Přepis řeči - RUNNING THE EXAMPLE SCRIPTS

0:00:13	i
0:00:14	i think this is gonna be the official theme tune L T
0:00:19	a
0:00:21	okay
0:00:21	so i gonna be talking about yeah
0:00:24	how to rank of the
0:00:26	and
0:00:27	so basically will be going
0:00:29	we have to recipes on on uh
0:00:31	yeah
0:00:32	the come our stuff one is results man as one a most wall street journal
0:00:36	we gonna go through some all the results management recipe
0:00:39	and well
0:00:40	well have a few digressions to explain
0:00:43	as much of the internal the cal as you need to know it's a kind of a understand that
0:00:48	that
0:00:48	we those to the installation process i'll describe the unix one because that's one most people will use but
0:00:54	it also has a visual studio windows one
0:00:56	uh
0:00:57	the scripts i I scripts or all and bash again for popularity reasons
0:01:02	D is kind of agnostic about the shell there's nothing unit that's really specific to anyone shell
0:01:07	no
0:01:08	suppose you want to download cal than you want to run it
0:01:11	a
0:01:12	you probably first go to this
0:01:14	location kaldi don't source forge don't net
0:01:17	i F will also work
0:01:18	we have a page of documentation that explains the
0:01:22	we much everything county relate to
0:01:24	uh
0:01:25	we use a source control program called sub ocean the command name is S the N
0:01:30	it'll typically be installed on most of the system you will have
0:01:35	uh
0:01:36	it's a S U N as a little bit like C V S but it's a more modern implementation
0:01:41	so to check out cal the you would just type this command on
0:01:44	and this will a
0:01:46	the a lot of stuff go by and the screen it will check out a bunch of directories the code
0:01:50	the screen
0:01:52	an installation instructions the documentation source
0:01:55	so on
0:01:56	uh
0:01:57	the installation instructions you just
0:01:59	look at the install file
0:02:00	and the installation is pretty simple that's like change direct to here
0:02:04	ron installed a S say
0:02:06	C D to here running
0:02:08	one can figure run me
0:02:10	there is an rather nonzero probability that something will go wrong
0:02:14	because it does kind of hope that set things there install
0:02:17	but we kind of provide instructions
0:02:19	a the common cases
0:02:21	and if it doesn't and stop please ask me and i'll tried the
0:02:24	help you to get it to install
0:02:26	uh
0:02:28	that there is a directory of kind of external tools and
0:02:31	we try to have a script to configure and uh
0:02:34	to download and make
0:02:36	all of these external tools so that you don't have to
0:02:39	worry about that yourself
0:02:40	these include as P H two right
0:02:43	two
0:02:43	because sphere files
0:02:45	i yeah rest M is a language modeling toolkit
0:02:49	uh
0:02:49	with a that them as i mentioned before we chose this because it has a right of be open license
0:02:54	very
0:02:55	limited features
0:02:56	openfst at such
0:02:59	so
0:03:00	but you done that you checked it doubt you try to install it
0:03:04	uh
0:03:05	no scratch that sense
0:03:07	uh
0:03:08	so
0:03:10	a and the future we gonna have version numbers and everything currently because we haven't yet come to version one
0:03:15	point are we just have trunk
0:03:17	which is the kind of version control thing for whatever your current code is
0:03:21	uh inside then you'll find
0:03:24	rules which is a place where we gonna download and compile various external to
0:03:29	the find has sars the source which is
0:03:31	well all of us source code is including the source for our documentation
0:03:36	and these are the subdirectories in there
0:03:38	that these are these are the names of the things that i showed you on that funny uh slide with
0:03:42	the rectangles
0:03:44	so this this is all the subdirectories of code
0:03:46	and E G the directory a script
0:03:49	they contain the results management and
0:03:52	and wall street journal scripts
0:03:53	the you can probably see it's it's and of hubris and the naming scheme here we we got we went
0:03:59	for their deep naming scheme because we
0:04:01	believe that eventually will be tons of script
0:04:04	and those directories
0:04:06	i
0:04:08	so
0:04:10	i was you've uh
0:04:12	then the installation in tools and that that's we just wanna get script
0:04:17	you got the source to
0:04:18	configure
0:04:20	the the configure script that sometimes configure script so these vast
0:04:24	scripts that also generated by things like to make
0:04:27	or whatever it is
0:04:29	but this one is just a hand generated wanted tries to find where you're
0:04:32	like atlas library or steal a pack libraries and if it finds it
