Speech Transcript - TWO-CHANNEL POST-FILTERING BASED ON ADAPTIVE SMOOTHING AND NOISE PROPERTIES

0:00:13	a with the money
0:00:14	yeah yeah
0:00:15	ladies and and you're
0:00:17	uh uh and and sends a long
0:00:19	communication code
0:00:21	to that
0:00:22	is due of a code good chinese got than all science
0:00:25	uh it's a great honour for me to speak up out to channel post got need based on what they
0:00:30	have to use the moody and noise properties
0:00:32	i have divided them my
0:00:34	presentation i been into four
0:00:36	house
0:00:37	you know first pass i would like to
0:00:39	say something about background and the motivation
0:00:42	in an S hat
0:00:43	i would like to insert
0:00:44	i would like to study a to channel all get getting at reasons in theory
0:00:49	in has three
0:00:51	two schemes are proposed
0:00:53	to improve noise reduction without introducing audible
0:00:56	so peace distortion
0:00:58	and in a last pass
0:01:00	experiments or out will be giving and we we are makes and conclusions
0:01:06	as as we know
0:01:07	uh
0:01:08	uh the noise the word it's becoming rather and with noise almost everywhere
0:01:13	and a background reduce use both speech quality and the speech
0:01:16	intelligibility
0:01:18	what's more small also increase the dissonance for T
0:01:21	so in practice
0:01:23	speech enhancement
0:01:24	uh including a single symbol channel speech enhancement and the channel which an enhancement is often in have twelve
0:01:33	oh we can't have a speech enhancement sort a spatial feeling and the whole to as in a in a
0:01:37	have a a in at the table
0:01:39	we found that
0:01:41	post-filter filter is more efficient for a diffuse noise field
0:01:46	and uh we compare the single channel speech enhancement with the model channel
0:01:51	filtering filter read and you know no uh table
0:01:54	uh from a male table we can find that we can find a to imitation of the single channel speech
0:01:59	enhancement
0:02:00	first
0:02:01	now station noise
0:02:03	of of them can all be well suppressed
0:02:05	second
0:02:06	speech intelligibility
0:02:08	can no be well
0:02:10	can be significantly improved
0:02:12	in generally
0:02:13	so
0:02:14	multi-channel a a is of them pretty for
0:02:19	we present a three gain functions that of a used
0:02:22	for post filtering
0:02:25	uh previous study
0:02:26	the we ask that
0:02:28	and noise only segments
0:02:29	we
0:02:30	the noise reduction should be in finny in theory
0:02:34	but
0:02:35	and we have a that use be M means that the noise reduction
0:02:39	is all and not a not
0:02:41	very
0:02:42	uh we want to study
0:02:43	the reason in the theory
0:02:45	this also
0:02:46	brings out a new
0:02:48	but your and how can we do to improve noise reduction without introducing audible speech
0:02:54	distortion
0:02:57	that's all i would like to speak about uh
0:03:01	background and the modulation
0:03:03	now that's ten to the first question why don't know it's action
0:03:07	is no in infinity
0:03:09	even even a noise only segment
0:03:12	we have to some soon two assumptions
0:03:15	first
0:03:15	the thing or a lot
0:03:17	i don't two microphones
0:03:18	a gaussian distribution
0:03:20	second
0:03:21	the noise it's it's and the at to model and that these two assumptions
0:03:27	the P we all the ground four but chi square distribution
0:03:31	with two degrees of three then
0:03:33	and in theory the real and the image only task of look the P both plus
0:03:40	it's a but it
0:03:42	oh what you me to be use in all level as distribution do you it makes it difficult to obtain
0:03:47	that a
0:03:48	that this should fusion of the cross
0:03:50	spectra
0:03:51	so we use of the gaussian distribution
0:03:55	to approximate a lower level as
0:03:58	distribution we present it in a
0:04:01	you know low location
0:04:06	a a as we know the all spectra and cross
0:04:09	spectral can be a a ten by averaging L
0:04:12	independent of friends
0:04:15	all the P
0:04:16	or a uh
0:04:18	P read all right
0:04:19	so
0:04:20	level
0:04:22	that this your job of the or those vector that it's speech and of problem
0:04:26	cross a
0:04:27	can be of a the has to be
0:04:29	we the pdf of low all spectra and a cross spectra
0:04:33	the pdf of the gain function can be a at and in theory
0:04:38	we brought of the pdfs of logging function in is a was the difference a a
0:04:44	smoothing factor of a
0:04:47	uh as can be seen from this figure
0:04:50	the
0:04:51	theoretical or out
0:04:53	face power while with the empirical with out
0:04:57	oh to obtaining the pdf of the gain functions
0:04:59	we can and a nice
0:05:01	the theoretical on means of noise reduction
