Speech Transcript - MPTRACKER: A NEW MULTI-PITCH DETECTION AND SEPARATION ALGORITHM FOR MIXED SPEECH SIGNALS

0:00:13	and run uh uh and uh you know functional
0:00:16	um uh for that
0:00:17	for thought of that
0:00:19	but are
0:00:19	uh
0:00:21	you
0:00:22	tech
0:00:23	the right
0:00:24	and going
0:00:24	for a a
0:00:27	know
0:00:27	oh
0:00:28	and like or use a course there are
0:00:31	yeah i and we talk that the real problem uh
0:00:34	i wouldn't never seen although all
0:00:36	yeah
0:00:37	so would change from code work
0:00:40	but not what is a type or in the uh D
0:00:43	if not from that but
0:00:45	and uh we really want you know are still uh try
0:00:50	oh a pitch detection the is uh
0:00:53	and essential for in
0:00:55	compute a know
0:00:56	all to racine
0:00:57	and not
0:00:58	uh
0:00:59	for to a meeting
0:01:01	but it all the tears system
0:01:04	mean a a out one source model don
0:01:07	you know that machines and can to use
0:01:10	really D
0:01:11	a a a a really can
0:01:13	do it when they
0:01:14	so uh
0:01:16	i'm the
0:01:17	oh
0:01:18	it what do you by the yet very easy uh an a as the cocktail party problem
0:01:23	and they are on uh
0:01:25	and it's very common from a you
0:01:27	uh
0:01:28	a a at work
0:01:30	vol
0:01:31	and to the model
0:01:32	a chat or uh
0:01:33	and
0:01:34	or or be
0:01:35	computational of the dreams
0:01:37	and how
0:01:38	so the real to replicate the
0:01:41	yeah
0:01:41	oh
0:01:42	so once or
0:01:44	one day
0:01:45	she
0:01:46	so that in the uh
0:01:48	a a a a a a a model
0:01:50	face
0:01:51	we have several stages a frequency the analysis
0:01:54	and and are looking for discriminative features
0:01:57	a model speakers namely each frequency once that all of that
0:02:02	spatial diversity
0:02:03	all that but with and then
0:02:05	some some of the
0:02:06	and grouping
0:02:07	at the end
0:02:08	have some um
0:02:10	a a well we're
0:02:11	a a one one or
0:02:13	oh i really would like to apply a one mask when we by this one by D
0:02:19	spectrogram of the mixture sure we can call over the underlying source
0:02:23	oh are we are interested in a single channel speech separation
0:02:27	and you have two sources and one is speaker
0:02:30	and we don't have any special favours in nation for T V
0:02:36	oh
0:02:37	so
0:02:38	and and is based on our previous work
0:02:41	we we ah i have a i don't know do you mention to the right model
0:02:45	basically uh
0:02:46	also uh
0:02:48	we propose
0:02:49	well well track information
0:02:51	another other their discriminative feature you small
0:02:54	sort all but there and semantic information you move them more right and we use them when we
0:03:00	separate um
0:03:01	so these are some prior knowledge can be trained
0:03:04	uh
0:03:06	i i one
0:03:07	so that
0:03:09	using um
0:03:10	um
0:03:11	but and that would be a record two sources
0:03:13	so
0:03:14	i
0:03:14	method that very well but
0:03:17	we did a
0:03:18	a good was that in order to have some estimate from on the line
0:03:22	each one
0:03:23	so
0:03:26	mean streets and we work and this uh well
0:03:29	uh
0:03:29	and then lot
0:03:31	a P to which has several
0:03:33	feature
0:03:34	a
0:03:35	five
0:03:36	i
0:03:37	yeah
0:03:37	pitch contours and also a them to individual source
0:03:41	so many our proposed works just only
0:03:45	it that the contour and
0:03:46	do want to
0:03:47	on the a
0:03:49	contour
0:03:49	two
0:03:50	individual
0:03:52	and uh
0:03:53	that is it is assumed that a one one of the on the line sort of is always
0:03:58	which make it prediction very
0:04:00	and also but that use a different from a a lot of papers
0:04:05	essentially for a
0:04:07	a music signal in which T
0:04:09	time
0:04:10	frequency continued to use more pronounced
0:04:12	such that the could and is very easy
0:04:16	or speech
0:04:17	and also this these uh uh uh what is it should be
0:04:20	we from the
0:04:21	a rebel
0:04:23	pitch action for single
0:04:25	speaker
0:04:26	but
0:04:26	we have another source
0:04:29	uh
0:04:30	and and have some sort of the
0:04:32	uh
0:04:33	but in each or and we need we do need to recover both
0:04:37	so
0:04:38	but you a high level
0:04:40	well
0:04:41	the i don't from and and D tracker
0:04:43	uh
0:04:45	a a for uh stages
0:04:47	section grouping separate