0:04:36	then a composite with that
0:04:38	and it and it detects certain like
0:04:40	certain systems like
0:04:42	cygwin and
0:04:43	mac os that have
0:04:45	particular setups that are common and then it handles those a separate
0:04:50	uh
0:04:51	uh it's good to talk minus J for when you make
0:04:54	decode because there's a lot of uh
0:04:56	tools and the code is rather template and so the compilation is a little bit slow
0:05:00	this makes it in parallel
0:05:03	you don't make test the ghost all the subdirectories and
0:05:06	runs all the programs that and with dashed test
0:05:09	we have a lot of testing programs
0:05:11	they're mostly uh
0:05:12	units S
0:05:13	to make sure that
0:05:15	all of the code is working things like
0:05:17	you have a matrix
0:05:18	multiplication or something you do the multiplication and you
0:05:21	verified that the answer was right
0:05:23	like that
0:05:25	uh
0:05:27	and there's also you if you can also type make well grind it runs a program called well grind to
0:05:31	check for memory error
0:05:33	and that that would
0:05:34	i mean right now there's no error but that would detect
0:05:37	if there with things like
0:05:38	and allocated memory
0:05:41	so suppose you've done and you to make you type make test and
0:05:45	i thing one wrong
0:05:47	so
0:05:48	you C D two
0:05:49	E jeez are S one and it's is where example script uh
0:05:53	this just seems that you
0:05:56	you know you member of the L D C a what have and you have
0:05:58	you have access to be
0:06:00	you did think this corpus i think for members that's like three hundred dollars the a lot of sites will
0:06:05	have it already
0:06:06	so the results management corpus it the a all really simple corpus
0:06:10	but
0:06:11	uh
0:06:13	we use a because it's really fast to run and it
0:06:15	because it's kind of an lvcsr like task is really medium vocabulary but
0:06:19	because it contains that words
0:06:21	and has a lexicon and everything it kind of but haze like a typical lvcsr system even though it's
0:06:26	or
0:06:27	uh
0:06:29	so that's be on some directory and you have to figure out what that directory is
0:06:34	at some point you have to pass it to one of the scrip
0:06:37	as a bunch of come than here that you're supposed to write you know you're not real expect the run
0:06:41	this directly it will just X it on you if you do that
0:06:44	it's
0:06:45	it's just a sequence of commands you're expected to run by had
0:06:48	because there's a high enough probability that any given one of the most failed that
0:06:53	you thought it wasn't good to a be over optimistic can just make it a single script
0:06:57	i mean the failure is the gonna be do to simple things like
0:07:00	maybe the wrong directory as some
0:07:02	but
0:07:03	anyway
0:07:03	so i i'm gonna go through what this run done S age that
0:07:08	the first thing is data preparation
0:07:10	and
0:07:11	so you will
0:07:13	you'll see that the door called data probably B to there
0:07:16	you know you know what the directory if your results management uh
0:07:20	data data is
0:07:21	this is up this is the ldc the
0:07:24	give it that uh
0:07:25	and it'll just do a bunch of stuff basic with convert thing whatever format this
0:07:29	corpus has
0:07:31	in to a format that deal like
0:07:34	and you know
0:07:35	Q waiting lists of file names like that
0:07:38	these
0:07:39	you C D out
0:07:40	uh
0:07:41	just for things that are created by this
0:07:44	that was the bunch of stuff actually in this directory that it
0:07:47	create
0:07:48	here's an example
0:07:49	S C P file
0:07:51	so it contains the utterance i'd B
0:07:54	and then
0:07:55	if is the pipe come on
0:07:58	from his the here's apply can man so this can is gonna be run whenever some program tries to read
0:08:02	this
0:08:03	now and C P far this of the concept that
0:08:06	the that's okay had is not really quite the same as C K's notion of an S C P file
0:08:11	not be explaining later exactly what that is
0:08:13	another think that's created here is
0:08:16	a decoding graph in fst format
0:08:18	in in some other scripts like in the wall street journal script
0:08:22	this stage wouldn't be creating any fsts it would just create an arpa
0:08:25	the because our M doesn't use an arc a we do like this
0:08:30	so uh
0:08:30	uh
0:08:32	and stage of data preparation uh
0:08:34	oh yeah it's is created in that directory to it comes from
0:08:38	stuff that's in the results management this
0:08:40	it'll create a lexicon for you in this form
0:08:43	is pretty of obvious
0:08:45	and will to ten into an F C
0:08:47	the call tools don't we deal with this directly
