would it be possible to detect if the incoming audio is music or speech? i guess it's very hard, but i was thinking about some methods:
using some kind of frequency detection using bonk (if the tempo is stable = music) env~ (most music are compressed nowadays) training a voice (using neural network?!?)
From the author of aubio: Use a few low level features, such as energy of low and high frequencies bands, spectral spread. In a second step, these approaches are often refined using machine learning techniques bayesian networks or support vector machines.
See for instance these papers: http://cobweb.ecn.purdue.edu/~malcolm/interval/1996-085/ http://www.aclweb.org/anthology/O/O08/O08-1015.pdf http://www.hindawi.com/journals/asp/2009/628570.html
i would like to achieve > 90% of accuracy if possible. any suggestions are welcome!
First of all, I would take it from another angle:
<this is one possible way, out of zillions> if it is speech or not. Thus if the speech recognizer has X % of recogniztion rate, you inherit that percentage. Now you heavily depend on the recognizer, some recognizers like teh default windows try to always match the input to some string, thus they are a bit of garbage in academic terms, what you need is a strong open recognizer that can tell you how % similar the sentence is to a target sentence in database.
Why do I suggest this angle?
- Cause' I don't wanna think "what is music". Speech is a language, it is
defined, it easy structured. Music? Noise is music, drone is music, ambient can be non rhythmical, what about an a Capella singing? Will it be music? and all those inherited philosophical issues. Furthermore, if you need more help maybe explaining the context will aid us, because if you only care for certain "music" can be easier. ALSO: if you have access the audio data, you can always extract (filter) the music.
</this is one possible way, out of zillions>
best, pedro
On Mon, Feb 7, 2011 at 5:43 PM, patrick puredata@11h11.com wrote:
would it be possible to detect if the incoming audio is music or speech? i guess it's very hard, but i was thinking about some methods:
using some kind of frequency detection using bonk (if the tempo is stable = music) env~ (most music are compressed nowadays) training a voice (using neural network?!?)
From the author of aubio: Use a few low level features, such as energy of low and high frequencies bands, spectral spread. In a second step, these approaches are often refined using machine learning techniques bayesian networks or support vector machines.
See for instance these papers: http://cobweb.ecn.purdue.edu/~malcolm/interval/1996-085/ http://www.aclweb.org/anthology/O/O08/O08-1015.pdf http://www.hindawi.com/journals/asp/2009/628570.html
i would like to achieve > 90% of accuracy if possible. any suggestions are welcome!
Pd-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
I already have a way to pause the voice recognition after X mistakes (based on time, score, ...). BUT why start the recognition if it's not someone speaking.
So maybe it's more "easy" to detect a voice and if it doesn't meet the criteria then it's music or something else (a phone is ringing, ...)?
Yeah. Trying to define music is essentially pointless. Defining speech would be easier.
However, since you mentioned neural nets, you could try to train a net on speech and music (there's an ANN object out there that I've tested) and see what happens. That would be a fun experiment, no idea how well it would work... you'd need a pretty huge training set for it to be even remotely (I would think).
——t3db0t
On Feb 7, 2011, at 1:38 PM, Pedro Lopes wrote:
First of all, I would take it from another angle:
<this is one possible way, out of zillions> if it is speech or not. Thus if the speech recognizer has X % of recogniztion rate, you inherit that percentage. Now you heavily depend on the recognizer, some recognizers like teh default windows try to always match the input to some string, thus they are a bit of garbage in academic terms, what you need is a strong open recognizer that can tell you how % similar the sentence is to a target sentence in database.
Why do I suggest this angle?
- Cause' I don't wanna think "what is music". Speech is a language, it is defined, it easy structured. Music? Noise is music, drone is music, ambient can be non rhythmical, what about an a Capella singing? Will it be music? and all those inherited philosophical issues. Furthermore, if you need more help maybe explaining the context will aid us, because if you only care for certain "music" can be easier. ALSO: if you have access the audio data, you can always extract (filter) the music.
