Speech Transcript - A TOOLKIT FOR SPEECH RECOGNITION RESEARCH

0:00:26	a a a i i i
0:00:27	no
0:00:30	i
0:00:31	a
0:00:32	i you know what
0:00:33	a i today
0:00:35	and people
0:00:36	and some people would be you know a to up to have an another's with
0:00:40	they on if an sup of for the rest of the live
0:00:43	i thinking that those prophecies is of been just slightly misinterpreted
0:00:47	and the event that they were referring to is this
0:00:49	a wonderful speech to okay
0:00:51	i
0:00:53	a that have in of actually that
0:00:55	you know it almost buys review thing it
0:00:57	so uh
0:00:59	i do i think that have anything and their estimates the significance of this that
0:01:04	a that's okay
0:01:05	okay
0:01:06	so first they just about the name
0:01:08	it's a it's some kind of coffee reference hence the little coffee being with uh
0:01:13	but had so
0:01:15	a
0:01:16	but
0:01:16	is just
0:01:17	whatever name we thought to
0:01:20	so uh
0:01:21	the structure of this uh this whole presentation is fess i'm gonna talk
0:01:25	for about
0:01:26	fifteen or twenty minute
0:01:28	just giving you know of you kind of from all sides of this tool K
0:01:32	and then we're gonna a people to escape in case they don't want to know more details than the have
0:01:36	a short break
0:01:37	and then
0:01:39	i not and uh on drug going to talk about a uh
0:01:43	some more called local stuff like
0:01:45	and i was gonna talk about some of the acoustic modeling code
0:01:48	and we'll talk about the uh matrix like
0:01:51	which just kind of independent useful
0:01:53	uh
0:01:54	speech
0:01:55	and then after that
0:01:56	uh
0:01:57	i'm gonna go through some example scripts that we have been try to get people
0:02:01	more of a you know
0:02:02	give people a sense of of how to use that
0:02:06	now
0:02:07	or the next slide
0:02:09	so
0:02:10	some important aspect of the project is it the
0:02:13	it's license under a you V two point uh which is the
0:02:17	style a license that basically allows it to do anything you want with it
0:02:21	there is only a uh
0:02:23	an acknowledgement a
0:02:25	close which as you have to acknowledge that
0:02:27	the code came from that but that
0:02:29	that's of that's
0:02:31	it's it's one of the most open up the standard lies
0:02:35	uh
0:02:36	the project of currently hosted on source forge which is the
0:02:39	standard place for these kinds of open source project
0:02:43	uh
0:02:44	we we it
0:02:45	some talk it's a very closely associated with a particular institution
0:02:49	our attention is for it to be more of a kind of
0:02:52	thing that lives
0:02:54	and the clouds out or and source for
0:02:56	i i shouldn't have use that will that that's to
0:02:58	that's just gratuitous that
0:03:00	but it yeah there it's very for it not to just be him a
0:03:04	the pet project of some particular little group but uh
0:03:07	that's to represent
0:03:08	the best of what's out there and and and we will can be participants as long as you can contribute
0:03:12	code under
0:03:13	this slice sense than that's great
0:03:16	uh
0:03:17	it's basically a C plus plus to at
0:03:19	the code compiles it a native windows and
0:03:22	and the common units but fun like can we're not claiming that a compile once on or you know
0:03:28	other we're problem but but it compiled from on the normal one
0:03:32	a
0:03:34	you have some documentation not as much as takes T K
0:03:37	and and and we have example script
0:03:40	these example scripts and not uh
0:03:42	there just for results also as one and and uh
0:03:45	wall street journal
0:03:46	but we're gonna have more to
0:03:48	they
0:03:50	they basically run from ldc that's
0:03:52	so once you have the this you can kind of point them to the disk
0:03:55	and just
0:03:56	get an idea of how it work
0:04:00	so
0:04:04	oh no i now i realise that we didn't look a large enough row
0:04:08	i think i think we just have a tie this thing to uh aggressively
0:04:12	if these were not guy
0:04:14	uh
0:04:15	yeah
0:04:16	so
0:04:20	okay somehow out i gonna go through the kind of a think that support this is just the current features
0:04:24	obviously
0:04:25	or tending to a lot more
0:04:27	so you can build a standard context-dependent uh
0:04:30	lvcsr system
0:04:32	you know with tree clustering
0:04:34	in that it's been written in such a way that it supports arbitrary context size is so you can go
0:04:39	to
0:04:39	quint phone oh what's have and it will uh
0:04:42	a work
0:04:43	without without pain
0:04:45	but the the training coding about fst based on a
0:04:49	our code compiled against openfst
0:04:52	for those of you who don't know up fst is
0:04:55	it's kind of like the eighteen T tells set it's open source
0:04:58	it's uh
0:04:59	a project uh
0:05:00	like google and some other
0:05:04	um
0:05:06	we can only only have max and like the had training
0:05:09	we haven't yet done lattice generation but at time
0:05:12	timeline line for adding discriminative training and lattice generation
0:05:15	and
0:05:16	this summer slash
0:05:18	like
0:05:20	uh
0:05:21	we we we support all kinds of linear and affine transforms you can imagine
0:05:25	i don't not all of these
0:05:27	necessarily involve uh
0:05:29	you know that tree version
0:05:30	what where you have a
0:05:32	multiple regression plot
0:05:34	that's just because we
0:05:36	are trying to avoid very complicated frameworks that would make that so difficult to use
0:05:41	so a lot of these just support point a single transform
0:05:45	we
0:05:45	all of these things also have examples scrip
0:05:48	so it's not just something that's in the code that
0:05:51	that we know work
0:05:51	something that you can also
0:05:53	get to
0:05:55	so
0:05:57	and trying to have a i did want to just is other tool kits as a little disclaimer here
0:06:01	that
0:06:02	we're not claiming that all of tool kids don't have any of these advantages to
0:06:07	but uh
0:06:08	waiting for clean coal code and modular design
0:06:12	uh
0:06:13	and and by module we we probably need something a little bit stronger than you would normally uh
0:06:19	normally imagine it's it's written in such a way that
0:06:22	it's not only easy to combine the various things that are in the
0:06:25	but it's easy to uh
0:06:27	kind of extend arbitrarily
0:06:29	and and we have avoid the kind of code where
0:06:32	when you add something
0:06:34	a bunch of other bits of code have to know about what you added then you have to modify all
0:06:38	kinds of
0:06:39	you know
0:06:40	all kinds of other
0:06:42	and
0:06:44	the part is a big uh
0:06:46	advantage i know but not a lot of uh
0:06:50	to gets such a completely free lies
0:06:53	and that that we don't really anticipate this being used for commercial purposes
0:06:57	uh
0:06:58	our understanding is that
0:06:59	a lot of research group
0:07:02	as a matter of principle they they won't