0:08:50	we deal with fst so the lexicon that you give to count
0:08:53	is gonna be an open
0:08:54	i
0:08:55	format
0:08:56	fast
0:08:57	that there also some uh is the speaker matt
0:09:00	so
0:09:01	this of that are inside the this a speaker I D
0:09:04	the file that contains a lot of the
0:09:06	this this is how to the
0:09:08	you know maps utterances just because and vice versa
0:09:11	there's no notion of like
0:09:12	masks of comments or
0:09:15	uh_huh
0:09:17	so
0:09:17	yeah that's content about turns idea is quite important important and D
0:09:21	in never there's no notion of like parsing file in thing like the last element is the utterance idea what's
0:09:27	of
0:09:27	you have to have an explicit uh list
0:09:30	and all of these that C P files than R kaiser index by this utterance are inside the you have
0:09:34	to decide on
0:09:35	uh
0:09:37	we are
0:09:38	that script also create a text format of the transcriptions but will convert this into an integer format kaldi eli
0:09:44	just just so the cal doesn't need to have
0:09:47	for all of the program some kind of match between the
0:09:50	text an integer form of the uh
0:09:53	the the uh word
0:09:55	so this is the transcript the text format of the trans
0:09:57	oops
0:09:59	uh
0:10:00	next step after to that it the prep stuff we
0:10:02	the pair the graphs
0:10:04	there's is gonna be a
0:10:06	a bunch of openfst if commands and here like scripts to convert
0:10:10	from uh
0:10:11	from the lexicon to the fst format
0:10:15	the
0:10:16	the lexicon actually contains the silent
0:10:18	and the scrip
0:10:20	the the script kind of at and it's not something that
0:10:22	very deeply embedded
0:10:24	in county
0:10:25	these these little files if you've ever used at indy toolkit or openfst you'll know what these are
0:10:31	there
0:10:32	symbol tables
0:10:33	so so it's uh this uh
0:10:36	this is the text form of zero the text form of one et cetera
0:10:39	and E P S for epsilon is always zero
0:10:43	this is kind of uh
0:10:45	a common thing an fst toolkits knows of that idea zero
0:10:48	so
0:10:49	so this is why phones or one based because
0:10:52	zero is always reserved for epsilon
0:10:56	uh
0:10:58	so
0:10:59	the the
0:11:00	create
0:11:01	yeah all of the F so openfst does have a capability to put
0:11:05	to put symbol tables on the fst so the fsts we kind of know what the words were
0:11:10	we haven't used that because it
0:11:13	it quickly becomes very difficult once you decide to have simple tables on the fsts we've
0:11:17	we basically use integer format throughout the data
0:11:21	which
0:11:23	uh
0:11:23	it outputs these files this is the
0:11:26	gee the grammar or that could be not the language model used for decoding
0:11:30	so the lexicon
0:11:31	L L just got this one big is the lexicon the disambiguation symbols
0:11:35	and if anyone has read the papers of uh
0:11:38	more riyadh i'll the described the standard recipe for fst based uh
0:11:43	yes uh
0:11:44	i don't know what that is
0:11:46	little symbols like hash one hashed to
0:11:49	that they put on the lexicon of the ends of words
0:11:52	to ensure that term eyes ability
0:11:54	uh
0:11:56	i i i i but i'm not going to that in more detail or it's gonna
0:11:59	suck up the entire time of the tall
0:12:04	uh
0:12:06	pairing integer list of silence and nonsilence phones i we we we created little files
0:12:11	tape things like this
0:12:13	isn't needed later on by the scripts because occasionally a scribble need to know what the I D's of the
0:12:17	silence phones
0:12:19	and because the kaldi tools will at integer formats it's gonna need that
0:12:22	as a list of integers
0:12:25	uh computing remote okay so this is
0:12:29	this is just a command to
0:12:31	and vocal kind of other script
0:12:33	that uh
0:12:35	compute the mfcc
0:12:37	and and here is the command and i believe actually this before
0:12:40	it's uh
0:12:43	it basically write cm mfcc to some disk
0:12:46	and then this is gonna be a text file
0:12:48	that
0:12:49	contains
0:12:51	on each line is gonna be utterance id
0:12:53	and then
0:12:54	the law that this
0:12:56	filename
0:12:57	cool on
0:12:58	integer offset so it so it can kind of
0:13:01	directly go to that
0:13:03	part of the file using F C
0:13:07	okay i think this is what i just said
0:13:09	this is the uh
0:13:10	script format that i mentioned
0:13:12	of course
0:13:12	the script format is very generic this whole thing doesn't have to be of this formant it's any
0:13:18	is anyone of our extended filenames might include a real file something of this form
0:13:23	pi whatever
0:13:25	i i
0:13:26	i'm showing you what the archive format looks like that really is binary data so that you can
