Přepis řeči - Hunting for Wolves in Speaker Recognition

0:00:06	okay um
0:00:08	and he said my name is laura so i'm a P H E student at the university of california berkeley
0:00:14	and i also work at the international
0:00:16	computer science
0:00:17	institute or icsi um as many of you know
0:00:20	and i would like to
0:00:20	just uh
0:00:21	also acknowledge my caught and charged ions it
0:00:24	who was uh
0:00:25	a fundamental part of of this work
0:00:28	um
0:00:30	so i just uh
0:00:31	quick overview of pretty standard start out with a
0:00:35	what we're trying to do and why
0:00:37	uh we just got
0:00:38	um related work
0:00:40	um and go our our approach
0:00:42	to the problem uh give you the results
0:00:45	uh
0:00:45	do a little bit additional analysis
0:00:47	and conclude um
0:00:49	with a summary and
0:00:51	future work
0:00:55	so i think we can
0:00:56	all agree that uh automatic speaker recognition
0:00:59	that some performance
0:01:00	depends on a number of factors
0:01:02	uh one of watch uh are intrinsic speaker characteristics
0:01:07	um
0:01:07	so there's no designs that if
0:01:10	you know as humans we notice that certain sneaker sound more like
0:01:13	similarly there's
0:01:14	uh
0:01:16	and you have that
0:01:17	uh
0:01:17	system's automatic systems will perform better or worse
0:01:21	for different speakers
0:01:22	um
0:01:23	so the goal of this work
0:01:25	it's too project watch
0:01:27	speaker pairs
0:01:28	well be difficult for automatic
0:01:30	speaker recognition systems two distinct
0:01:32	uh we did some preliminary work um
0:01:35	what yeah that
0:01:37	speaker pairs that are are hard for one system are also hard for others
0:01:41	it's and
0:01:42	um
0:01:43	and of course you can use the system and
0:01:45	select speaker pairs and you'll probably do really well
0:01:48	but uh we wanted to
0:01:49	it's a away from using any one system and and said um
0:01:53	have a general approach
0:01:55	and just use features that will hopefully capture uh
0:01:59	oh
0:02:00	degree of
0:02:00	uh
0:02:01	speaker similarity
0:02:03	um
0:02:04	the motivation
0:02:05	uh
0:02:06	besides
0:02:07	being an interesting task
0:02:09	it is
0:02:09	to potentially better focus
0:02:11	uh
0:02:12	the research and reduce the amount of data needed to estimate
0:02:15	system performance but
0:02:21	so there's a couple times of
0:02:22	related work um the first has to do with the idea of different speakers uh causing different problem
0:02:28	um the infamous
0:02:30	uh
0:02:30	george washington
0:02:31	you paper um
0:02:33	categorise speakers based on system performance
0:02:37	so you have
0:02:39	fig out
0:02:40	um
0:02:41	you call the large number of false rejections as target speakers
0:02:46	you have lamb
0:02:47	uh who cause a large number of false acceptances as target speakers
0:02:53	well as you call a large number of false acceptances and impostor speakers
0:02:58	and finally or default well behaved
0:03:01	she
0:03:02	um
0:03:03	in this work
0:03:04	uh we don't actually distinguish between them available
0:03:07	uh
0:03:08	since we're looking at speaker pairs
0:03:09	um
0:03:11	but
0:03:11	we want more on the title because
0:03:13	hunting for lance didn't
0:03:15	didn't sound so good
0:03:18	um a couple other
0:03:20	there's been other other work done on uh dealing with these speakers
0:03:24	but it may be difficult
0:03:25	um there's been work that shown that
0:03:28	oh their performance differences
0:03:30	between high and low pitch
0:03:31	speakers
0:03:32	um and then there been some uh worked on
0:03:35	that
0:03:36	uh tried
0:03:37	two
0:03:37	uh
0:03:38	well the method to deal with this
0:03:40	problem speakers
0:03:44	the other elements of related work
0:03:46	uh that's relevant
0:03:47	is
0:03:47	um
0:03:48	and the whole
0:03:49	features that are used
0:03:50	two
0:03:51	describe speakers or characterise speakers
0:03:53	so you can draw varies from a lot of different types of work obviously uh