0:07:04	you stuff that has no commercially license because this say
0:07:07	is this research can the commercial by the
0:07:10	and now
0:07:11	or of the license will
0:07:14	uh
0:07:15	have example scripts which were which were uh
0:07:19	standing documentation
0:07:21	and
0:07:22	that this whole community building think that the people involved in cal is currently uh
0:07:27	it
0:07:28	it's a group of people mostly vol
0:07:30	who are to the previous to you works so
0:07:33	myself are are not a bunch of guys from but
0:07:36	and and if you others
0:07:38	case
0:07:39	uh
0:07:40	but we open to new participant
0:07:43	and uh
0:07:45	well what what what we're hoping for mainly is not just people who come to be a line not to
0:07:49	of code but
0:07:50	the people who really want to understand the whole thing
0:07:53	you can contribute a significant amount
0:07:56	um
0:07:58	the
0:07:59	it's okay is especially good for stuff that involves a lot of linear algebra
0:08:03	it has a
0:08:04	very good matrix like be the andreas going to talk about
0:08:07	so if you want to do stuff that involves a lot of a matrix and vector
0:08:12	H
0:08:13	are
0:08:14	do
0:08:15	also uh
0:08:16	of course we we compile pile against the openfst library so
0:08:20	you can do have T stuff you know at the code
0:08:23	uh
0:08:24	its built in
0:08:26	a scalable way
0:08:27	well
0:08:30	it doesn't explicitly interact with any power level is a parallel by
0:08:34	it doesn't
0:08:36	it doesn't interact with them at weird do use or or um
0:08:39	um P i i think
0:08:40	"'cause" we felt that that that would just lock it into particular kinds of system
0:08:44	so uh
0:08:45	but
0:08:46	all the a
0:08:47	it's been in in such a way that uh it should still work efficiently when everything is very large scale
0:08:53	you have a lot of day
0:08:55	our our intention is to it and all of the state-of-the-art methods
0:09:00	for lvcsr things like
0:09:01	discriminative training
0:09:03	a standard
0:09:04	all of the standard adaptation
0:09:07	uh
0:09:09	but uh i think i say
0:09:11	on the next slide
0:09:12	uh
0:09:13	something that we not kinda doing in the in the immediate future
0:09:17	it's things like online decoding which
0:09:19	what i mean by that is uh
0:09:21	where the data is coming in say from a microphone or telephone
0:09:25	and it's some kind of interactive application
0:09:27	because you could use it to do that and building a decoder isn't that hard in this framework
0:09:32	but uh
0:09:34	i basic target audience is uh
0:09:37	speech recognition researchers who want to work
0:09:40	on the speech rec oh
0:09:42	other than
0:09:43	rather than those who uh
0:09:47	oh have a mock i was learning what everyone was looking at a multiscale to enter the room um and
0:09:52	disrupted that's all right if very present
0:09:56	i
0:09:56	oh i
0:09:59	i
0:10:00	i
0:10:00	okay
0:10:02	i
0:10:04	so i
0:10:06	we set some people lately have uh
0:10:10	this become popular recently take do a kind of life unwrapper for C plus plus code
0:10:15	the idea being that you can uh
0:10:17	more easily write your script
0:10:19	however we we've avoided that approach because
0:10:22	probably because it's a hassle to do the the wrapping
0:10:25	and nobody ever understands house we were
0:10:28	probably because uh
0:10:30	it just forces people to learn a new language and
0:10:32	probably those who just want the colours think that everyone knows by
0:10:36	that
0:10:37	uh
0:10:39	so
0:10:40	we support the kind of
0:10:42	flexibility and configurable ability of that in different ways
0:10:46	but partly uh
0:10:48	i think it'll become clear later so perhaps will
0:10:53	will will will leave to lake to those ask
0:10:56	so we don't have back would training their in there are no immediate plans to do it
0:11:01	and i some people i think some people like for back for kind of religious reason
0:11:05	but uh
0:11:06	i don't believe any was demonstrated the viterbi be is worse
0:11:10	and it just so and we need to use with a be
0:11:12	for uh
0:11:14	we because you can write the alignments to this compact lee
0:11:17	and then
0:11:18	on
0:11:23	really
0:11:24	okay
0:11:27	really interesting
0:11:29	but i i i even even not let this as
0:11:32	like
0:11:32	just a single hypothesis
0:11:34	makes it if
0:11:36	okay
0:11:37	so we'll have to think about that i mean it's not like it's really hard to do
0:11:40	but it just wasn't something that we had planned
0:11:42	uh
0:11:45	one
0:11:47	uh_huh
0:11:49	oh okay
0:11:50	well it's at the state level
0:11:53	but we it's not really this the
0:11:55	i stay i mean pdf
0:11:56	index but
0:11:57	that you little bit more precise not because uh
0:12:00	you just right out the state sequence it's fine for model training but then
0:12:04	if you wanna work work the phone sequence the penny how tree work
0:12:08	it might not be implied by the state sequence of then we have these identifiers the also contain the phone
0:12:13	and the transition
0:12:15	oh it's and it's a it's an integer a list of it just but those
0:12:19	in integers
0:12:21	are not quite the states there
0:12:22	something that can be mapped to the state also to the phone
0:12:28	uh
0:12:30	so i'm just gonna describe a how this
0:12:33	came to be we had this work in two thousand nine
0:12:36	a a lot of uh focus was on
0:12:38	as G M N
0:12:41	um
0:12:42	we that the supper we we were using that some guys some brno a university of technology
0:12:47	including a on draw look at another
0:12:49	uh
0:12:50	they built this
0:12:51	uh infrastructure for uh
0:12:54	for training as gmms that was it was written in C plus plus but it rely don't he's T K
0:12:58	system
0:12:59	and i also built a a and F E F S T based code
0:13:02	so that we could be code our own C plus plus code with access to the matrix like
0:13:07	um
0:13:08	so we kind of calling that crow took D
0:13:12	and and we wanted to release that
0:13:14	recipe
0:13:15	you know in is some kind of open source way but we realise that
0:13:19	the rest P was just too hard to encapsulate because the had he's T K had our stuff
0:13:24	as a lot of script
0:13:26	so we we wanted to create something that
0:13:29	good support this stuff and was easy to encapsulate so we we an entirely new uh
0:13:35	uh
0:13:37	the next summer we were entirely new toolkit that is
0:13:41	you know that we that
0:13:43	we wanted everything to be clean and unified
0:13:45	and to have a nice use shiny C plus plus
0:13:48	speech rec my
0:13:51	i think that's the uh
0:13:53	i think that's this a
0:13:55	slides a last somewhere
0:13:57	are two thousand ten we had another workshop and or no
0:14:00	where we uh
0:14:02	that a lot of coding
0:14:03	and and the vision at that time which and i realise is very unrealistic
0:14:07	was that we
0:14:08	we have a complete working system with example script
