| 0:00:15 | oh |
|---|
| 0:00:16 | on the |
|---|
| 0:00:18 | traditional speech recognition |
|---|
| 0:00:20 | and the first paper |
|---|
| 0:00:27 | i think |
|---|
| 0:00:30 | speech |
|---|
| 0:00:42 | mar |
|---|
| 0:00:44 | vol |
|---|
| 0:00:47 | huh |
|---|
| 0:00:48 | really |
|---|
| 0:00:50 | actually i |
|---|
| 0:00:51 | yeah |
|---|
| 0:00:57 | okay so this |
|---|
| 0:00:58 | the stock |
|---|
| 0:01:00 | you will notice that some |
|---|
| 0:01:02 | overlap with the present |
|---|
| 0:01:04 | that is |
|---|
| 0:01:06 | a couple the signal and i'll try to point out the differences |
|---|
| 0:01:11 | the |
|---|
| 0:01:12 | i'll start with the motivation |
|---|
| 0:01:14 | and how we prepared what the data is now we prepared it |
|---|
| 0:01:19 | we actually and i |
|---|
| 0:01:21 | average talks in this |
|---|
| 0:01:24 | in this |
|---|
| 0:01:26 | we actually use more than just a cepstral |
|---|
| 0:01:28 | a system so it's i'll explain what the systems are using |
|---|
| 0:01:33 | that i describe the results |
|---|
| 0:01:36 | and some questions for that should be multiple question |
|---|
| 0:01:40 | future work |
|---|
| 0:01:44 | so the as we all know sre has a |
|---|
| 0:01:49 | just telephone speech is all that |
|---|
| 0:01:51 | however |
|---|
| 0:01:53 | our keynote presentation yesterday |
|---|
| 0:01:56 | and how to train |
|---|
| 0:01:59 | interviews with a microphone |
|---|
| 0:02:02 | recorded speech |
|---|
| 0:02:04 | and |
|---|
| 0:02:05 | a total of twenty ten sre and this was still distributing the data as to |
|---|
| 0:02:12 | where telephone speech |
|---|
| 0:02:15 | eight khz sampling rate the you're not coding which is a lossy |
|---|
| 0:02:20 | oh coding scheme that is works well for |
|---|
| 0:02:24 | for telephone speech |
|---|
| 0:02:25 | so one of the things that we're gonna looking at is the effects of those |
|---|
| 0:02:31 | two factors and what you remove the word relaxes |
|---|
| 0:02:34 | constraints out that the data is encoded |
|---|
| 0:02:37 | and this is the part where there's you know where there's overlap with bills talk |
|---|
| 0:02:43 | from tuesday |
|---|
| 0:02:44 | and i should point out is a difference not |
|---|
| 0:02:47 | we actually did not need of the system front-ends |
|---|
| 0:02:51 | for acoustic modeling the same |
|---|
| 0:02:54 | we always use a telephone from that |
|---|
| 0:02:56 | still see some interesting differences |
|---|
| 0:02:59 | that's the part that is different |
|---|
| 0:03:01 | the studied |
|---|
| 0:03:05 | and then we look at the second variable which is |
|---|
| 0:03:09 | how much better it can get if you actually use an asr system |
|---|
| 0:03:13 | some of our systems use a speech recognition and of course the quality of the |
|---|
| 0:03:18 | speech recognition is also |
|---|
| 0:03:19 | partly a function of the quality of the audio and |
|---|
| 0:03:24 | that is the second variable that you also |
|---|
| 0:03:30 | so the history of this work is that during the a train ten sre development |
|---|
| 0:03:35 | cycle which of course based on |
|---|
| 0:03:38 | what part as a really data |
|---|
| 0:03:41 | we notice that |
|---|
| 0:03:43 | you to write it is important control in the recording of the interviews |
|---|
| 0:03:47 | the |
|---|
| 0:03:49 | it's not be |
|---|
| 0:03:50 | no danger audio |
|---|
| 0:03:52 | recordings |
|---|
| 0:03:54 | using only |
|---|
| 0:03:56 | it's |
|---|
| 0:03:57 | the U haul as a result coding |
|---|
| 0:04:01 | because |
|---|
| 0:04:03 | the little boy |
|---|
| 0:04:04 | the |
|---|
| 0:04:06 | the compression best |
|---|
| 0:04:08 | here |
|---|
| 0:04:09 | this |
|---|
| 0:04:10 | so that this is |
|---|