0:03:57	speaker recognition approaches have use a variety of features
0:04:01	uh certainly not an exhaustive list here but things like pitch and energy distributions are dynamic
0:04:06	um
0:04:07	prosodic statistics
0:04:09	uh jitter and shimmer
0:04:11	and
0:04:12	in looking at
0:04:13	perceptual speaker characterisation or discrimination your find
0:04:16	a lot of formant frequencies and bandwidths and dynamic features
0:04:20	and um
0:04:23	other acoustic parameters that influence voice individuality include the pitch frequency contour and fluctuation
0:04:29	again the formant frequencies and long term average but
0:04:36	so our approaches
0:04:37	as
0:04:38	fairly straightforward
0:04:39	um basically we compute feature values over some speech data
0:04:44	uh
0:04:44	corresponding to marry speaker
0:04:46	and then using these feature values compute a measure similarity uh for all speaker pairs
0:04:53	and
0:04:54	the and looking at these measures look at the uh
0:04:58	speaker pairs
0:04:59	that have the highest and now we
0:05:01	um
0:05:02	values
0:05:03	in terms of these
0:05:04	uh similarity measures
0:05:05	and compare for performance and those speakers
0:05:08	uh to all
0:05:14	so the features
0:05:15	we consider here uh first of all
0:05:17	pitch
0:05:18	that sadistic
0:05:19	um i mean median
0:05:21	range and mean average slope
0:05:24	much we
0:05:25	you know
0:05:25	okay
0:05:26	um
0:05:27	jitter and shimmer are the relative at an
0:05:30	the average perturbation of generic
0:05:32	and a five point amplitude perturbation quite
0:05:34	question version
0:05:39	uh formant frequency statistics
0:05:41	you mean and median of the first three formant
0:05:43	um
0:05:44	i'll be doing that we work with it
0:05:46	eight kilohertz
0:05:47	so um
0:05:48	although
0:05:48	higher formants
0:05:49	might be useful we we didn't calculate them here
0:05:55	uh and he nonunion energy
0:06:00	uh long term average
0:06:02	spectrum energy statistics
0:06:03	uh including the mean standard deviation
0:06:06	range
0:06:07	slow
0:06:07	and local peak day
0:06:10	um
0:06:11	and we did a fourteenth order lpc analysis
0:06:14	and uh found the frequencies
0:06:17	from
0:06:17	the coefficient right
0:06:19	uh both with and without a minimum magnitude requirement which is essentially a limiting the bandwidth
0:06:26	and uh then we to calling frequency it and
0:06:28	in the middle
0:06:29	histogram
0:06:32	and finally we have B mode and median
0:06:35	spectral
0:06:41	so we have
0:06:42	all these features well what measures do we use
0:06:45	um for the scalar features of what
0:06:47	almost all of them are
0:06:49	uh we simply took the absolute or percent difference
0:06:54	um also we in addition to using the formant frequencies individually we looked at some of the formant frequencies
0:07:02	and we also looked at doctors of
0:07:04	formant frequencies and
0:07:05	but euclidean distance between the vector
0:07:08	and finally for the histograms of frequencies
0:07:11	uh
0:07:12	we calculated the correlation
0:07:15	as a matter
0:07:16	so there's there's
0:07:17	two different ways you can compute the single measure for speaker pair
0:07:21	uh the first is
0:07:23	to take for every speaker
0:07:24	take all their uh feature values over the
0:07:27	conversation sides but are available
0:07:29	and just
0:07:30	get an average feature value over the conversation
0:07:33	and then compute the measure between these average values for each speaker
0:07:38	um the other approaches to take
0:07:40	and the conversation by conversation basis between two speaker pairs
0:07:44	for two speakers
0:07:45	uh compute the distance measure first and then averaged over the conversation pairs
0:07:51	and um
0:07:52	and the result types and i just present whatever method
0:07:56	gave
0:07:56	better
0:07:59	larger different
0:08:05	so the data that we use
0:08:07	um
0:08:08	really feature measure calculation and
0:08:11	speaker pair selection
