Re: NSSpeechRecognizer and Speech Recognition calibration
On Dec 26, 2008, at 4:56 AM, Christopher Corbell wrote: I'm working on an accessibility app for the visually impaired and was hoping to use NSSpeechRecognizer. I've found it extremely difficult to get NSSpeechRecognizer to behave predictably on my system. Does anyone on the list have experience with this class success with the Speech Recognition system preference panel? Any tips or tricks? I find that that calibration dialog for the Speech Recognition settings doesn't work at all for me. I'm using a pretty standard external microphone (built-in to a Logitech Webcam) with an intel Mac Mini. I can see my signal just fine and I'm speaking clearly in as accent-neutral a way as I can, and still none of the test sentences ever highlights. Is a headset mic typically required, or is there some other gotcha here? It must be your particular setup. I've been doing SR ever since it debuted (Mac OS 8.x days) and have not had trouble when words/phrases are unique enough (as yours clearly are). When I give NSSpeechRecognizer a very small and unambiguous command set, I find it badly misses the mark. For example I might have Play, Next, and Stop in my command set, and it will interpret Next as Play, but it will never interpret Play as a command - pretty unusable, I'm hoping it's just a calibration issue. Since the calibration dialog isn't working for you, it's not surprising that it's getting your phrases confused. Make sure to get your setup working in the calibration area first. One last note - is there any way to do proper dictation with this class or will it only recognize the preset command list you give it? I'm thinking for example of prompting for a file name to save to, or a term to search on - it would be nice to have true dictation, otherwise I'll resort to providing an alphabet as a command set so the user can spell it out (assuming I can get that to work). No. And, you definitely do _not_ want to add letters to your language model. English letters have too many cases where sounds are extremely similar: 'B', 'C', 'D', 'E', 'G', 'P', 'T', 'V', 'Z' for probably the largest set. When I worked on numeric input, I had to offer two modes (two different speech models driven by user-preference). For example, 'sixteen' and 'sixty' were often confused. This got better over time though, but still not 100%. For users that had trouble, they could switch to the other model in which they needed to speak individual digits instead: 'one six' and 'six zero'. Now the phrases were unique enough to remove any confusion. You really only have two options: (1) The user has a 3rd-party dictation solution or (2) your solution uses words/phrases for letter input. For example the military alphabet (alpha, bravo, charlie, etc.) which was designed to work over very low-quality audio situations. ___ Ricky A. Sharp mailto:rsh...@instantinteractive.com Instant Interactive(tm) http://www.instantinteractive.com ___ Cocoa-dev mailing list (Cocoa-dev@lists.apple.com) Please do not post admin requests or moderator comments to the list. Contact the moderators at cocoa-dev-admins(at)lists.apple.com Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/cocoa-dev/archive%40mail-archive.com This email sent to arch...@mail-archive.com
Re: NSSpeechRecognizer and Speech Recognition calibration
On Fri, Dec 26, 2008 at 5:56 AM, Christopher Corbell chriscorb...@gmail.com wrote: I'm working on an accessibility app for the visually impaired and was hoping to use NSSpeechRecognizer. I've found it extremely difficult to get NSSpeechRecognizer to behave predictably on my system. Does anyone on the list have experience with this class success with the Speech Recognition system preference panel? Any tips or tricks? I find that that calibration dialog for the Speech Recognition settings doesn't work at all for me. I'm using a pretty standard external microphone (built-in to a Logitech Webcam) with an intel Mac Mini. I can see my signal just fine and I'm speaking clearly in as accent-neutral a way as I can, and still none of the test sentences ever highlights. Is a headset mic typically required, or is there some other gotcha here? When I give NSSpeechRecognizer a very small and unambiguous command set, I find it badly misses the mark. For example I might have Play, Next, and Stop in my command set, and it will interpret Next as Play, but it will never interpret Play as a command - pretty unusable, I'm hoping it's just a calibration issue. I'm afraid I don't know a lot about this stuff, but since nobody else has answered yet, I'll give it a shot. From what you describe I'd guess that your gain is not set well. You need a good strong signal when you speak, but it absolutely must not clip. If it's too quiet then I think that the recognizer won't realize when you're speaking, or won't be able to understand what you're saying when you do. If it's clipping then it will get a lot of distortion. Try to make it so that your speech just reaches the top of the green area. I've never used speech recognition extensively but I have played around with it a fair amount, and with a decent microphone that's properly calibrated, I haven't really had much trouble with it misunderstanding me, certainly not to the degree which you describe. You might also try recording yourself and then playing it back. Listen to see how it sounds. If it's distorted or otherwise of low quality then that will hurt recognition badly. One last note - is there any way to do proper dictation with this class or will it only recognize the preset command list you give it? I'm thinking for example of prompting for a file name to save to, or a term to search on - it would be nice to have true dictation, otherwise I'll resort to providing an alphabet as a command set so the user can spell it out (assuming I can get that to work). Alas, no dictation. The great thing about Apple's speech recognition is that it's speaker-independent and requires no training. However the state of the art for such things is that dictation requires training to a particular speaker's voice before it can work. So Apple's recognizer, by virtue of its capabilities in other areas, loses out on this capability. You'll have to go with your alphabet command set or some other idea that works with a predefined set of recognizable phrases. Mike ___ Cocoa-dev mailing list (Cocoa-dev@lists.apple.com) Please do not post admin requests or moderator comments to the list. Contact the moderators at cocoa-dev-admins(at)lists.apple.com Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/cocoa-dev/archive%40mail-archive.com This email sent to arch...@mail-archive.com