0:04:50	interpolation
0:04:51	but it can be a little what
0:04:53	uh oh
0:04:54	but was by we need only to resist
0:04:56	also
0:04:56	some some sort of a
0:04:58	ah
0:04:59	to making feature
0:05:01	P
0:05:01	and
0:05:02	so a separate
0:05:03	i i and
0:05:04	in interpolating for a ah
0:05:06	we uh one
0:05:10	oh uh i think there's
0:05:11	stage a a uh a a more pitch detection
0:05:15	oh i mean why by the work of a client
0:05:19	oh
0:05:19	he's group but that propose a distortion measure for in
0:05:25	so uh basically
0:05:27	uh a as a why uh
0:05:29	there is at all
0:05:30	it of course so you in but that thing think the white one
0:05:36	section
0:05:36	we we our goal is to me white these source and B the new signal is the text of the
0:05:43	the of the signal
0:05:44	and D or its deviation of the spectral densities for on the line sources
0:05:49	and they have a i
0:05:50	but are aggressive or model
0:05:53	two source are and then
0:05:55	we need mean one
0:05:56	that that in order to record the each
0:05:59	a line
0:06:00	a time
0:06:01	yeah why the same concept
0:06:03	i
0:06:04	instead of uh you a C I
0:06:07	you with a sinusoidal model and which are more suitable for the past well
0:06:11	that's
0:06:12	so we yeah
0:06:13	she thought what
0:06:14	you know the new signal
0:06:16	yeah
0:06:17	uh so one
0:06:18	yeah it's of two
0:06:20	a a you know the one source a
0:06:23	and our goal
0:06:24	for
0:06:25	detection they
0:06:26	to minimize these
0:06:28	a distortion
0:06:30	so uh
0:06:32	for this that for the
0:06:34	the uh uh uh a classic paper by mac will like they show that
0:06:37	we can
0:06:39	a group symmetry uh "'cause" that this that the in terms of sinusoidal modeling using
0:06:45	some of some
0:06:46	a a sound a does signals
0:06:49	or or a a a a thing to of peaks
0:06:51	the spectrum
0:06:52	and that we we present
0:06:54	i a a a you all uh be though
0:06:57	and that
0:06:59	that the you of L O I E the uh
0:07:02	location of P
0:07:04	for the presentation or the sinusoidal model
0:07:07	but the the peaks
0:07:09	don't occur exactly i at in with bit integer all
0:07:13	uh
0:07:14	fundamental frequency yeah another to out where here
0:07:17	and B to a parameter in order to to a exactly match
0:07:23	so uh
0:07:24	you we have to and i don't for and to uh the location of the
0:07:29	and so
0:07:31	because we do not have access to the location along the line source P
0:07:35	so we apply these
0:07:36	approximation which we found the what pretty well in right
0:07:40	so
0:07:41	and to uh
0:07:44	or are bits that separate
0:07:46	say
0:07:47	i
0:07:47	a and then you paris
0:07:49	and then be assign peak
0:07:51	each data source
0:07:53	and then they are very close the
0:07:55	sign no ha of the peak to each individual sources
0:07:58	and then and
0:07:59	oh
0:07:59	the
0:08:01	only problem to the the to me might you the two pitch a
0:08:05	a uh
0:08:06	uh
0:08:07	points
0:08:07	so we we we my station
0:08:09	and we got some um
0:08:11	estimation for the
0:08:12	i but one source for each
0:08:16	a are you uh
0:08:17	yeah idea of how
0:08:18	to uh because that
0:08:20	a whole one a for more speak to a the signal
0:08:24	ah
0:08:24	we have
0:08:25	a source here
0:08:26	a first one had a week one up to twenty eight
0:08:29	the second one
0:08:30	nine three
0:08:31	i think and he he's there cool are
0:08:34	so with the and the more people are integer all
0:08:37	pitch frequency
0:08:38	are not exactly a query in the more people are integer of the fundamental frequency
0:08:43	mean and the white out uh from to you play around with these
0:08:48	and the order to get these uh
0:08:51	and do the thinking for that in
0:08:54	a a a a a a can with these uh
0:08:57	uh
0:08:58	and and sort them than the signal we
0:09:00	uh to minimizing
0:09:03	i the second
0:09:03	stop
0:09:04	it
0:09:04	i of the power was also
0:09:06	a we detect a peak detection now or
0:09:09	grouping
0:09:10	that
0:09:11	pitch a a a a a a a one
0:09:13	so
0:09:13	but yeah he is a large a a a a i don't
0:09:16	long detection
0:09:17	i in two D can "'cause" you don't want to point
0:09:20	you want to be the curve to it
0:09:23	so
0:09:23	well you here uh that's you in the first frame
0:09:27	but we search for a uh and these two or more