0:14:12	you know the end of the sum
0:14:14	but that that kind of didn't really materialise a had a lot of pieces
0:14:18	but we didn't really have a complete working system so
0:14:22	after uh
0:14:24	i kind of obligated to
0:14:26	you know
0:14:27	and is the system and and and we had a help from others thing especially on that
0:14:31	and and doing a lot of coding after that
0:14:34	so
0:14:37	uh
0:14:40	when we go to the next slide
0:14:42	it's a it's only been officially really something like last week
0:14:46	that's when we actually uh got all the legal approvals and
0:14:49	put up on source forge
0:14:51	this is just a list of the people i don't think i'm gonna go through all the names
0:14:55	this the list of all the people who are rich then uh code specifically for D
0:14:59	that's of the list the people who done various other things or it's so help the in various ways
0:15:04	and
0:15:05	uh
0:15:07	i would describe exactly have for each one but i'm kind of scared i've left someone of one of these
0:15:12	lists
0:15:13	and
0:15:14	i i i i
0:15:15	i i just let you read it
0:15:18	um
0:15:20	a lot of these people are
0:15:21	have some connection to bird or you invested to of technology
0:15:25	oh people but the in uh or
0:15:28	like that
0:15:30	so that this is a
0:15:32	this is is a rather messy diagram
0:15:34	i i just wanted
0:15:36	i want to give you some idea of what the dependency structure of kaldi was but i decided to put
0:15:40	side information and to here so
0:15:42	the area of these uh
0:15:45	of these rectangles is roughly proportional to how many lines of code
0:15:49	there are
0:15:50	so
0:15:51	the these think the thing that we can pile again
0:15:54	so open a fist is the C plus plus library
0:15:57	uh
0:15:59	at let's C left that refers to the math libraries that we can pile again
0:16:03	uh
0:16:06	and
0:16:06	and the rough dependency structures thing on top of things that and on them but
0:16:10	is very approximate
0:16:11	so
0:16:13	for instance he's various
0:16:14	fst the algorithms that we've extended of an fst with
0:16:18	uh
0:16:19	stuff relating to tree clustering for decision tree
0:16:23	uh
0:16:24	that for leading to hmm topology
0:16:27	decoder decoders
0:16:29	language modeling thing this is a small box because really all it does is uh
0:16:34	compile a marketing to enough
0:16:37	i two that
0:16:38	uh
0:16:41	you tell this at this is mostly i O stuff as various frameworks for io
0:16:45	that
0:16:46	will be explained later run kind of after a break
0:16:49	so we can allow people
0:16:50	skate
0:16:51	this is the matrix like we so this
0:16:54	a lot of this is just wrappers for stuff that's here
0:16:57	but if any i don't know if any of you are familiar with
0:16:59	with the steal a pack and blast and those things
0:17:02	but their C libraries that
0:17:04	for C plus plus program a slightly painful to work with "'cause" they have all of these arguments like the
0:17:09	rose the columns
0:17:10	tried
0:17:11	and the thing you wanna do is this very long line of code
0:17:14	and uh
0:17:16	so there's no notion of like a matrix as an object
0:17:18	so this kind of ad that abstraction and it is it is significantly easier to use
0:17:23	then of this make the
0:17:26	uh
0:17:27	this is feed sure
0:17:29	preprocessing and you know
0:17:30	going from a web file to mfcc that's that's fair
0:17:34	uh gaussian mixture models a diagonal and full
0:17:39	subspace gaussian mixture models this is
0:17:41	the reason might talk
0:17:42	the
0:17:44	linear transforms
0:17:45	things like fmllr M L R S T C
0:17:49	hlda
0:17:50	things of that nature
0:17:52	vtln is in here to kind of the
0:17:54	linear form of vtln
0:17:57	uh
0:17:57	all of these things that he had these are kind of you know directories that contain
0:18:01	command line programs that tells you a bit about the structure of the toolkit which is that we have
0:18:06	which really more than a hundred command line program
0:18:09	and each one does a fairly specific thing
0:18:12	wanted to avoid this phenomenon where you have a program that kind of allegedly does one thing
0:18:17	that really is controlled by button really an option
0:18:20	and has rather complicated behavior depending which upset you give it
0:18:24	so this is part of the mechanism that we use to ensure
0:18:27	uh
0:18:28	the everything's configurable an easy to understand
0:18:31	is no python layer but that's a lot of uh
0:18:34	programs as simple function
0:18:37	and on top of this
0:18:38	is the
0:18:39	shell scripts
0:18:40	so to do a not actual system building a recipe
0:18:45	what are example scripts currently only do is it's the bash script
0:18:48	and that you know has a bunch of variables and bash to keep track of iteration and stuff
0:18:53	and it and it runs the job
0:18:55	but invoking
0:18:56	from the command line
0:18:57	because the different ways you could do this if you if you love perl up a python or whatever you
0:19:01	as to i
0:19:02	but that's how a script
0:19:04	and and something that
0:19:06	i really haven't included on this diagram but it's kind of parts of the
0:19:10	dependency structures this some
0:19:12	tools that we rely on so
0:19:15	uh
0:19:17	for language modeling
0:19:18	i D thought we use i R T L them just because of license issues but probably you on use
0:19:23	that's i lm if you
0:19:25	wanna do a lot of a language modeling stuff
0:19:27	uh things like as P H two pi
0:19:29	to and
0:19:30	to uh
0:19:31	in separate data from the L
0:19:34	and so on so that the you
0:19:35	we actually we actually have a
0:19:37	and of can
0:19:38	and installation script that will automatically obtain those things are so the scripts can run
0:19:43	without you having to manually install stuff
0:19:46	and your sis
0:19:48	so i'm just gonna
0:19:49	briefly summarise the matrix like tree under will be talking more about it later but the plan was
0:19:55	to allow people to escape after this initial segment
0:19:58	case the not that the boat to that they one here about this stuff
0:20:01	but uh
0:20:03	as i said it's a C plus plus rap for a blast and seal at pat
0:20:07	and we've
0:20:07	well why should say on really has gone to a lot of trouble to ensure that it can compile
0:20:12	and the various
0:20:13	different configurations the what
0:20:16	libraries you have your system
0:20:18	so it can either the work from blast plus C lap pack
0:20:21	or from a less or using
0:20:23	entails M K L
0:20:25	the reason is that on some systems you might have one but not the other
0:20:29	i i less is an implementation of blast that's the
0:20:32	kind of optimized to the specific a hardware
0:20:35	automatically
0:20:37	is is generally a more
0:20:39	so
0:20:40	the code that we've rat
0:20:41	includes
0:20:42	generic matrices like square matrices