0:13:31	see
0:13:32	uh
0:13:34	yeah
0:13:35	but but in some cases there would be text i'm you you
0:13:38	you can give it the option to write in text and and very often it'll be a nice line by
0:13:42	line format
0:13:45	yeah i
0:13:47	uh
0:13:48	yeah so
0:13:50	i think i mentioned this before
0:13:51	the script
0:13:53	this key is an important concept
0:13:55	because there's this concept of uh
0:13:58	a collection of objects indexed by key in this case the string
0:14:02	think of a little bit like an S T L map
0:14:05	where you know it would be a map from string to whatever object
0:14:10	so
0:14:11	the archives in the script both the kind of make use of this concept
0:14:15	and i think that concept a little bit more detail in the next slide but the script format is the
0:14:19	key and then
0:14:20	some kind of extended filename blah blah blah
0:14:24	i think i mentioned this before but the types of extended filenames include actual file
0:14:30	a command
0:14:31	piping output
0:14:33	pipe symbol then the command which is like in
0:14:36	which is the input motion and out but this is only a pretty
0:14:41	very inputting from applied
0:14:43	an an offset into a file which is uh
0:14:47	which is useful where you where you want to write a big archive but have random access into
0:14:52	uh
0:14:52	so
0:14:54	this might seem like a very
0:14:55	in minus things i think it's important that
0:14:57	if you as you want to ever use count it's and to understand this how this work
0:15:01	so
0:15:02	there's the concept of a table
0:15:04	and this table doesn't really correspond to any like concrete objects or class it
0:15:08	some a generic comes that the idea or is
0:15:11	a collection of objects of some known type it's type known and of of
0:15:15	all indexed by
0:15:17	key which is the string
0:15:19	we we define a key is the non empty space free string for
0:15:23	that
0:15:24	and i was we have to make its space free so
0:15:26	otherwise we get it all kinds of issues
0:15:28	a
0:15:29	so
0:15:30	so
0:15:31	there was a street template plated class of that somehow relate to tables
0:15:35	is the table right ear
0:15:36	sequential table read or and random access table with that
0:15:40	so this two ways you this three ways you can do something with a table
0:15:44	you can write a table
0:15:46	and what you like you do with this is you
0:15:49	you'd say write me something with this key
0:15:52	and this object
0:15:53	that's gonna write it to the table
0:15:56	a any in you keep doing that
0:15:58	you can read a table chilly
0:16:01	which means repeatedly give the next key and giving the next subject
0:16:05	are you can random act you can do random access on a table which means
0:16:09	do you have a object this key and so no if so
0:16:12	give me the object
0:16:13	that's how you interact at is the templates the template it on
0:16:17	i gonna describe the next
0:16:19	like what they're ten it on
0:16:21	then not actually template on the object
0:16:23	it's it would be most natural to template
0:16:26	on the object that's in the table
0:16:28	but the problem is that doesn't work very well with uh
0:16:32	kind of fundamental types like integers and so on
0:16:34	because it "'cause" that normal cal the object
0:16:37	they have a read function and a right function have a particular
0:16:41	behavior
0:16:42	it's common all of them
0:16:43	but we can't just to see using the everything we want to read and write will have that form
0:16:47	because how would be writing to do is a how would write as T L like
0:16:52	and it and it would be ridiculous and my pin to somehow have to derive a class that's an integer
0:16:57	and give it a
0:16:58	thank
0:16:59	of the integers not class
0:17:00	last
0:17:01	so
0:17:02	we tend like um what we call a holder
0:17:04	a hold class as a cost that has set and read and write functions
0:17:09	uh
0:17:10	and it has a type that T inside it
0:17:13	that
0:17:13	is the actual type of the table whole
0:17:16	so
0:17:18	you know knowing all of this stuff
0:17:20	is
0:17:20	if you if the i lost to by not because you know a C plus plus are really doesn't matter
0:17:25	because this is i'm just it's how the channels of this like as am works but uh
0:17:31	you don't need to know this to understand the how the whole thing work
0:17:35	so
0:17:37	i think that's as an example of how
0:17:39	i the C plus plus level you use that the table comes
0:17:43	so
0:17:44	we we introduce things of terminology here that may seem a bit annoying but
0:17:48	eventually becomes clarifying
0:17:50	and i are specify or is a string that tells the table code had to read a table of check
0:17:56	uh