0:08:12	uh we use p2p of neatness
0:08:15	followup evaluation data
0:08:17	um so this is all interview data
0:08:20	um
0:08:21	which is recorded on microphones me limit it to just
0:08:24	you have a your microphone
0:08:26	quality purposes
0:08:28	and uh just uh the sign out um almost all the speakers have four conversations available
0:08:33	um there's a handful with
0:08:35	three or five
0:08:36	but
0:08:37	um
0:08:38	because this is
0:08:39	that's multiple conversation
0:08:41	one
0:08:42	and then once we have the speaker pairs selected um
0:08:46	we evaluate performance using the uh data from the
0:08:50	nist two thousand the evaluation short too short three condition
0:08:54	uh so this
0:08:55	data varies from the
0:08:57	uh other did the pilot data in
0:08:59	a couple respect
0:09:01	um
0:09:02	in addition to
0:09:02	possibly being an interview
0:09:04	uh it can also be
0:09:06	speech from a telephone conversation
0:09:09	and in addition to having uh the lab of the your microphone
0:09:12	channel there are other microphones
0:09:14	and as well the telephone
0:09:22	so um
0:09:23	available we had uh
0:09:26	submissions that were shared by participating site
0:09:29	and so thank you to everyone who share their submission
0:09:32	um
0:09:33	be sure to short precondition originally had i think maybe around ninety thousand miles or so um
0:09:39	i had to remove the child that
0:09:41	correspond to speakers that weren't in the selection data
0:09:44	and that's you that your left
0:09:46	about fifty five thousand trial
0:09:49	and then furthermore when you just
0:09:51	sub select
0:09:51	oh and only keep trials corresponding to some percentage of speaker pairs
0:09:56	uh you got around four thousand or eleven thousand trials love
0:10:00	i know
0:10:02	and
0:10:03	we only keep target trials for speakers to show up in one of the
0:10:11	so how do we evaluate the system performance
0:10:14	um
0:10:15	they're various metrics you can use what we look at here are to be uh
0:10:19	minimum detection cost function
0:10:21	and
0:10:21	which of course is the
0:10:23	a weighted
0:10:24	some of uh
0:10:26	with relative weights
0:10:27	for errors
0:10:29	uh what this is all done with the um
0:10:31	two thousand a cost
0:10:32	so it's not the low false alarm like the two thousand ten evaluation
0:10:37	and then since we're looking at
0:10:38	impostor speaker pairs uh we look at
0:10:41	T false alarm rate which of course is
0:10:43	simply
0:10:44	for a given decision threshold the number of false alarm errors
0:10:49	that occur are out of it
0:10:50	total number of possible
0:10:52	target track nine target
0:10:56	so
0:10:58	for every other system submission that we have
0:11:01	we first
0:11:01	uh
0:11:03	just looking at the trials for the most
0:11:05	or least similar speaker pairs
0:11:07	we can keep the change in dcf relative to
0:11:10	what it is for all speaker pairs
0:11:13	and then uh
0:11:15	take all these
0:11:16	um
0:11:17	system
0:11:18	differences and average over the system
0:11:20	so the results are just
0:11:22	to
0:11:23	a typical overall try
0:11:24	and
0:11:25	from the system
0:11:26	um
0:11:27	and then with the false alarm rate we
0:11:29	uh for the all speakers' we look at a decision threshold that
0:11:33	uh generates a false alarm rate of one
0:11:35	sign
0:11:35	and then at the same decision threshold um
0:11:39	see what the false alarm rate is for the most and least
0:11:41	similar speaker pairs
0:11:44	and of course if we're actually taking uh if more similar speaker pairs
0:11:48	actually corresponds to
0:11:50	more difficult to distinguish speaker pairs
0:11:52	then we expect these changes
0:11:54	and the dcf
0:11:55	and
0:11:56	false alarm rate to be
0:12:02	so
0:12:03	here the results uh when you look at
0:12:05	one
0:12:05	sound of speaker pairs
0:12:07	uh in each case
0:12:08	the
0:12:09	top row
0:12:10	corresponds
0:12:11	to uh
0:12:12	the
0:12:13	least similar speaker pairs
