I have a question related to "audio mining", the process of analyzing audio clips for the purpose of searching and other processing. I see references on the web (eg, this and this) that suggest that the technology exists and is in use, but only on large (and expensive) scales.
Most of these systems allow you search the audio database with text queries, which requires the database to have undergone speech-to-text analysis. Speech-to-text, particularly for arbitrary (and unfamiliar) speakers is incredibly difficult, and so it's not surprising that these systems are not mainstream or cheap yet.
But I'm interested in a subset of this problem: single-speaker, speech-only search. That is, I record a bunch of audio clips of me speaking, and then I submit a query--in the form of spoken words. So it's only one speaker, and the system never has to interpret the audio as text.
I know this simpler problem is not trivial, but it seems like it should be much simpler than the commercial systems I've read about. And I can't see any examples of this simpler system.
Anybody?
Just a pointer to a class of technologies that you've already probably examined, but the various "recognize songs from short samples" technologies are being deployed (Gracenote's MusicID(sm) is the one I know of), it might be worth seeing if that gives you any search terms which are useful.
ReplyDeleteYeah, Muscle Fish has a sound-effect-clip-searchy-thingy, SoundFisher, which is similar. They say
ReplyDeleteIt is also not adapted to speech recognition or speaker identification tasks, although occasionally and under very specific (usually impractical) situations, SoundFisher may provide some utility for music and speech-related sound management.
I conclude from that, and from other stuff I've read, that audio mining tools are depressingly domain-specific, and tend not to cross over well.
I have no idea what you're talking about - but need to get your address again.
ReplyDeleteHope you had a safe trip back and all is well in your world!
Peace,
Jason