| 0:04:12 | it was a problem for system |
|---|
| 0:04:15 | at least |
|---|
| 0:04:16 | so we started dating around |
|---|
| 0:04:18 | different |
|---|
| 0:04:19 | the effects of different |
|---|
| 0:04:20 | audio |
|---|
| 0:04:23 | and then we got lucky because we have another project that was independent of S |
|---|
| 0:04:27 | are going on sri time |
|---|
| 0:04:30 | which basically give us access to the fullband the original recordings of a portion of |
|---|
| 0:04:36 | mixer data was the basis yes or interview |
|---|
| 0:04:42 | and so we basically created and version of the interview data answer rate |
|---|
| 0:04:49 | and the rest mixture |
|---|
| 0:04:50 | and so on that which pointed to some interesting results which we actually recorded |
|---|
| 0:04:56 | and the rhino workshop following |
|---|
| 0:05:00 | okay |
|---|
| 0:05:01 | S three |
|---|
| 0:05:02 | right |
|---|
| 0:05:03 | but there but the results were |
|---|
| 0:05:06 | you know we have this |
|---|
| 0:05:07 | limited data set and it wasn't the complete dataset because there were still a microphone |
|---|
| 0:05:13 | data that was available to us in full bandwidth |
|---|
| 0:05:16 | so we set aside and actually last year |
|---|
| 0:05:19 | just released the complete sre |
|---|
| 0:05:24 | and microphone actually using the phone calls |
|---|
| 0:05:28 | in the in sixteen khz |
|---|
| 0:05:32 | and then we decided okay this is now we can basically look at this and |
|---|
| 0:05:36 | proper you know the complete evalset we can actually gets results |
|---|
| 0:05:42 | so that's |
|---|
| 0:05:44 | so we have two data sets were set is the three-way data and so the |
|---|
| 0:05:49 | original us rewrite data we repartition that's not what we use portion for the well |
|---|
| 0:05:53 | for training purposes such as training that by adding to the background data and intersession |
|---|
| 0:05:59 | variability training data |
|---|
| 0:06:01 | and we help we held out at a set of forty eight females and thirty |
|---|
| 0:06:05 | four male speakers |
|---|
| 0:06:07 | for development testing and that's the data for |
|---|
| 0:06:11 | so we found all the possible trials that data |
|---|
| 0:06:14 | and |
|---|
| 0:06:15 | i remember that the data |
|---|
| 0:06:17 | classified into short |
|---|
| 0:06:19 | conversations |
|---|
| 0:06:21 | and |
|---|
| 0:06:22 | we have those two conditions a long conversations were truncated actually |
|---|
| 0:06:26 | because |
|---|
| 0:06:28 | yeah |
|---|
| 0:06:28 | ten |
|---|
| 0:06:30 | condition |
|---|
| 0:06:31 | so these are the number of trials that resulted from that strategy |
|---|
| 0:06:36 | this was actually again this was the development set |
|---|
| 0:06:40 | is that so by the time to |
|---|
| 0:06:42 | developers |
|---|
| 0:06:44 | channel |
|---|
| 0:06:47 | that S which and data is again hasn't released version and using the extended trial |
|---|
| 0:06:53 | set so large number of one |
|---|
| 0:06:57 | oh actually with a wide band versions of both these phone calls recorded |
|---|
| 0:07:05 | you know microphone as well as |
|---|
| 0:07:08 | the |
|---|
| 0:07:10 | wideband |
|---|
| 0:07:12 | a number of them are only gonna look at all conditions that involved microphones both |
|---|
| 0:07:18 | training |
|---|
| 0:07:18 | so this |
|---|
| 0:07:19 | total five conditions |
|---|
| 0:07:26 | well |
|---|
| 0:07:27 | and in this presentation just with a lot of all focus on the eer hold |
|---|
| 0:07:35 | yeah |
|---|
| 0:07:36 | results and |
|---|
| 0:07:37 | the paper has there's the dcf results |
|---|
| 0:07:41 | and only if you have these |
|---|
| 0:07:43 | how the results differ but they differ qualitatively so |
|---|
