thing.
- kbutler 10y ago
  > That phrase set was explicitly designed for text entry (typing) experiments. I'm sure that set was designed for <del>typing on full-sized, physical keyboards,</del> not touch-screen mobile devices. (Thanks for the correction!) Also, even though speech is faster than typing on a touch-screen mobile device, it's a lot easier to correct the errors that inevitably happen via typing. Sometimes it's impossible to verbally correct errors or enter unrecognized words (or names!).
  - kentlyons 10y ago
    It does predate capacitive touchscreen keyboards. However it wasn't made explicitly for full size physical keyboards. They explicitly had soft keyboards in mind. Here is the paper that introduced the phrase set [1] at it refers to this soft keyboard they made previously [2]. [1] http://www.yorku.ca/mack/chi03b.html http://www.yorku.ca/mack/chi03b.html [2] http://www.yorku.ca/mack/p25-mackenzie.pdf http://www.yorku.ca/mack/p25-mackenzie.pdf
geon 10y ago
When I use speach to set reminders, I usually have to prepare mentally really carefully, so I use the exact phrasing for datetime it recognizes without me hessitating.
- daveguy 10y ago
  This. If you don't prepare the correct phrasing you will be saying it at least twice and probably fumbling with the keyboard to get it into the right mode to accept voice again -- waiting for the acknowledgement that it is listening. Voice interfaces are currently incredibly cumbersome. There is no way to reconcile "3x faster" with actual excruciating experience that is voice input. Edit: one time when it was faster is when I had a string of 4-5 alarms that needed to be set. Of course I used the touch interface for prompting the voice input.

by gok 10y ago

kentlyons 10y ago

(similar disclaimer: I've done a bunch of text input work for mobile) That phrase set was explicitly designed for text entry (typing) experiments. While not optimal, it does allow for more direct comparison to a large body of previous studies using the same phrase set (and similar procedures). Having said that, keyboard input methods that provide suggestions/corrections at multi-character or word level features probably also benefit from longer words since the recognizer has more signal to work with. Revisiting the assumptions that went into that initial phrase set (character at a time input) in light of modern text input techniques might be a good thing.

kbutler 10y ago
> That phrase set was explicitly designed for text entry (typing) experiments. I'm sure that set was designed for <del>typing on full-sized, physical keyboards,</del> not touch-screen mobile devices. (Thanks for the correction!) Also, even though speech is faster than typing on a touch-screen mobile device, it's a lot easier to correct the errors that inevitably happen via typing. Sometimes it's impossible to verbally correct errors or enter unrecognized words (or names!).
- kentlyons 10y ago
  It does predate capacitive touchscreen keyboards. However it wasn't made explicitly for full size physical keyboards. They explicitly had soft keyboards in mind. Here is the paper that introduced the phrase set [1] at it refers to this soft keyboard they made previously [2]. [1] http://www.yorku.ca/mack/chi03b.html http://www.yorku.ca/mack/chi03b.html [2] http://www.yorku.ca/mack/p25-mackenzie.pdf http://www.yorku.ca/mack/p25-mackenzie.pdf

geon 10y ago

When I use speach to set reminders, I usually have to prepare mentally really carefully, so I use the exact phrasing for datetime it recognizes without me hessitating.

daveguy 10y ago
This. If you don't prepare the correct phrasing you will be saying it at least twice and probably fumbling with the keyboard to get it into the right mode to accept voice again -- waiting for the acknowledgement that it is listening. Voice interfaces are currently incredibly cumbersome. There is no way to reconcile "3x faster" with actual excruciating experience that is voice input. Edit: one time when it was faster is when I had a string of 4-5 alarms that needed to be set. Of course I used the touch interface for prompting the voice input.