0:17:57	and his an example of one
0:17:59	uh
0:18:00	yeah yeah K call on this finally
0:18:02	so
0:18:03	the table code is gonna part this and
0:18:06	when it reads this it's as okay
0:18:08	yeah telling me that this is an arc
0:18:10	um thing that has the
0:18:11	key object key object
0:18:14	and this is an extended file name but tells you had to open a pipe of
0:18:18	or
0:18:18	open a tree
0:18:21	so
0:18:23	now this is the tight name
0:18:24	sequential table read template it on this holder tie
0:18:28	so this is
0:18:29	if were reading something of type in thirty two
0:18:33	so this is the use of the object name
0:18:36	the and initialize that we're giving it this string
0:18:39	so
0:18:40	it's soon as you initialise the object it it's
0:18:42	opening the high
0:18:43	it's say we gonna read from this
0:18:46	so
0:18:47	now we now we using the subject with thing what
0:18:49	for blah blah about
0:18:52	what is what this code is doing that's getting each key and to and from the sequential table read
0:18:58	and of course this and since this is the sequential table read that's what this subject
0:19:02	expect us to do
0:19:03	so the point is that
0:19:06	the maybe error it's right
0:19:07	some of the objects may not be there
0:19:09	sometimes you know something may go wrong
0:19:12	this
0:19:13	the template it code is gonna handle that so you're kind of
0:19:16	user level code
0:19:18	just see that as a
0:19:19	sequential access
0:19:22	i think this
0:19:23	uh
0:19:25	a stuff that have already told you
0:19:28	a there is some things that the table code has to do there were little bit tricky
0:19:32	one one of these things as
0:19:34	a very often you
0:19:35	once to do random access on objects that are
0:19:38	in an archive in that our K may maybe in a pi
0:19:41	as use a lot of high
0:19:43	and and the problem is that
0:19:44	suppose to some reason you ask you query a key that was not in the arc
0:19:49	in order it's of tell you know it wasn't in the arc
0:19:52	it's gonna have to read each one in the archive
0:19:54	go to the end of the pie and then saying no
0:19:57	but that means that i doesn't know that you're not gonna ask for something else to so has got the
0:20:01	store all of that stuff and member
0:20:03	so
0:20:04	in in order to uh
0:20:07	stop it from having to do this
0:20:09	you can specify and the are specified thing a little
0:20:12	common S calm cs S
0:20:14	a options that what tell it
0:20:15	this archive is sorted on key
0:20:18	are we gonna call this archive in sorted or
0:20:22	so basically that gives the code enough information to know that
0:20:25	i it doesn't have to store all the stuff and memory in it can still kind of
0:20:28	be correct
0:20:30	i'm gonna go a little bit fast is reduced
0:20:33	uh
0:20:34	i think we went through this computing mfccs
0:20:38	monophone training
0:20:40	so
0:20:42	you would invoke this script
0:20:44	uh
0:20:46	we gonna go through the script a little bit
0:20:48	it's set some some very than bash the directory were what are you doing your experiment
0:20:53	the features i think we so one of the strings for
0:20:57	this is
0:20:57	and are specified that i mentioned before
0:21:00	this
0:21:01	this part tell the that
0:21:03	we're gonna and separate this stream as an archive this tells the had to open the stream
0:21:08	and of course this is a i that's another colour the command has its own thing
0:21:12	sometimes it can can even be nested but beyond one level of nesting
0:21:16	be the shell escaping would become to thing
0:21:19	that
0:21:23	hi this is applied
0:21:25	what's so in fact this is an output is always that puts on the right
0:21:28	so what this is a it i think that out says
0:21:31	it's reading in this
0:21:33	this script file that says where the features a
0:21:36	and its output thing to an are kind of on the standard up
0:21:39	so and then this says that this whole thing is a pie
0:21:43	so this park gets interpreted by the program that
0:21:46	is given that
0:21:50	yeah
0:21:51	you can used to it
0:21:52	as you i
0:21:53	oh
0:21:55	huh
0:21:55	so
0:21:56	and that
0:21:57	is going to the monophone training script
0:22:00	uh we create a file called the X
0:22:02	slash last let's top L
0:22:04	that specifies the hitch an apology
0:22:07	to be uh
0:22:09	to the uh
0:22:10	the cow
0:22:12	so
0:22:13	i mean you you can this file for a fairly self explanatory a script repeat that
0:22:19	uh
0:22:21	there is uh
0:22:22	is it of the three state and then this is the kind of final state that
0:22:26	call of that the last state always has an X a probability of one
0:22:32	uh
0:22:33	this is a week amount to initialize the uh G M and