0:20:46	also packed symmetric matrices where where you uh
0:20:50	have a symmetric matrix the only store the lower triangle
0:20:53	and it's like this this this
0:20:56	order
0:20:57	and uh pack triangular matrix
0:20:59	there are other formats that last and C web back supports but these are the ones that we for what
0:21:04	most
0:21:04	applicable to
0:21:06	speech processing like we don't are a lot of sparse make sure
0:21:10	and traditional
0:21:13	so
0:21:14	this uh and i like we also includes things like S P D an F S C
0:21:18	i fifty isn't supply any of those libraries but we uh
0:21:22	we we uh got permission from rick come out of our microsoft
0:21:25	to uh
0:21:26	use this code
0:21:27	so he has a good "'em"
0:21:30	um
0:21:32	something about the matrix like the even if you don't buy into the whole to kit
0:21:36	if you need a C plus plus matrix library it's probably a
0:21:40	is probably quite good in fact it's surprising that there it doesn't seem to be a lot out there
0:21:45	that fills this nice just there's blues
0:21:48	but that it's a rather weird library and i i don't think a lot of people like
0:21:55	um
0:21:57	okay if you what the about open F is key
0:22:00	so i i seem and one he knows what what fsts are
0:22:03	it in T had this command line tool kit
0:22:06	but i don't believe they ever released
0:22:07	source
0:22:08	so one some of those guys went to google they decided to have one that was uh
0:22:12	for open source and it's a patch lies
0:22:15	that's why we as part there is reason we made out the a you license
0:22:18	because we figured that
0:22:20	to to use up pin fst there's no real point in having a
0:22:23	different license "'cause" it just gives the law my head
0:22:26	so we went for the same one
0:22:28	ah
0:22:30	so yeah
0:22:31	we can pile against its some that for is the decoder
0:22:35	it doesn't use like a special decoding graph format
0:22:38	use is the uh same memory structures the openfst
0:22:43	and the by the way open F to C has a lot of templates and stuff so that
0:22:47	is not just one fst for and there's a lot of them
0:22:49	so if you want to do you could uh
0:22:52	kind of template your decoder run some fancy format that would be let's a compact or dynamically expanded or some
0:22:58	like
0:22:59	we're not gonna go into that in detail today
0:23:02	so we actually implemented various extensions to openfst
0:23:07	some of the recipes the perhaps not totally in the spirit of openfst because
0:23:12	those guys have a particular recipe that they do
0:23:15	and i was is just a little bit different for
0:23:20	later on i can explain why
0:23:21	i feel that there are good reasons for uh i don't know if those guys would agree with
0:23:26	uh
0:23:29	so
0:23:31	if you with the by about io
0:23:33	it's of the controversial decision among the group to U C plus plus three
0:23:38	in the end we decided to do it probably because openfst also does it
0:23:42	uh
0:23:43	something you know a lot of people prefer sea base i L
0:23:46	but but but we do this
0:23:48	uh
0:23:48	we support binary in text mode formats a little bit like htk so that each
0:23:53	object in the toolkit
0:23:55	as a function that will
0:23:57	right and it takes a little argument binary tech
0:24:00	so it it'll just
0:24:01	put its output it's data out of the stream in binary or text mode
0:24:05	any in each object also has the read function that does the same thing
0:24:08	so
0:24:09	ah
0:24:11	it's of the standard thing in many talk at the used and final made in various ways
0:24:15	like this can mean the standard input standard output
0:24:18	it is just a command
0:24:20	and this is what how it knows that it's
0:24:22	can
0:24:23	uh
0:24:24	this is the
0:24:25	and off that into a found meaning it will
0:24:28	it will open the file fc to that position
0:24:31	it's is uh useful for reasons that will be described later
0:24:36	uh
0:24:39	so this this archive format is it
0:24:41	quite fundamental part of the way uh
0:24:44	kaldi work
0:24:45	and i think
0:24:46	i've just cry i'm gonna describe this more later in a another talk with the basic concept is
0:24:51	you have a collection of objects let's imagine that they're matrix
0:24:55	and there you are there are indexed by a string
0:24:58	where the string might be let's say an utterance id
0:25:01	so you want to have some way to
0:25:04	to access this collection of uh
0:25:06	strings and matrices
0:25:09	and you might there might be a couple of different ways you could do that you might wanna go sequentially
0:25:12	through the
0:25:13	as an accumulation of some
0:25:15	we might want to do random access
0:25:17	so there's a whole framework for doing this
0:25:20	uh
0:25:21	basically the reason is so that your
0:25:23	the most of the calico doesn't have to worry about
0:25:27	things like opening files and ever conditions and
0:25:30	you know that doesn't have to be a lot of logic about that in the command line programs because it's
0:25:34	all handled by some
0:25:36	generic framework
0:25:37	but apart from this we tried to avoid
0:25:39	generic framework
0:25:42	ah
0:25:44	the tree building clustering code
0:25:46	we it's based on
0:25:47	very generic
0:25:49	clustering the can something like
0:25:51	i guess hard to model whatever they call it
0:25:53	uh
0:25:54	so it doesn't that that that internal code doesn't assume a lot about what your trees
0:25:59	it is suitable build decision trees in different ways including
0:26:02	like sharing the true
0:26:04	and asking questions about the central central phone
0:26:07	it's like that
0:26:08	um
0:26:10	it's very scalable to white context for example quint phone
0:26:13	i know a lot of the
0:26:16	it it's hard to write code that was scaled to queen phone because if you have to enumerate all of
0:26:20	the context
0:26:22	that's kind of it's hard hard to go to
0:26:24	a but uh
0:26:25	we basically avoid ever enumerating those con
0:26:29	uh as an example of a
0:26:30	how we make use of this general C
0:26:33	and the wall street journal recipe we uh
0:26:35	we increase the phone sets of the in the were asking about the phone position and the stress
0:26:41	i
0:26:41	a "'cause" the know he's to K supports this "'cause" i thing you had a
0:26:44	have a paper marked with
0:26:45	he about doing that
0:26:47	so uh
0:26:48	but but uh if the phones that much larger than that probably
0:26:52	an approach based on enumeration of context would start
0:26:55	i
0:26:57	you don't think so no i mean like it was a thousand thousand keep this day
0:27:03	right
0:27:04	okay well i
0:27:06	okay
0:27:09	uh
0:27:10	okay hmm and transition modeling co
0:27:13	so
0:27:14	we've
0:27:15	we try to have an approach where
0:27:17	a piece of code only needs to know
0:27:20	the minima needs to know
0:27:21	so so the hey gmm and transition modeling code doesn't really have any notion of a pdf it's purely