0:09:31	the second row
0:09:32	i
0:09:33	or and that the reference of any P
0:09:36	i
0:09:36	one one in in it or what any um
0:09:39	pitch and be
0:09:40	we group and to get a
0:09:42	and re
0:09:44	oh
0:09:45	a a a a a very
0:09:47	to be for another
0:09:48	and
0:09:50	or not
0:09:51	it can not be grouped into used uh
0:09:53	um
0:09:54	first
0:09:55	so no one one another core
0:09:57	a very uh uh uh
0:09:59	uh uh
0:10:00	that that for you try to each candidate
0:10:03	and
0:10:04	and and i got from five like
0:10:08	now the second stage is that
0:10:10	separate
0:10:11	so uh
0:10:13	we or that
0:10:14	do the separation be uh
0:10:16	she mean not be track
0:10:19	and then compared the
0:10:20	we you know we will try
0:10:23	if the longest track
0:10:24	yeah have in these uh
0:10:26	to to the a representation
0:10:28	i is that we do we do a right
0:10:32	and and the longest track
0:10:33	smaller than a threshold
0:10:35	he
0:10:36	sound in to one group
0:10:37	you
0:10:38	if not then you read a a of them to the sec
0:10:41	i we we basically
0:10:43	separate the uh
0:10:45	individual tracks
0:10:46	two source
0:10:49	and that the that
0:10:50	state
0:10:51	yeah
0:10:52	because
0:10:52	we have to the
0:10:54	you know
0:10:55	we have a problem my
0:10:57	there for the a or some sort of interpolation in order to the record me stand the mean
0:11:03	pitch frequencies
0:11:04	and some time here
0:11:05	you here that
0:11:07	that might be you to i'm voice signal power like is like that
0:11:13	about that
0:11:14	you to a second and the these uh
0:11:17	and data
0:11:17	using the relation
0:11:19	requiring covering
0:11:20	max
0:11:21	uh uh
0:11:23	so
0:11:24	oh
0:11:25	or are some than those from
0:11:28	we
0:11:29	tire uh a
0:11:31	he's try
0:11:32	and here this is another uh nice
0:11:35	frequency
0:11:36	ah
0:11:36	and uh we also have another uh a heuristic parameters that with
0:11:42	track or the overlapping they can be
0:11:44	don't to want source
0:11:46	so you the presence of two source
0:11:48	and i a lot of
0:11:50	the
0:11:50	oh
0:11:51	the to make the pitch contour which of the exact to the uh uh uh a reference uh one
0:11:58	no so uh that's still
0:12:00	i
0:12:01	that a E
0:12:02	can and detect on the line each one
0:12:05	so but yeah
0:12:07	but not well um
0:12:08	results
0:12:09	we are sure
0:12:11	i one ninety
0:12:13	oh
0:12:15	like you know
0:12:16	um i
0:12:17	a combination of gender
0:12:18	me mail
0:12:20	in a email
0:12:21	maybe a met
0:12:22	a we with that the uh
0:12:25	are be to interference rate you to zero to eighteen db
0:12:29	uh hamming i mean window of black
0:12:31	i it is that
0:12:32	to bring the of the ten millisecond
0:12:34	where live
0:12:36	a a a new to segment the signal
0:12:38	a reference speech
0:12:39	uh uh are a using the uh
0:12:42	talking method which is very what was five
0:12:44	and accurate
0:12:45	and uh the uh the white three
0:12:48	a very or uh maybe ross or right
0:12:51	and a your mention
0:12:52	a previously
0:12:53	a a voiced unvoiced or rate
0:12:56	and
0:12:56	separation error
0:12:57	we compare this the with the uh one of the or a the to was back
0:13:02	de leon wiring groups that have
0:13:05	um
0:13:06	um there
0:13:07	have a have applied some sort of gammatone few trained with a channel
0:13:11	and and another at that or or or or a a a proposed by captain in night
0:13:15	for
0:13:16	of course
0:13:16	a sort of a
0:13:18	harmonic suppression or
0:13:20	for the so
0:13:21	yeah
0:13:22	or or or a result uh
0:13:24	for error rate versus the target to interference ratio
0:13:28	ah
0:13:29	have three sets of lot
0:13:31	which how one so the result for target
0:13:35	and the battle for
0:13:37	we are and
0:13:38	um
0:13:39	we have to lines here a dynamo you go
0:13:42	and the uh
0:13:44	so that one and um and and uh
0:13:46	for one of those and
0:13:48	uh the proposed method
0:13:50	uh
0:13:50	the the that we but that's stand for the uh
0:13:54	hmmm
0:13:55	and
0:13:55	i have some good
0:13:57	you know but
0:13:58	and it can for all
0:14:00	can be nation of mixtures
0:14:02	a and five factor
0:14:04	a a a from the two other techniques um
0:14:07	so if we can see
0:14:08	uh
0:14:09	and there for the target
0:14:11	if we a signal
0:14:12	L
0:14:13	so