0:12:15	so
0:12:16	performance is improving
0:12:18	and this road is
0:12:19	um
0:12:20	corresponds to most
0:12:22	similar speaker pairs so that
0:12:24	the
0:12:24	performance is getting worse
0:12:26	um
0:12:27	so we notice that we are able to find uh
0:12:30	features and and measures that
0:12:33	we also like
0:12:34	uh
0:12:35	speaker pairs
0:12:36	with the desired
0:12:37	um
0:12:37	and and see
0:12:39	um
0:12:41	if we then
0:12:42	compare uh performance on one side
0:12:45	to performance on five percent you can see that it's less
0:12:49	pronounced when you include
0:12:50	more speaker pairs
0:12:52	um
0:12:53	in some cases you
0:12:55	you have these negative
0:12:56	or
0:12:57	opposite trends from what you
0:12:58	back
0:13:04	oh i i
0:13:05	i pretty much mention all these points the only thing to note is that
0:13:08	um
0:13:09	changes in performance are not uniform
0:13:12	across site submissions
0:13:13	so that is
0:13:14	uh
0:13:14	one issue
0:13:20	um okay so here's would adopt her for one system um when we use
0:13:25	uh be euclidean distance between
0:13:28	uh vectors of the first three formant
0:13:31	uh dyslexic
0:13:32	figure pair
0:13:33	um
0:13:34	you solid line
0:13:36	correspond to uh the most similar speaker pairs being you
0:13:40	and the dashed lines are very similar
0:13:43	uh right is one percent and green is five percent
0:13:47	and the black line is
0:13:48	the
0:13:49	uh case for all speaker pair
0:13:52	um
0:13:54	we know it is
0:13:57	that
0:13:58	uh in this particular instance
0:14:00	um
0:14:00	there's
0:14:02	a bigger difference
0:14:03	uh when looking at
0:14:04	the
0:14:05	uh
0:14:06	leave
0:14:06	similar
0:14:07	speaker pairs
0:14:08	and the most similar speaker pairs
0:14:11	uh are much closer to
0:14:12	formance overall speaker
0:14:15	um
0:14:16	although that doesn't happen all of the time it is
0:14:19	uh
0:14:20	certainly the general tendency
0:14:22	uh
0:14:22	to have this larger
0:14:23	larger gap in this direction
0:14:29	and
0:14:30	here's another
0:14:31	example that
0:14:31	shows
0:14:33	that it doesn't always hold
0:14:34	um
0:14:36	and this is
0:14:36	uh a different system and a different feature measure that's the percent difference of median energy in this case
0:14:43	and um
0:14:48	you you get better separation here
0:14:50	uh it there is and how much separation there is
0:14:53	and
0:14:56	across
0:14:58	so
0:14:58	we've been able to do some stuff but we expect we could probably do even better
0:15:03	if we use uh more knowledge
0:15:05	speaker system
0:15:06	so
0:15:08	um
0:15:09	we decided to just
0:15:10	simply use gmm
0:15:12	since uh they show up obviously in a lot of
0:15:15	um system
0:15:16	um so we adapted uh
0:15:19	speaker specific gmm
0:15:21	um and then calculated the
0:15:24	uh
0:15:25	cal divergence
0:15:26	between them has to be measured speaker similarity
0:15:30	when we do that
0:15:31	not surprisingly we get uh
0:15:33	better results
0:15:35	um
0:15:36	the previous
0:15:37	charts all had just one from negative fifty percent to fifty percent
0:15:40	so you can see already that
0:15:42	there are larger difference
0:15:44	watches as i said what we would expect
0:15:49	um
0:15:50	here's
0:15:51	that curve for a system using the key algorithm
0:15:55	um
0:15:57	again you can see that these are larger differences from the all performance
0:16:02	and we again
0:16:04	uh see this asymmetry where
0:16:07	uh
0:16:09	a bigger gap
0:16:10	for the dissimilar pairs one for the somewhere
0:16:20	so as i mentioned uh we're
0:16:23	tend to be more successful at selecting easy to distinguish speaker pairs
0:16:27	uh
0:16:28	and possibly because these pairs may be easier to
0:16:31	fine
0:16:32	um
0:16:32	one possible explanation would be that
0:16:35	if you have a a speaker pair
0:16:37	that is very dissimilar
0:16:39	and um
0:16:40	terms of pitch or formant frequencies
0:16:42	that
0:16:43	you know a big difference is probably going to mean that