0:27:27	it purely does what it needs to do
0:27:30	and the rest to separate
0:27:32	so
0:27:32	this is probably pretty standard approach you you develop a uh
0:27:36	you specify prototype to paul
0:27:38	it's apology for each phone is that how many states what the transitions are
0:27:43	uh
0:27:44	and we make the transitions the
0:27:46	separate depending on the uh
0:27:49	depending on the pdf
0:27:50	so so that if the pdfs into states are different than the transitions out of those
0:27:54	states are separately estimated
0:27:56	is this is just the most
0:27:58	specifically that you can estimate the transitions without having your
0:28:02	decoding graph blowup
0:28:04	it's not believing clear that this matters but
0:28:07	uh
0:28:08	we just felt that it was that we should do the best we could on
0:28:12	uh
0:28:13	they're mechanisms would sending these youth hmms into fsts because
0:28:17	all of the training decoding is fst basically kind of have to have an fst representation of these
0:28:24	uh
0:28:25	it's is something that we touched on a a
0:28:28	and i are F S T so what you would normally imagine is that the F it has input symbols
0:28:32	that are the
0:28:33	the pdf so some symbol the represents the P D and the output symbols of the word
0:28:39	but the problem with that is let's suppose you uh
0:28:43	you want to find out what the phone sequence
0:28:45	it's all well well and good if your
0:28:48	if if your phone had separate tree
0:28:51	so that so that it was could for each state which phone it belong
0:28:55	but but what if you had a larger phone set and you wanted to have a shared tree room
0:28:59	and that wasn't you know one to one mappings
0:29:01	oh there was in the mapping you need so
0:29:04	so we have a input labels on the fsts the encoded bit more information
0:29:08	uh
0:29:10	and this is also useful in training the transitions because
0:29:12	sometimes just the pdf labels wouldn't you of you quite enough information
0:29:17	the train the transition
0:29:20	uh
0:29:21	there's a couple of different ways to create decoding graphs
0:29:24	for
0:29:24	for uh training purposes you have to create a lot of these things at the same time
0:29:29	and combining the fst algorithms using script
0:29:32	would be quite inefficient because you have the overhead of process creation
0:29:37	so
0:29:37	we uh
0:29:39	we call the openfst algorithms of the C plus plus level combine them together
0:29:43	so that uh
0:29:45	you can create your decoding graphs for
0:29:47	training
0:29:49	uh
0:29:50	and and we typically put them in one of these archive
0:29:54	like basically a big file concatenated together with little keys in it
0:29:58	on disk
0:29:59	so that you don't have the I O of
0:30:01	accessing hundreds of little file
0:30:03	training use of the viterbi path
0:30:05	these graphs
0:30:07	uh
0:30:08	for test time
0:30:09	we we we didn't we didn't use this approach of C plus plus because it there's just no point
0:30:14	we uh
0:30:16	it's basically scripts and i'm gonna goes wannabe scripts later for those words
0:30:21	um
0:30:23	that's the least scripts that create the decoding graph recalls some openfst tools but some of our own
0:30:28	and that relates partly to a difference in recipes
0:30:31	but uh
0:30:32	i'll talk more about later
0:30:34	after great
0:30:36	so
0:30:37	and i was gonna talk later about some of the acoustic modeling co
0:30:41	i'm just gonna give a brief summary
0:30:43	are gmm code is
0:30:45	it's very simple it's not part of some big framework
0:30:47	it kind of but like an
0:30:49	and object that has you know the means the variances
0:30:52	it can evaluate like it's for you give it the feature
0:30:55	but it doesn't
0:30:56	and her from some
0:30:58	generic acoustic model class and it doesn't at ten
0:31:01	that's a kind of know about things like linear a it just sits there
0:31:04	and and things like we a transform
0:31:07	they kind of have to access the model and do what they want
0:31:10	the the reason for that is that if
0:31:12	the gmm knows too much
0:31:14	them whatever you do that's fancy
0:31:16	you have to then change the gmm code
0:31:19	and it just
0:31:21	it's is not my situation
0:31:24	so uh
0:31:26	yeah we have a separate class for gmm stats accumulation
0:31:29	and doing that they
0:31:31	so
0:31:32	for for a collection of gmms like an gmm gmm system
0:31:36	we have a class that pretty much behave similar to a vector a G M at
0:31:41	so it's
0:31:42	it's a fairly simple thing
0:31:43	there's no notion of name of a state that is just an integer
0:31:47	and then really we've avoided having
0:31:50	like names and names for things in the co
0:31:52	exit
0:31:53	jurors
0:31:54	uh_huh
0:31:57	oh this this low case vector just refer to the S T L vector
0:32:01	but there is an upper case vector to that
0:32:03	but does something in a matrix like
0:32:06	i
0:32:08	well the code is never been case in that as the code we
0:32:12	i i even on windows
0:32:14	uh
0:32:15	i
0:32:19	yeah
0:32:20	okay
0:32:21	we've got quite a lot of linear transform coder
0:32:24	uh
0:32:26	lda hate lda
0:32:28	again and fitting on the fence with regard the naming of this technique
0:32:32	i don't wanna and anyway
0:32:33	i
0:32:35	uh
0:32:36	another these multi name okay
0:32:38	uh olympia version of each other i mean we tried regular vtln is
0:32:42	yeah everyone knows that it's kind of tricky to get it to work
0:32:45	it was that you'll anyone that worked better in the N
0:32:47	uh it is something new that
0:32:50	it's a kind of a replacement for vtln that what's a little bit better
0:32:53	i gonna
0:32:54	explain what it is uh at a later date
0:32:57	mllr
0:32:58	uh
0:32:59	a lot of this
0:32:59	so
0:33:02	one this transform the global the with the way we handle them as well
0:33:06	it just becomes part of the feature space
0:33:09	so it's just
0:33:09	start of the matrix on disk and this
0:33:12	use a lot of plight so the way it actually works is that this matrix
0:33:15	is multiplied by the feature as part of a high
0:33:18	my seem like you're right obviously there is silly way to do it from a computational point of view but
0:33:23	it just makes the scripts really convenient
0:33:25	to uh
0:33:26	do
0:33:27	uh
0:33:29	yeah so when i say they're applied in a unified way what what i mean is that the co the
0:33:33	estimates any of these transforms
0:33:35	there really outputs just to make trick
0:33:37	so uh
0:33:38	you know there's no like
0:33:40	and some a lot transform J
0:33:43	that's just
0:33:44	well okay yeah there is so for the uh regression tree one
0:33:47	i
0:33:48	but for but for the global one it's just it's just a matrix
0:33:52	i
0:33:52	i mean that's with the point of contention among as that to whether to do it this way
0:33:56	but uh
0:33:57	some of a style that it was important to keep the simple case is simple