</this is one possible way, out of zillions>
best, pedro
On Mon, Feb 7, 2011 at 5:43 PM, patrick puredata@11h11.com wrote: would it be possible to detect if the incoming audio is music or speech? i guess it's very hard, but i was thinking about some methods:
using some kind of frequency detection using bonk (if the tempo is stable = music) env~ (most music are compressed nowadays) training a voice (using neural network?!?)
From the author of aubio: Use a few low level features, such as energy of low and high frequencies bands, spectral spread. In a second step, these approaches are often refined using machine learning techniques bayesian networks or support vector machines.
See for instance these papers: http://cobweb.ecn.purdue.edu/~malcolm/interval/1996-085/ http://www.aclweb.org/anthology/O/O08/O08-1015.pdf http://www.hindawi.com/journals/asp/2009/628570.html
i would like to achieve > 90% of accuracy if possible. any suggestions are welcome!
Pd-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
-- Pedro Lopes (MSc) contact: pedro.lopes@ist.utl.pt website: http://web.ist.utl.pt/Pedro.Lopes / http://pedrolopesresearch.wordpress.com/ | http://twitter.com/plopesresearch _______________________________________________ Pd-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
Sometimes data beats algoriths, but really depend on the scope.
I read this in SIGGRAPH a couple of years ago: (was regarding image matching and blending)
"there's no good algorithm to do this, only millions of data in the database".
:) p.s.:There's a state of the art research project in Portugal, with speech recognition and realtime translation of streaming TV, and what they use is: massive recording data from the European parliament reunions. Massive data.
On Tue, Feb 8, 2011 at 4:22 AM, Tedb0t lists@liminastudio.com wrote:
Yeah. Trying to define music is essentially pointless. Defining speech would be easier.
However, since you mentioned neural nets, you could try to train a net on speech and music (there's an ANN object out there that I've tested) and see what happens. That would be a fun experiment, no idea how well it would work... you'd need a pretty huge training set for it to be even remotely (I would think).
——t3db0t
On Feb 7, 2011, at 1:38 PM, Pedro Lopes wrote:
First of all, I would take it from another angle:
<this is one possible way, out of zillions> if it is speech or not. Thus if the speech recognizer has X % of recogniztion rate, you inherit that percentage. Now you heavily depend on the recognizer, some recognizers like teh default windows try to always match the input to some string, thus they are a bit of garbage in academic terms, what you need is a strong open recognizer that can tell you how % similar the sentence is to a target sentence in database.
Why do I suggest this angle?
- Cause' I don't wanna think "what is music". Speech is a language, it is
defined, it easy structured. Music? Noise is music, drone is music, ambient can be non rhythmical, what about an a Capella singing? Will it be music? and all those inherited philosophical issues. Furthermore, if you need more help maybe explaining the context will aid us, because if you only care for certain "music" can be easier. ALSO: if you have access the audio data, you can always extract (filter) the music.
</this is one possible way, out of zillions>
best, pedro
On Mon, Feb 7, 2011 at 5:43 PM, patrick puredata@11h11.com wrote:
would it be possible to detect if the incoming audio is music or speech? i guess it's very hard, but i was thinking about some methods:
using some kind of frequency detection using bonk (if the tempo is stable = music) env~ (most music are compressed nowadays) training a voice (using neural network?!?)
From the author of aubio: Use a few low level features, such as energy of low and high frequencies bands, spectral spread. In a second step, these approaches are often refined using machine learning techniques bayesian networks or support vector machines.
See for instance these papers: http://cobweb.ecn.purdue.edu/~malcolm/interval/1996-085/ http://www.aclweb.org/anthology/O/O08/O08-1015.pdf http://www.hindawi.com/journals/asp/2009/628570.html
i would like to achieve > 90% of accuracy if possible. any suggestions are welcome!
Pd-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
-- Pedro Lopes (MSc) contact: pedro.lopes@ist.utl.pt website: http://web.ist.utl.pt/Pedro.Lopes / http://pedrolopesresearch.wordpress.com/ | http://twitter.com/plopesresearch _______________________________________________ Pd-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
Pd-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
-
patrick -
Pedro Lopes -
Tedb0t