0:22:37	initialize that with the dimension of thirty nine outputs puts the here
0:22:41	and this also outputs a tree very trivial tree that doesn't really have any splits and it
0:22:46	and that's how we handle a monophone system
0:22:48	even a monophone system has a decision tree
0:22:51	it's just so that you don't have you know all the code is you five
0:22:55	uh
0:22:57	see if we have
0:23:02	okay
0:23:02	creating decoding graphs for training
0:23:05	or all of the kind of training script have a command of this form that
0:23:09	it creates an archive that have what has all of the fsts one for each are
0:23:13	and we do this
0:23:14	as a separate come "'cause" otherwise it would be too slow we'd only do on each iteration
0:23:19	a little bit too slow so
0:23:21	i take that the initial model
0:23:23	the lexicon a fist C
0:23:26	uh
0:23:27	trained a all this of the transcriptions an integer format
0:23:31	and that the put goes to this sprite that it just use that it and puts it in a
0:23:36	and that file
0:23:38	so uh
0:23:39	this is just the format of the dot track not try file it's just an integer at uh
0:23:44	transcription where we've can all of the strings so their integer
0:23:48	numbers
0:23:50	no of people like that
0:23:52	a
0:23:56	you okay so
0:23:57	the very first stage of uh monophone training is the flat start where
0:24:01	you uh
0:24:02	and of in
0:24:03	divide the utterance equally
0:24:05	a to the number of phones or whatsoever
0:24:07	and uh
0:24:08	create a an alignment a once to that so
0:24:13	yeah output of this program is something called alignment
0:24:16	which is
0:24:17	basically for each utterance it's a vector of integer
0:24:20	in to those integers is an id D that i touched on earlier we call a transition i D
0:24:25	it's something that behaves roughly similar to the P D F
0:24:30	index of P D
0:24:31	i D
0:24:32	but it has a little bit more information so you know the phone you know what the transition lot
0:24:36	so it kind of contains sufficient
0:24:38	information to to to update data
0:24:42	so we put that into this
0:24:43	program gmm max that
0:24:45	a light the suffix a means that it read an alignment
0:24:49	"'cause" the different versions of this program that we alignments that read in uh
0:24:53	posteriors gaussian in little posters and different thing
0:24:57	so
0:24:59	it takes the model it take the feature this of the shell variable is good bye
0:25:04	it read than this stuff from the input put an input
0:25:07	and the outputs of this
0:25:10	so but by the way
0:25:13	whenever something has a arc on it or or the C P O
0:25:16	that
0:25:17	that's an are specify or or doubly specify that means that as a collection of objects being passed around indexed
0:25:23	by key
0:25:24	but if you don't see that
0:25:26	like here and is just a file is just a single stream
0:25:29	is not there's no notion of index
0:25:34	a
0:25:35	there the
0:25:36	i think a cover this
0:25:38	a this oh you and that's is the gmm mm update
0:25:41	so it takes the
0:25:44	the original late to outputs the you model
0:25:50	so
0:25:50	the that this is the viterbi stage of training
0:25:53	what what we do during training is on so on selected to rate it's iterations we redo the alignment
0:25:59	we don't necessarily do that every iteration simply because
0:26:02	this is the
0:26:03	this is the thing that takes most to the time
0:26:06	and and it "'cause" it
0:26:09	if you have multiple gaussian Z
0:26:11	uh
0:26:12	this is not the only thing that's going on in training so it makes sense to uh
0:26:16	not do it every
0:26:17	so
0:26:19	i think this is pretty obvious that should be to that she's here
0:26:23	it
0:26:24	i you give it the beam
0:26:25	with the model this is the yeah this is this stream that that has all the fsts on it
0:26:32	features
0:26:33	and uh
0:26:35	it's gonna right
0:26:36	it's gonna
0:26:37	sorry oh that's a as an option i mentioned briefly options
0:26:41	on these are specify or or in this case a double is just five it so that a right in
0:26:44	text format
0:26:46	the default is binary but you could do common be if you want to emphasise that
0:26:51	uh
0:26:51	you monophone training we re align on almost every iteration because
0:26:56	thing i found that that would better or something thing or maybe it's because you usually have single gaussian
0:27:01	uh during right
0:27:03	i i think
0:27:04	after that you system is they do to but you don't have to
0:27:07	we so often so typically during it kind of you pocket triphone training we'd only
0:27:12	realigned three or four time
0:27:14	uh
0:27:15	so mix up to increase the number of gaussian