0:34:01	and to it to avoid having a
0:34:03	a framework
0:34:04	for the cases one was an S
0:34:07	uh
0:34:10	okay decoders
0:34:11	well of the decoders that we currently have use
0:34:14	fully expanded F S is one i mean when i say for the expanded i mean is down to that
0:34:18	H M and state level with
0:34:20	so loops represented as uh
0:34:23	actual you know if sdr
0:34:26	i know there's a lot of way to do this and initially
0:34:28	one of the thoughts we had
0:34:29	would be that
0:34:31	you know we wouldn't have the self loop so we might even have
0:34:34	representations of the states the and then it was just so much simpler to do it this way
0:34:38	this is what we have now
0:34:41	we have three decoders but by decoder we mean they uh
0:34:44	C plus plus code that does decoding
0:34:47	it's not necessarily the same thing as a command line decoding
0:34:50	we have three decoders on the spectrum simple too fast
0:34:53	and the reason for this is that
0:34:54	once you have a complicated fast decoder is almost impossible to the to debug
0:34:59	so if something goes wrong you can always just one the simple one
0:35:02	you know and you can find out if it's a decoder issue
0:35:06	uh
0:35:07	decoded
0:35:08	we wanted to make it so the decoder doesn't as you too much about what you're model model selection
0:35:13	so it again decoder has no idea of gmm hmms it doesn't even know about features
0:35:18	all that
0:35:20	all the decoder knows about is give me the likelihood or
0:35:24	score level
0:35:25	for this
0:35:26	uh frame index
0:35:28	and this pdf in that
0:35:30	so it so interface that the decoder seizes is almost like a matrix
0:35:35	the matrix of uh
0:35:37	of floats but i'm is is not represented that way because you want to
0:35:41	you know you want to have it on them on
0:35:43	i
0:35:45	so yeah this is the decodable interface an interface that the
0:35:49	it's a very simple interface that says give me the likelihood for this you know time in this frame and
0:35:54	like
0:35:54	how many time frames are the
0:35:57	and how many pdf index is that that's almost all the interfaces
0:36:01	but this this is the interface at the decoder requires so the idea was to implement you know
0:36:06	L fantastic a model
0:36:08	and you
0:36:09	uh
0:36:10	i
0:36:11	in in a very matter what interface of that model is
0:36:13	you create a small object that satisfies the decodable interface
0:36:17	and knows how to get the likelihoods from your and L fantastical model
0:36:21	and then you uh
0:36:23	you instantiate the decoder with that are you give that
0:36:27	so uh
0:36:31	the gmm wrapping okay
0:36:34	yeah so i come online decoding programs a very simple we don't have like multipath or anything
0:36:39	we don't have uh
0:36:42	we don't we don't know than to support multiple types of model
0:36:46	an example decoding program is
0:36:48	decode with the G M and
0:36:51	but no
0:36:52	with number multiple class adaptation
0:36:54	yeah so does the simple thing
0:36:55	and then if you want to support let's a multi-class
0:36:58	mllr fmllr
0:37:00	we uh have a separate come online prague
0:37:03	yeah the idea is that
0:37:04	there might be people coming into the project might want to be able to understand that come online program
0:37:09	and we don't one that once a make the barrier to entry too high
0:37:12	we got the
0:37:13	support the overhead of having to maintain two parallel decoders
0:37:17	keep it relatively simple to understand any given one
0:37:24	uh
0:37:24	we support the standard types of features
0:37:27	mfcc and plp features are quite similar to
0:37:30	K one
0:37:31	we've
0:37:32	we put in a reasonable range of configure ability but
0:37:35	i mean being realistic with respect to how much people are really working on this stuff i mean i think
0:37:40	most people are doing research on this would probably be coming out with their own features
0:37:44	so we don't support every possible
0:37:46	combination of it
0:37:47	for every possible change
0:37:49	we only we we dwell format because there i reasoning is
0:37:53	your you can always it's find the external program to convert it and
0:37:57	do it as part of a high
0:38:02	sorry
0:38:07	well we cannot htk and i won't from uh we don't there's no more that we support
0:38:13	uh
0:38:15	yeah
0:38:15	i mean
0:38:16	i i basic concept to have people use the system is
0:38:19	as a complete system
0:38:21	because once you start supporting model you know in a conversion just get work
0:38:26	but yeah that's the he's tk K features as a as a special case
0:38:30	uh
0:38:31	we typically will right features another large objects to a single very large file of relates to this archive format
0:38:37	so the form of the file as a key space then your object
0:38:41	and another key a space that object
0:38:43	and uh
0:38:45	we have efficient mechanisms to read such files
0:38:48	the the the two normal cases are firstly sequential access
0:38:51	we want it's rate over the things an archive
0:38:54	exactly random access and the the different ways to do that one is
0:38:58	you can write a separate file that has little
0:39:00	point doesn't of the file
0:39:02	another is that
0:39:03	you can kind of simulate random access even though you're really going sequentially
0:39:07	if you know that the keys are sorted
0:39:10	uh and another way is if the file isn't isn't that big
0:39:13	you can do random access by just having the code go through the whole file
0:39:18	stalled objects and memory
0:39:19	that's not just scalable but
0:39:21	for for a lot of uh
0:39:23	types of all kinds it really doesn't matter
0:39:27	oh yeah so the feature
0:39:29	feature level processing like adding deltas that from a lot
0:39:32	typically each one of those the separate program so you have like a sequence of programs and apply
0:39:37	and again that's a bit inefficient but
0:39:39	it's not like it's really consuming more than ten percent of your C P U so
0:39:43	you just don't care that much this has been written with
0:39:46	ease of use in my
0:39:49	uh
0:39:50	like i said there's a lot of command line tools this is an example of uh
0:39:54	a command line and this backslashes of this
0:39:57	the cell
0:39:58	so uh
0:40:01	this this is one of the many programs
0:40:03	the plp would be a separate command line
0:40:06	this is just you know
0:40:07	an option
0:40:08	either the two command line arguments in this uh
0:40:11	i gonna be explaining later on or about what these mean with this
0:40:14	directed to write these things to it
0:40:16	and archive on the
0:40:18	a key object key object
0:40:21	and then
0:40:22	i don't know this is the input
0:40:23	we have to read it
0:40:25	and then this is telling it to write an archive and also
0:40:28	and i C P file that
0:40:30	kind of has little pointers into the okay