0:14:14	a a you is to a a voice as all
0:14:17	uh
0:14:19	he to incorporate these he met that you can be it and propose anything for
0:14:23	the in the unvoiced the only work who worked know so we have to in here
0:14:29	um
0:14:30	a
0:14:31	kinetic
0:14:31	see
0:14:32	a a a and it's factor uh
0:14:35	oh a very well fit to the other that
0:14:37	for all combination me male
0:14:40	in in a minute
0:14:41	and male email
0:14:42	um week
0:14:44	and the point only in terms of uh
0:14:47	separation error
0:14:49	uh
0:14:50	we we see a you
0:14:51	i two method
0:14:53	uh
0:14:53	how our method
0:14:55	very robust against one
0:14:57	a a a a a sort of the
0:14:59	to like or
0:15:00	and uh uh uh you get in separation performance for to uh
0:15:05	a method
0:15:07	yeah so uh
0:15:09	there are a number of issues that should be risk
0:15:12	a about this for uh so
0:15:15	i have a problem but two pitch contours are are crossing each other how
0:15:19	we can assign and two different sorts of than the pitch contours are very close
0:15:24	a green
0:15:25	i don't know even
0:15:26	oh uh uh are the to system can separate them are our uh we are working to improve the performance
0:15:32	in by applying some prior knowledge about a speakers
0:15:36	i believe we can also apply the spatial diversity another at another clue
0:15:41	we can uh yeah i meant to to do small and improve the for one
0:15:45	and some prior knowledge about
0:15:47	the
0:15:48	there uh
0:15:49	i been working on a bayesian inference method
0:15:52	the performance
0:15:53	yeah
0:15:53	i would like to time is uh
0:15:56	a a called me
0:15:58	not
0:15:58	and the only one who provided codes
0:16:01	i three D for a a a a researcher or so
0:16:04	for really how to compare
0:16:07	a with a
0:16:08	so
0:16:09	and that
0:16:09	ah
0:16:11	right now week
0:16:12	the whole of the code it's some demos from a my page
0:16:16	i is
0:16:17	really
0:16:18	and that one so
0:16:19	and now finally to you uh
0:16:22	a to taking any question or comment about
0:16:32	i
0:16:48	but you come in a little bit i
0:16:50	do you mean in terms of separation
0:16:52	a sequential grouping problem of interest
0:16:55	i think that that just to they're not actual
0:16:57	separation and six that's right means voice
0:16:59	yes
0:17:00	the uh
0:17:02	a pitch you want to do
0:17:05	a speaker
0:17:06	and and tracker as signs it to the seconds
0:17:09	that as the separation
0:17:11	yeah i a this classification be two class
0:17:14	and i i i i are or correctly that we would also some that's make any attempt
0:17:18	to
0:17:18	to solve the problem
0:17:19	that
0:17:21	a i i i i one and that method that it if you need to do a different at contours
0:17:26	for
0:17:27	to
0:17:28	no
0:17:29	sure
0:17:30	and
0:17:31	yeah yeah maybe a
0:17:33	and a
0:17:34	we then and can the feast five
0:17:37	the fact that you're are getting two contours from two speakers i got it at home
0:17:42	it sounds
0:17:43	okay
0:17:47	so something about two you of to look at it that you like to
0:17:50	translating to do most mcclay speak
0:17:53	a charter
0:17:53	to to look at the model look at that you
0:17:56	use
0:17:57	was that translates to like C
0:17:59	and is that being online have uh
0:18:04	just it something i duration of tracks so you you okay yeah i i i i
0:18:09	uh marking cooking of S
0:18:11	a i think it duration of sentences are about seven
0:18:15	to two sec
0:18:17	two to sec
0:18:22	yeah
0:18:23	oh
0:18:26	i
0:18:30	i
0:18:31	i
0:18:37	but the method automatically a as the and voices
0:18:41	so when you have these uh
0:18:46	five
0:18:48	the
0:18:49	so uh
0:18:50	yeah as and what
0:18:52	so we
0:18:54	reference
0:18:55	a a a a a i i the uh
0:18:58	and with anything specifically to recognise and work
0:19:02	a fact that you don't have a
0:19:04	a a to here
0:19:05	uh
0:19:06	yeah i that we don't have any
0:19:08	yeah
0:19:09	which
0:19:11	i think yeah
0:19:13	i thank you very much your

MPTRACKER: A NEW MULTI-PITCH DETECTION AND SEPARATION ALGORITHM FOR MIXED SPEECH SIGNALS

Speech Analysis

Presented by: Hossein Radfar, Author(s): Hossein Radfar, University of Toronto, Canada; R. M. Dansereau, Carleton University, Canada; Wai-Yip Chan, Queen's University Belfast, Canada; W. Wong, University of Toronto, Canada