0:27:17	is maybe slightly against the whole called a philosophy but
0:27:20	it's just an option to the update program
0:27:24	uh
0:27:25	the way we allocate gas since we don't have a constant number of gaussians per state
0:27:30	we
0:27:31	we provide uh
0:27:33	it it's a power law it's proportional to the count
0:27:36	and this shouldn't be no but by the should be not point to i don't know why that
0:27:40	way
0:27:41	uh
0:27:42	it it's just slightly better than having a constant number
0:27:46	so yeah just schedule we used to allocate the guest in that's typically
0:27:50	you start from a set the number
0:27:52	you linearly increase
0:27:54	and then it
0:27:55	levels out it would probably be more natural to increase with the log
0:28:00	kind of increase of the power law something but
0:28:02	it just didn't work as well
0:28:03	was but it to do a linear
0:28:08	i
0:28:09	uh
0:28:10	okay
0:28:11	so a triphone training
0:28:13	the first stage is we
0:28:15	we align all of the data of that we uh
0:28:18	for the monophone we use the subset because this is no point
0:28:22	so just small system
0:28:23	so we re all of the data and we've output alignment
0:28:27	we
0:28:27	we we Q my a special kind of stats for training the decision tree
0:28:31	what this is
0:28:32	for each unique tries
0:28:34	triphone context in this case
0:28:36	it's gonna a malay single gaussian
0:28:39	well the stats for a single gaussian and this is gonna and was to train the tree the standard way
0:28:44	so that the just stuff in the script that kind of
0:28:47	automatically that some automatic clustering produces question
0:28:51	we don't use hundred or questions is as the hassle
0:28:54	find them
0:28:56	and and this
0:28:58	it a these a producing various files that will be read like D
0:29:01	so a lot of the actual control of how the tree get set up is some of the script level
0:29:07	a building the tree this is the colour command the bill the tree
0:29:12	what that's actually does is the it goes that to fifteen hundred leaves
0:29:15	and then it kind of clutches it like down a little bit
0:29:18	but by nonpredictable amount because
0:29:20	yeah chills threshold it uses to
0:29:23	you the clustering of to the initial splitting
0:29:26	is the same as the kind of last successful split
0:29:29	so you can't quite predict have big it'll be but normally it's tricks by twenty percent
0:29:34	or is what you
0:29:35	it's give you
0:29:37	so you initialise the model
0:29:39	for this tree this this this program doesn't know if it's gonna be a gmm or of for for various
0:29:44	gmm or S gmm you gonna create
0:29:47	oh to separate program to initialize the model
0:29:50	uh
0:29:52	is a nice feature of the whole alignment
0:29:54	onset
0:29:55	you can actually take can a and produce for one model
0:29:58	and converted it to kind of be valid for another model
0:30:01	so that means that
0:30:03	you can avoid it's a certain amount of uh we generating a
0:30:10	okay so if you want to decode you have to build the decoding graph
0:30:14	this is a this is the how we and be a graph generation
0:30:18	and the think that is doing first to compose is L with G
0:30:22	it's a as minimize is you get a L G
0:30:25	that's an
0:30:26	this some
0:30:27	so stuff for disambiguation symbols going on
0:30:30	uh
0:30:31	if if you and are gonna go through that
0:30:35	then you have to compose the in the context of christian
0:30:38	it kind of expands the file that a context-dependent phone
0:30:42	and
0:30:43	that's a kind of dynamic generation of uh
0:30:46	the context of T going on that happens within member in here
0:30:50	and not gonna go
0:30:51	and more do sell uh what's going on here
0:30:54	and then this last one
0:30:56	uh make eight trends use so that this hey jeff T that
0:30:59	on the
0:31:01	a basic we expand that the hey jim and
0:31:03	so on the right to the context-dependent phones on the left
0:31:06	you got the P D S but uh adds all the stuff that network
0:31:09	so this is grading to see that does that and the last just to uh
0:31:14	compose hates with C L G
0:31:18	uh
0:31:19	it's M and eyes i
0:31:20	yeah
0:31:21	oh in we did that without self loop so at the end
0:31:24	i this is to to just
0:31:26	make these more memory efficient
0:31:28	we don't we wait till the very end had the self
0:31:34	so
0:31:34	this just goes prefix prefixes
0:31:36	for the decoding script
0:31:38	we create a shell variable that tells that what the features will be
0:31:42	then we invoke this
0:31:43	uh
0:31:44	program that's three decoders and
0:31:47	this what the affect the uh come man it could be jen and decode code
0:31:50	simple faster or D
0:31:52	it's the the kind of medium one