0:40:32	so that you can efficiently access the features by random access
0:40:36	um um
0:40:39	so
0:40:39	so yeah another example of another feature of this is that as only one option here we we we have
0:40:44	no more than a few options on any given come on
0:40:47	i mean it's a local program i support
0:40:49	less the channel
0:40:51	it's not it's not a very can different to at is more driven by how you combine these grow
0:40:57	a
0:40:58	oh you something else about this whole archive a uh formalism is that
0:41:02	this C plus plus level code in the individual come line tools
0:41:06	we doesn't have have to worry too much about high uh
0:41:10	you can just treat
0:41:11	the uh
0:41:13	when to get something like this there's
0:41:15	there's very short uh
0:41:17	statements in the C plus plus that will it's a rate over a
0:41:20	stuff
0:41:20	so it doesn't have the
0:41:22	think too much about the error conditions
0:41:26	but yep
0:41:32	fst festive generation
0:41:35	okay that as another part of the talk later on
0:41:44	well
0:41:45	for training
0:41:47	there's there's a command line program that will
0:41:49	kind of do the fst generation for you and generate lots of the left S to use one for each
0:41:53	file
0:41:54	yeah so for testing
0:41:56	it's it's a script the calls a fist openfst programs an our versions of openfst for
0:42:03	so
0:42:05	i'm gonna go through that script later one and another part
0:42:07	top
0:42:09	a a are you this decide this is not obvious you know a lot stand the script
0:42:13	but this is just to get people some idea
0:42:16	oh of uh
0:42:17	of how we do do training
0:42:19	so you know this is the bashed script it's doing a loop over the iterations
0:42:24	uh and this one is estimating ml mllt up
0:42:29	i suppose this script review the bias and sorry man i
0:42:33	but as that we are is the colour i've yet
0:42:35	so a
0:42:38	so if it's that one of iterations that we do a lot C
0:42:42	then uh
0:42:44	so we have on disk
0:42:45	some uh alignment this is like steak level alignment
0:42:49	it's in a mark at i've
0:42:51	from my that i mentioned
0:42:52	so this converts them to posteriors
0:42:54	just an average of trivial way by thing that each
0:42:57	each one has a posterior of one
0:43:00	this takes the this
0:43:01	this gives a zero weight to the file and
0:43:03	that's would be a
0:43:04	this would be a variable and by
0:43:07	uh
0:43:08	yeah so this takes away the uh
0:43:10	you the silence is there a posterior
0:43:12	and this is an accumulation program
0:43:14	that uh
0:43:16	this would be the model that's the thought fit of the features as the abashed variable that would be
0:43:21	that elsewhere where
0:43:23	uh this
0:43:24	a a hmmm
0:43:25	i think that's refers to the standard input
0:43:28	me that's reading an our cat from the standard input and that
0:43:30	output by the
0:43:32	you this
0:43:32	mean that's writing an archive to standard it out but
0:43:35	so
0:43:35	yeah yeah output of these programs is passed by up pi
0:43:41	uh
0:43:42	all all of the error and logging out but goes to the standard error uh
0:43:45	because we've kind of used with that it out but for this type stuff
0:43:49	so
0:43:50	so we just directing the logging up
0:43:52	there
0:43:53	so then this is a separate program that does the mllt the estimation
0:43:58	it takes in uh
0:43:59	let me see
0:44:01	uh it's it's it's computing some kind of make
0:44:04	and then uh
0:44:06	because then am a lot T yeah
0:44:08	what i i have to you can the transform
0:44:10	you have to change the means of your model so
0:44:13	we have a separate we like to get everything separate
0:44:16	so you know transforming the me the separate operations so we have a separate program for that
0:44:21	and then
0:44:22	we have to compose the L B M T transform with the previous one
0:44:26	so this is another will program that does that
0:44:29	so this with was setting another bash variable able to make
0:44:32	the ah features correspond now to the
0:44:35	new ml L you a melody features
0:44:38	so
0:44:40	so as you can see that this is the very and bash
0:44:43	and it's
0:44:43	this would be passed as a command line arguments to one of the program
0:44:47	and it's a command involving a pie that actually vol
0:44:51	calling to separate cal be uh
0:44:54	program
0:44:55	each for their own argument
0:44:57	so obvious you can guess from the names of those programs what they're doing
0:45:01	and then of "'cause" uh it seems to have features sub
0:45:04	oh yeah i think we were estimating the ml T on a subset of features
0:45:08	so this is like the same as this but it's the
0:45:10	it's using less
0:45:12	the data
0:45:15	so i think i
0:45:17	i spoke about these issues but for
0:45:21	oh yeah so uh
0:45:24	we had example scripts results management and was to general and these run from the ldc
0:45:29	distributed this
0:45:31	uh now we found in the literature just some uh
0:45:35	some some uh baseline
0:45:37	these numbers are numbers are just the basic context system
0:45:42	with i think uh mean normalization
0:45:44	we have of course more advanced things but
0:45:47	those you know because it had to find in the literature the same thing
0:45:50	we just giving you the unadapted adapted
0:45:53	so it's a
0:45:54	slightly better than this number will can right someone a two thousand
0:45:58	and that the hates you K paper from ninety four
0:46:01	a has a funny but a number for this was the gender dependent system
0:46:05	so uh
0:46:05	so i think basically would doing the same as
0:46:08	you expect given the same out
0:46:11	i mean
0:46:12	uh
0:46:13	i was hoping you know the set of this help project that the results would be but uh
0:46:17	for issues relating to the tree in can phone and stuff but
0:46:20	you know that in we give a senate
0:46:22	so
0:46:22	it it's working there's no major but
0:46:26	uh did of the
0:46:28	okay next slide
0:46:32	uh just the not on speed and coding is used
0:46:35	use a bigram numbers and the "'cause" the baseline we'll bigram numbers
0:46:38	we can't yeah yeah code with the full
0:46:41	with the full uh trigram language model that
0:46:43	distributed with the wall street journal corpus
0:46:46	because the fsts uh
0:46:48	they get to large
0:46:50	we have a "'cause" to with pruned track
0:46:52	but that's why we're coding the bigram numbers
0:46:54	so
0:46:56	hopefully by the sum we gonna
0:46:58	as the couple of things we can do that we both working on one is to have a decoder that
0:47:01	does some kind of on the fly
0:47:03	pensions so that we can uh
0:47:05	the code directly with that
0:47:07	and the other to have a just generation so we can we score
0:47:11	the decoding speed is for these was to just don't numbers is about twice as fast as real
0:47:16	and a "'cause" that's on a good machine