0:05:04	and amount of musical noise as link
0:05:06	oh think that some pearl
0:05:08	we present
0:05:09	a a like to you meets of the noise action
0:05:13	the the only
0:05:14	the like to call me miss off noise reduction can be a but by
0:05:18	can but expected value of the gain function
0:05:22	and we proud to the
0:05:24	noise reduction of a different values of uh
0:05:29	was a different well to of the
0:05:31	smoothing factor in these figure as can be seen from this figure go of the smoothing fact that increases this
0:05:37	the noise reduction can be improved
0:05:40	but this that E we can i as is question why don't noise reduction is no in you even at
0:05:46	noise only segments
0:05:48	and if and only if
0:05:50	that's the in fact uh approach is one noise reduction
0:05:54	can be in
0:05:58	but this study sounding max can be summarised as follows
0:06:02	if not move in fact uh it's no last in that
0:06:06	you car noise if you have well you to a like heart of the gain function having large very
0:06:12	it's better to use a large value of a less smoothing factor to increase both noise reduction and we use
0:06:17	musical noise
0:06:19	we have to make a chair off between and
0:06:22	noise reduction and estimation pass
0:06:25	but properly
0:06:26	so acting this smoothing factor
0:06:29	now a less to a second question how can we do to improve noise reduction
0:06:34	without introducing audible speech distortion
0:06:38	we propose to study in
0:06:40	in this paper
0:06:41	the first thing is that adaptive time-frequency frequency keys P
0:06:45	as we know for a two channel post filtering out regions
0:06:49	the sudden and change of the system only a "'cause" at a a that speech on site and off size
0:06:55	so it's better to use a small value of the smoothing factor
0:06:59	and that it's that
0:07:00	speech on side and all side
0:07:02	which use a tools that screen
0:07:05	for that it a speech on site
0:07:08	the smoothing factor
0:07:11	it it is it i mean the by the signal-to-noise ratio
0:07:15	for that is that a speech all side
0:07:17	the smoothing in fact uh it's gradually increased from zero to one
0:07:23	that's take again
0:07:24	if sample
0:07:26	we
0:07:27	the it's now
0:07:28	we
0:07:29	that let us the don't it's out of the proposal
0:07:32	D
0:07:33	we can find a from this
0:07:35	the
0:07:36	a little figure
0:07:38	and at it that speech on side
0:07:43	and desired
0:07:44	so based on side a all size
0:07:46	that's smoothing in fact uh has a small value
0:07:49	but is very close to zero
0:07:52	and a noise only segments
0:07:54	the smoothing factor fact is close to one
0:07:57	the
0:07:58	the without out i expected it
0:08:04	the second the skiing is the adaptive which or noise floor action
0:08:09	in order to mask me car noise can much speech enhancement
0:08:13	all thing use a constant with joe noise for all
0:08:16	with a even better
0:08:18	the after the equation
0:08:20	i have a bit on this
0:08:22	psychoacoustic fat
0:08:23	is difficult
0:08:25	for a home to mask a of noise
0:08:28	that's
0:08:29	is is the of
0:08:30	it's better that of further suppress the
0:08:32	tonal no and so we propose a modified gain function
0:08:37	two
0:08:38	uh we use the number a
0:08:40	K L to it at the tonal all components
0:08:44	but using in a modified gain function
0:08:46	the noise reduction
0:08:48	could be improved without introducing audible as
0:08:51	speech
0:08:52	enhancement
0:08:53	because we only increase the amount of noise
0:08:56	with action and of peaks of U
0:08:59	estimated noise power spectral density
0:09:03	that's the an example
0:09:06	the noise
0:09:07	no no it's it's be thing else like your
0:09:11	why noise
0:09:12	and then a noise it's is added to clean speech
0:09:17	and a segmental signal-to-noise ratio about an pain is a base
0:09:22	uh that's this end
0:09:24	to is
0:09:27	is all ills
0:09:29	uh the first one it's the noise this speech
0:09:34	right
0:09:34	a speakers no
0:09:36	hmmm
0:09:40	so
0:09:43	uh
0:09:45	the are we can also see that with the the the experiment bits out from this period
0:09:50	from this
0:09:51	period right
0:09:53	uh
0:10:10	uh
0:10:10	we can find that
0:10:12	the enhanced speech in enhanced speech
0:10:14	with a a that you race still noise for action yeah
0:10:18	the thing else like a around as the operator of
0:10:21	and with the in hand with the proposed
0:10:25	D
0:10:25	and that the thing is like component
0:10:27	and you can pretty T
0:10:29	suppressed
0:10:39	uh uh
0:10:40	but any i would like to compare to propose a re than with the
0:10:44	additional signal to channel post filtering