0:16:46	the system is not going to
0:16:47	to use them
0:16:48	um but on the flip side if you're trying to figure out what makes the speaker pair
0:16:52	difficult um just like you know any single feature may not be enough
0:16:56	to capture
0:16:58	um
0:16:59	that information
0:17:06	so using them the K L divergence measure
0:17:10	we took a closer look at
0:17:11	uh the speaker pairs but are selected
0:17:14	and
0:17:15	as we in the mostly similar so in addition to like you know the one percent by
0:17:19	the group
0:17:20	uh also looked at
0:17:21	three percent ten percent twenty percent
0:17:24	and um and this data and there were a hundred fifty speakers overall
0:17:29	uh
0:17:29	leading to uh uh eighteen hundred
0:17:32	unique speaker pair
0:17:34	for same sex
0:17:35	first
0:17:38	and uh one thing we noted
0:17:40	is that
0:17:41	in
0:17:42	in the groups of me
0:17:44	uh least similar speaker pairs
0:17:47	if you look at a group
0:17:49	with what the larger values
0:17:50	of the divergence
0:17:52	um
0:17:53	we would expect
0:17:54	to be easier to distinguish
0:17:56	the majority of them are male
0:17:57	um
0:17:58	but
0:17:59	if you look at any one group about seventy five percent of
0:18:01	speaker pairs in the group will be
0:18:03	uh mail
0:18:04	on average
0:18:10	uh to a lesser extent we notice the opposite tendency when we look at
0:18:14	uh
0:18:16	more similar speaker pairs which
0:18:18	uh somewhat tend to be more female
0:18:21	um
0:18:21	the
0:18:23	one
0:18:24	and
0:18:24	three percent
0:18:25	still have more male pairs
0:18:26	but
0:18:28	the other group
0:18:29	have more female
0:18:34	um
0:18:35	this
0:18:36	you know maybe part of the reason why uh
0:18:38	system performance
0:18:39	typically better
0:18:40	and male
0:18:41	um
0:18:42	and it just in that you know males may may
0:18:45	exhibit a greater range of differences
0:18:48	between them
0:18:49	so that
0:18:50	there are likely to be more
0:18:51	the similar
0:18:52	a male speaker
0:18:59	and finally so looking at these groups
0:19:02	um we notice that there is a tendency define two types
0:19:05	speaker
0:19:06	uh there are speakers who frequently appear as members of difficult to distinguish
0:19:11	uh
0:19:11	speaker pairs
0:19:12	and speakers who occur frequently as members of
0:19:15	easy to distinguish speakers
0:19:18	um
0:19:20	in fact there are fifteen speakers you never appear in the most
0:19:23	similar group
0:19:25	and twenty four speakers you never appear in the most
0:19:28	dissimilar group
0:19:30	um
0:19:32	i forgot
0:19:32	but i think this is
0:19:33	this twenty four speakers there's ten male and forty female
0:19:40	so this
0:19:41	uh tends to support the idea that there are these walls in there
0:19:45	uh
0:19:45	speakers who are
0:19:47	are more difficult
0:19:48	um
0:19:48	or more similar to other speakers
0:19:55	so just a summary of what i mentioned
0:19:57	uh first of all it is possible
0:19:59	project
0:20:00	uh what speaker pairs will be difficult for a
0:20:03	typical speaker recognition system
0:20:04	to distinguish
0:20:06	um
0:20:07	for the features
0:20:08	that we considered here would catch
0:20:10	formant frequency of the the best ones seem to be uh the the euclidean dist
0:20:14	between the first
0:20:15	uh three formant frequency
0:20:18	um but the best measure overall was the more uh complex
0:20:21	uh
0:20:22	cal divergence measure between
0:20:24	uh
0:20:25	speakers this
0:20:25	fig gmm
0:20:27	um i mentioned of course that we're typically more successful at identifying dissimilar speaker pairs
0:20:33	and that in addition to
0:20:35	to being able to um
0:20:37	you know finding speaker pairs
0:20:39	uh
0:20:40	using these measures can provide potentially useful information
0:20:43	about a speaker's tendency to be
0:20:45	similar or dissimilar to other
0:20:52	so future work
0:20:53	um
0:20:55	one thing to try is testing combinations of multiple feature measures