0:47:18	so i mean this is june so that you don't get more than zero point one degradation from
0:47:23	versus a white B
0:47:26	a the wall street journal script takes a few hours on a single machine using
0:47:30	we problem lies on to three C be used
0:47:32	this is just an example script we didn't want to include things like you serve in the example script
0:47:37	because then it wouldn't run on uh everyone's machine
0:47:40	the was they would be fast if you were doing a parallel
0:47:44	yeah
0:47:47	uh_huh
0:47:54	if it in member it well as well
0:47:56	well
0:47:58	but ten gig
0:47:59	i i mean
0:48:00	i S you know everyone knows that F is T compilation tend to up a bit
0:48:04	it's not like
0:48:06	if you have the size of the model you can just about compiler
0:48:12	i i don't recall that it's a trigram one for most journal
0:48:15	i i
0:48:17	and then we go how many was but i think
0:48:19	i don't think that the our stuff is any worse than you know normal if T
0:48:23	that ups that fully expand of thing
0:48:27	oh yeah okay results management
0:48:29	this is a
0:48:31	use she came results
0:48:33	or take an uh
0:48:35	from uh
0:48:36	this is i think this is basically the hey K are each K the be but he's real us to
0:48:40	take it from a paper of mine like in ninety nine or something
0:48:43	"'cause"
0:48:44	i just couldn't find in the read me file from are C K on all of the test
0:48:49	and the average as you can see the average is the same
0:48:52	so
0:48:53	with the same algorithms are getting the same result as H
0:48:55	okay
0:48:56	uh
0:48:59	yeah and it and the decoding we on the setup is about zero point one times real
0:49:09	yeah
0:49:09	yeah
0:49:16	yeah the test set are quite
0:49:20	oh yeah
0:49:21	is a very small test that a handful of words that are of
0:49:25	uh this is this page is mainly
0:49:27	just to give you some idea of the kinds of things that are in our example scripts we have a
0:49:31	bunch of
0:49:32	different configuration of this of the standard configuration
0:49:35	well this is the standard configuration because this is what within the htk baseline line
0:49:40	uh
0:49:41	um
0:49:42	adding M L T doesn't we seem the hell
0:49:45	sorry adding is T
0:49:46	see
0:49:47	as they we the hell
0:49:48	i
0:49:51	a a a a a it's well i think nine frames plus lda that you makes it worse but then
0:49:56	when you do
0:49:57	uh F T C on top of that
0:50:00	if that you better than here and so that this was the kind of this was that I B M
0:50:04	recipe P
0:50:05	so
0:50:07	sorry this with I B M is to be so i i guess that must been some interaction between these
0:50:11	two parts of the recipe
0:50:13	that somehow made it work
0:50:14	i i don't know if it's a generalized to other trade other test set
0:50:17	we gonna find out
0:50:19	uh that's placed nine frames plus hlda
0:50:22	triple deltas plus hlda
0:50:24	triple deltas plus
0:50:26	lda D A plus a lot C
0:50:28	this this
0:50:29	quite good
0:50:30	uh sgmm cyst these are all and adaptive
0:50:33	have a separate slide for uh adapted exp
0:50:37	um
0:50:39	if is doing it
0:50:41	and that's
0:50:41	it's stated otherwise that oh yeah okay so this is but utterance adaptation this is per speaker
0:50:47	so
0:50:48	this was four point five my and before uh
0:50:51	adaptation so it really doesn't help if you do it but i'd sir rights and that's because this too many
0:50:55	parameters
0:50:57	in to model a
0:50:58	this is doing the same thing per speaker gets a lot but uh
0:51:01	its exponential transform is
0:51:03	again i'm not gonna describe what it is is something vtln one
0:51:07	uh and it gets quite a bit but uh
0:51:09	this is a this vtln and of the kind of many a version of vtln i believe
0:51:14	it is that thing to improve quite a lot
0:51:16	and of got improvement is more pronounced on the per utterance level because
0:51:20	uh
0:51:21	in know it it's just like a constrained form of a from a loss of the only point is
0:51:26	to do it
0:51:27	to do when you have less they
0:51:29	uh
0:51:30	splice nine frames for cell day sex to transform
0:51:34	a from well thing i from a lot
0:51:36	we only did some of these put speaker because it wouldn't help of the
0:51:39	uh
0:51:41	as you can see that the well of different combinations this is as gmm including the
0:51:45	speaker offsets sets the and thumbs if you member
0:51:48	and it does help so
0:51:50	so uh i think rick was saying that that is wasn't working for him but it seems to be working
0:51:54	for us
0:51:55	three point one five goes to uh
0:51:59	where is it to point six eight
0:52:01	i i must of uh forgot to fill this line and
0:52:04	it's is that's gmm plus a from a la
0:52:07	but no speaker vectors
0:52:09	yeah
0:52:10	a per speaker
0:52:12	yeah
0:52:13	i think i have these numbers but i think i must not put in i think a best number was
0:52:17	like to point for
0:52:20	point three
0:52:21	uh
0:52:23	so general plug for cal
0:52:26	uh
0:52:27	i believe it easy to use i mean i have the scripts didn't scale you guys up as if you
0:52:31	traction is that once you understand them
0:52:34	everything becomes quite simple
0:52:36	but
0:52:37	it kind of does that you that the sound has speech works like if you some under who does
0:52:42	is randomly
0:52:43	moving the script you know changing configurations of
0:52:45	the
0:52:46	you're not uh
0:52:47	it's not gonna work
0:52:48	it it doesn't like
0:52:50	it doesn't or to magically know that the features you have a not combat your model
0:52:56	so so you can have to know what you doing from a speech science point of view
0:53:00	but
0:53:01	it's quite uh
0:53:02	it's easy to use that the C plus plus
0:53:04	flash to
0:53:06	software engineer
0:53:08	uh
0:53:08	it's easy to extend and modify
0:53:10	you can reduce should be go changes are give them back to
0:53:13	the cal group
0:53:15	uh
0:53:15	we open to including other people's
0:53:17	stuff
0:53:18	so that may give you most citation
0:53:21	so this
0:53:21	is i really
0:53:23	the and the this first part so
0:53:26	you can get up and have a drink and after a few minutes
0:53:29	well
0:53:32	yeah has documentation cal D duck source forge dot net
0:53:36	uh uh okay if is not as good as H K and probably being realistic will never be
0:53:41	what will do
0:53:42	is will
0:53:43	of able lies the F but the he's to K has use and point people to the he's K documentation
0:53:48	so then about eight and say that have he had then
0:53:51	yeah i know me
0:53:52	i
0:53:54	i
0:53:55	i
0:53:55	i use C
0:53:58	see i
0:53:59	i
0:54:02	but okay
0:54:03	we can have a shot rate you we can have a drink
0:54:06	and just a pair you're not in uh
0:54:09	that committed to it
0:54:10	and then
0:54:12	uh uh we've have a gonna talk up to what
0:54:14	the fact

A TOOLKIT FOR SPEECH RECOGNITION RESEARCH

Kaldi Workshop

Presented by: Dan Povey, Author(s): Dan Povey