0:31:54	there
0:31:54	the G and decode simple is mainly that for debugging
0:31:57	"'cause" it's so simple that
0:31:59	you know the can be anything wrong with it sorry
0:32:01	we just we just compare the to
0:32:05	i
0:32:05	yeah
0:32:06	so
0:32:08	decoding coding be missed twenty it's is the beam min kind of language model
0:32:12	scale
0:32:13	we only a the acoustic the language model
0:32:16	uh
0:32:17	this is just a get more human readable out put
0:32:21	the model
0:32:22	the F S T
0:32:25	features
0:32:26	this isn't a W specify specifies had to write transcription
0:32:31	it's says do it
0:32:31	and text for format
0:32:33	but D the can be integers
0:32:35	we're gonna have to change them to uh
0:32:38	the text before scoring if one
0:32:40	format
0:32:41	and this is a this is the the alignment
0:32:44	this is
0:32:45	it is a really useful like Q just decoding
0:32:47	but
0:32:48	i you might want to do adaptation late to using using that
0:32:51	decoding us that these supervision
0:32:53	so it just kind of
0:32:54	we just always produce a "'cause" it doesn't cost
0:33:02	okay so
0:33:04	i think that's basically comes to the end of this talk
0:33:07	a given you a very vague idea of have the scripts work what's and them
0:33:11	or we see there's a lot more details that you'd the
0:33:13	to find out before using them but a lot of that stuff is in the documentation
0:33:17	you have to kind of dig around the documentation
0:33:20	i've been told that it's not very uh clear where to star
0:33:23	but there is a lot of it there so if you just willing to read at all
0:33:27	a the thought also it it have heavily cross reference
0:33:31	so
0:33:31	if you five something that kind of related to what you need
0:33:34	that usually be a link that you can click on that will take it to what you do
0:33:39	i
0:33:41	okay so that's it
0:33:49	for any question
0:33:52	uh
0:33:55	yeah
0:34:03	uh
0:34:04	you never really deals directly with
0:34:07	with any of those symbols because all in integers
0:34:10	so
0:34:11	it really doesn't matter as as can and what it is so yeah i seen that you could do any
0:34:16	t-f eight
0:34:21	well those are those have to be
0:34:24	does have to contain no white space and been on them
0:34:28	so
0:34:29	i don't i don't
0:34:31	yeah it's not is not gonna worry but you T F a long is not white space i think that
0:34:35	in never checks but it's actually ask
0:34:38	but i mean
0:34:39	i think you to if a is that's that if it's not
0:34:43	but is no we gonna be white speech "'cause" it's all always more than a hundred and twenty eight
0:34:47	i don't
0:34:49	but it it should be a a for any have a
0:34:51	i i don't really think is a good idea to put you T F A in those things
0:34:54	because i mean
0:34:58	don't
0:34:59	you have old fashion
0:35:04	uh_huh
0:35:08	yeah
0:35:09	i i i think it should work but
0:35:11	you gonna be concerned about this
0:35:13	about this
0:35:14	shell the "'cause" it could be that the shell of doing some kind of
0:35:17	manipulation on the lines of that foundation go weird characters
0:35:21	i don't know with the will work
0:35:23	you know but
0:35:26	but it's
0:35:27	it should be easily changeable to handle that if it really becomes an issue
0:35:31	uh
0:35:38	uh_huh
0:35:42	uh
0:35:44	i believe that can i think our i reasoning right
0:35:46	that that i don't recall call it things ever been tested
0:35:49	a pulse go
0:35:49	so it it's two percent
0:35:53	well the five that i shall can have those probabilities but
0:35:56	i think the perl script that
0:35:58	great the lexicon actually
0:36:00	has a flag but is post except them
0:36:02	or least that that one point but
0:36:04	there of that the but it's just it's just that you know that you line for script so it's not
0:36:08	like
0:36:13	yeah yeah so it that really care whether a lexicon has probability is just an F T
0:36:18	so yeah
0:36:23	the
0:36:25	uh
0:36:26	yeah
0:36:37	ooh
0:36:38	well uh
0:36:40	they can get very large and
0:36:42	i i mean you with ones that we like
0:36:44	one and a half ago
0:36:45	i think that was with a a somewhat and trigram lm
0:36:50	they do get very big but at some point we gonna create coders that
0:36:53	a from that problem
0:36:56	i mean i think the festive
0:36:57	from a as a it's great because he D back its simple but
0:37:02	maybe the memories
0:37:04	we're gonna work on
0:37:08	if there's no more questions i guess we can call it a day
0:37:11	oh one more
0:37:14	i guess for but if there
0:37:20	oh
0:37:21	you have to redo them music they own oh oh
0:37:26	oh
0:37:28	yeah
0:37:32	yeah

RUNNING THE EXAMPLE SCRIPTS

Kaldi Workshop

Přednášející: Dan Povey, Autoři: Dan Povey