0:21:00	because the the method for selecting similar speaker pairs
0:21:03	um i did a little bit of work on this
0:21:06	uh where i just
0:21:07	basically
0:21:08	assigned a rank
0:21:09	according to each
0:21:10	uh feature manager and then some of the ring
0:21:12	over the speaker pairs and and
0:21:14	did selection that way
0:21:16	and and that that improve
0:21:18	result
0:21:19	uh another extension is to um instead of focusing on impostor speaker pairs
0:21:24	see if you can find uh figure out what
0:21:27	target speakers will be difficult
0:21:28	uh for the system to correctly right
0:21:32	and one thing but um certainly needs to be investigated is
0:21:36	uh
0:21:37	the you lack
0:21:38	assistant see uh
0:21:39	behaviour for the things but of speakers across
0:21:42	um
0:21:43	different
0:21:44	uh
0:21:45	system
0:21:46	um
0:21:46	we may be able to find potential trend
0:21:49	in behaviour across classes or types of stuff
0:21:52	of course with
0:21:53	uh
0:21:54	the
0:21:55	site submissions that we used here uh
0:21:59	almost all of the submissions are in fact
0:22:01	fusion of multiple systems
0:22:03	so might need to do a more of a breakdown
0:22:05	uh to
0:22:07	um
0:22:08	really get out
0:22:09	that
0:22:10	sure
0:22:13	okay
0:22:14	that's all i have thank you
0:22:16	hmmm
0:22:24	sh
0:22:38	thank you larry
0:22:39	presentation
0:22:41	i have a question about your
0:22:43	formant extraction
0:22:44	uh
0:22:45	do you have
0:22:46	don
0:22:47	nation
0:22:48	for all the old volumes
0:22:50	or
0:22:51	did you controls that's your extraction
0:22:53	uh
0:22:54	to the east
0:22:55	you you you
0:22:56	does volume
0:22:57	or
0:22:58	four
0:22:59	different
0:22:59	the
0:23:00	you do
0:23:01	one type of volume
0:23:03	because you know
0:23:04	it's
0:23:04	my question is
0:23:06	it is uh
0:23:07	this
0:23:07	use the volume
0:23:09	or you you do
0:23:11	um extraction according to the volume
0:23:14	now so uh we didn't it was just the over the entire file so it's it's definitely you could probably
0:23:19	get much better estimates
0:23:21	and what we what we actually did
0:23:23	because
0:23:24	the the the problem is that
0:23:26	uh
0:23:27	uh you have a lot of disturbance according to the volume for now
0:23:30	so
0:23:31	uh but i think that
0:23:33	uh
0:23:34	i think that it is more the sample
0:23:36	and
0:23:36	no
0:23:37	phonological information
0:23:39	that's value
0:23:40	yeah
0:23:40	the speaker information
0:23:42	so of course uh yeah of course this is
0:23:44	convolving the the phonetic the
0:23:46	phonetic
0:23:47	with
0:23:47	with the speaker
0:23:48	okay thank you
0:23:55	oh
0:23:57	oh
0:23:58	oh
0:23:59	hmmm
0:24:01	oh
0:24:05	oh
0:24:06	oh
0:24:07	oh
0:24:08	two
0:24:11	what
0:24:12	oh
0:24:13	hmmm
0:24:14	oh
0:24:16	oh
0:24:17	oh
0:24:18	oh
0:24:19	oh
0:24:20	i
0:24:21	pairs of what
0:24:22	oh
0:24:23	extracts or just
0:24:25	oh
0:24:26	uh_huh
0:24:28	oh
0:24:29	oh
0:24:31	oh
0:24:32	oh
0:24:33	sure
0:24:36	oh
0:24:37	fig
0:24:39	oh
0:24:39	two
0:24:40	oh
0:24:41	oh
0:24:42	oh
0:24:42	hmmm
0:24:43	four
0:24:44	oh
0:24:45	oh
0:24:47	or
0:24:48	oh
0:24:52	oh
0:24:54	sure
0:24:56	sure
0:24:59	thanks
0:25:04	hmmm
0:25:06	no
0:25:08	we will
0:25:08	we should
0:25:09	uh
0:25:11	four
0:25:12	the
0:25:12	the
0:25:13	the
0:25:14	yeah
0:25:16	uh
0:25:17	you
0:25:19	uh
0:25:19	uh
0:25:22	uh
0:25:24	right
0:25:25	uh
0:25:25	good
0:25:27	oh
0:25:28	i'm sorry
0:25:30	are you talking about
0:25:32	i
0:25:33	figure
0:25:33	paper
0:25:34	and i think you mean
0:25:36	uh
0:25:38	oh
0:25:40	cool
0:25:42	yeah
0:25:44	yeah i know but it was definitely the case but it was
0:25:48	oh
0:25:49	right
0:25:52	hmmm
0:25:53	hmmm
0:25:58	oh

Hunting for Wolves in Speaker Recognition

SESSION 7: Speaker and Language recognition - Evaluations and performance testing

Přidáno: 14. 7. 2010 11:08, Autor: Lara Stoll (International Computer Science Institute and UC-Berkeley), George Doddington (), Délka: 0:25:59