0:10:48	a low is at of an no set setup
0:10:49	a some as as for
0:10:51	we used
0:10:52	to map foreigns produced by boat at that distance be seen a two microphone is half a meter
0:10:58	the reverberation distance of a no is about one meter the noise speaker it's look at about four meters away
0:11:04	from the center of the two microphones
0:11:07	that it's that a speech
0:11:08	it's located in front of a center my from at a distance of a
0:11:13	have a meter
0:11:15	we with of the measured coherent
0:11:18	of the noise because it in
0:11:21	you know task to no
0:11:22	using in thought he to lie
0:11:25	as a comparison
0:11:26	we
0:11:27	we problem
0:11:29	the
0:11:30	the coherence of the diffuse noise field
0:11:33	using in a sorry the lie
0:11:35	we can see from this figure that
0:11:38	the
0:11:39	the coherence of the noise has a small value
0:11:42	so
0:11:43	we can assume that that
0:11:45	the speech
0:11:46	no the noise it's great it
0:11:50	we present the comparison and out of the set of mental as and not improvement
0:11:55	and the pesq improvement in these two tables
0:11:59	as can be seen from these tables
0:12:01	with the proposed as a read and
0:12:04	consistent improvements of pulls the
0:12:07	segmental a single to noise ratio
0:12:10	and the pesq Q can be a but a
0:12:15	now that's ten to the conclusion
0:12:19	in practice
0:12:21	uh uh that's kind a
0:12:22	the conversion
0:12:24	now we can't i'm sort of quite channel why a noise reduction is no if a need you and
0:12:29	at noise it's only segment
0:12:33	uh
0:12:33	we also propose
0:12:35	to help re
0:12:36	two
0:12:37	improve noise reduction without introducing audible speech distortion in practical consideration
0:12:44	one it's the adaptive time-frequency smoothing scheme
0:12:48	and the D out a it's the noise property G mean adaptive joe noise floor selection D
0:12:55	but any i would like to emphasise that
0:12:57	but two schemes could be applied to any speech enhancement and then need to be
0:13:02	then need to estimate also of cross vector
0:13:06	and the gain function
0:13:08	thank you for a attention
0:13:12	i
0:13:15	do real very
0:13:17	i we
0:13:17	but from many questions
0:13:19	so any questions
0:13:24	any any questions from all
0:13:28	know
0:13:31	i is it possible to try to demonstration restriction
0:13:34	be sorry
0:13:35	yeah to to use the microphone yeah
0:13:43	do do you use you in
0:13:45	excuse
0:13:46	so there's the question for you
0:13:48	well
0:13:48	right
0:13:49	do you use a computer nor is used a little D
0:13:53	where a diffuse noise
0:13:55	action really
0:13:56	uh
0:13:57	you know
0:13:58	i actual or just them the
0:14:00	you you used to do to a with or me right so for the assumption
0:14:04	or the with some correlation white
0:14:08	so
0:14:09	or just one the ring are
0:14:11	do do you
0:14:13	a how we you square X i experiment incorporating
0:14:16	click you the we so chose
0:14:19	uh
0:14:20	um
0:14:22	the direct and noise
0:14:24	or in any kind of the voice such just but we're always of or something
0:14:31	you someone
0:14:36	uh i think this is a very good question
0:14:38	because
0:14:39	uh
0:14:40	uh i think post filtering the post that that we don't i expect it a coherence space
0:14:46	but feel tuning at queen and we have
0:14:49	we have
0:14:49	to was size of that the noise is
0:14:52	and at it
0:14:53	and the and noise
0:14:54	the noise field is
0:14:56	uh
0:14:58	if a noise is created a i think is
0:15:01	is small
0:15:03	it's more efficient to use the
0:15:06	uh
0:15:07	a
0:15:11	to use
0:15:13	sorry
0:15:18	to use a special of of to me a reader
0:15:25	you O any questions
0:15:28	you used to solve system okay
0:15:31	uh what the what we the demo
0:15:34	yeah phone
0:15:37	okay okay it's
0:15:49	means
0:15:58	i
0:15:59	oh
0:16:03	oh
0:16:08	uh this is that you hand to speech without using Z
0:16:11	the base you or noise for
0:16:15	selection action and this is a but i
0:16:18	oh
0:16:19	oh
0:16:20	i
0:16:22	yeah
0:16:23	oh
0:16:24	i
0:16:25	i
0:16:26	i
0:16:28	and a or a signal so components
0:16:31	ask
0:16:32	as press
0:16:35	and then or or
0:16:36	or and you know
0:16:38	the noise is i
0:16:39	i
0:16:40	oh
0:16:45	ah
0:16:52	okay any question
0:16:55	Q and a any questions
0:16:57	remote more questions
0:16:59	okay if not the strings so speaker

TWO-CHANNEL POST-FILTERING BASED ON ADAPTIVE SMOOTHING AND NOISE PROPERTIES

Industrial Technology for Speech Processing Applications

Presented by: Chengshi Zheng, Author(s): Chengshi Zheng, Yi Zhou, Xiaohu Hu, Xiaodong Li, Institute of Acoustics Chinese Academy of Sciences, China