| 0:07:47 | just one |
|---|
| 0:07:48 | say |
|---|
| 0:07:50 | the number so |
|---|
| 0:07:52 | stick to |
|---|
| 0:07:55 | okay here so we prepare the data |
|---|
| 0:07:57 | so we have the first condition that's the baseline condition that you can look exactly |
|---|
| 0:08:01 | right the euler coding |
|---|
| 0:08:04 | yeah that sounds this is how the data was delivered to us |
|---|
| 0:08:08 | yeah |
|---|
| 0:08:12 | we ourselves to a version of the data |
|---|
| 0:08:15 | where we took you based on what that the data we downsampled to eight |
|---|
| 0:08:20 | but we live in the east yeah you're holding |
|---|
| 0:08:24 | to avoid these the loss |
|---|
| 0:08:27 | right |
|---|
| 0:08:28 | and |
|---|
| 0:08:30 | we did not you builds a condition to use cost |
|---|
| 0:08:35 | this you saw |
|---|
| 0:08:37 | one thing we noticed socks is that we can use talks a lot of things |
|---|
| 0:08:42 | and actually there's different down sampling are |
|---|
| 0:08:46 | you provided once and you should try to see how the unit with your actual |
|---|
| 0:08:50 | task |
|---|
| 0:08:51 | what i think that you have to be very careful sauces that |
|---|
| 0:08:55 | they are not backward compatible so if you take the latest version of sauce will |
|---|
| 0:09:00 | not the same as well as the one that you might have used five years |
|---|
| 0:09:04 | ago so in fact very careful with keeping older versions around |
|---|
| 0:09:09 | to make sure that doesn't are we have to use an older version |
|---|
| 0:09:13 | this off to |
|---|
| 0:09:15 | get |
|---|
| 0:09:15 | results |
|---|
| 0:09:17 | there were some things that have |
|---|
| 0:09:18 | we tried |
|---|
| 0:09:20 | so i'm just one you're about so |
|---|
| 0:09:22 | maybe a little less |
|---|
| 0:09:24 | harshly not saying that you should use it all but you should use it with |
|---|
| 0:09:27 | great care |
|---|
| 0:09:28 | and that's why we have the sixteen khz |
|---|
| 0:09:31 | yeah |
|---|
| 0:09:32 | oh |
|---|
| 0:09:34 | then you're holding |
|---|
| 0:09:36 | a just a little bit more detail so we have basically we have a |
|---|
| 0:09:41 | a portion of the |
|---|
| 0:09:42 | five |
|---|
| 0:09:44 | available to us |
|---|
| 0:09:45 | seconds |
|---|
| 0:09:46 | flat encoding |
|---|
| 0:09:48 | actually |
|---|
| 0:09:49 | forty four khz |
|---|
| 0:09:51 | so we downsampled to sixteen and |
|---|
| 0:09:55 | and we use the this the segmentation tables |
|---|
| 0:10:00 | the segments for the development |
|---|
| 0:10:02 | some spot checking to make sure actually have exactly matching |
|---|
| 0:10:07 | dataset that matches best sex |
|---|
| 0:10:12 | that's all |
|---|
| 0:10:13 | that probably |
|---|
| 0:10:16 | now in our system |
|---|
| 0:10:18 | the first thing we do with all the microphone data is we find a wiener |
|---|
| 0:10:23 | filter doesn't formant by |
|---|
| 0:10:26 | from start with at least but is run evaluations |
|---|
| 0:10:32 | but |
|---|
| 0:10:33 | for |
|---|
| 0:10:34 | icsi ogi T |
|---|
| 0:10:37 | some |
|---|
| 0:10:38 | actually bottom that's |
|---|
| 0:10:42 | then |
|---|
| 0:10:44 | then we used a speech activity detection method that was that we're not problems but |
|---|
| 0:10:50 | seem to do |
|---|
| 0:10:51 | reasonably grounded was actually inspired by the point those two thousand and we use |
|---|
| 0:10:56 | combination of a |
|---|
| 0:10:58 | yeah hmm based speech activity detection |
|---|
| 0:11:01 | and we saw that |
|---|
| 0:11:03 | provided by |
|---|
| 0:11:05 | but the important thing is that we did not want to introduce |
|---|
| 0:11:09 | segmentation as confounding variables are comparison so we took the original segmentations which where |
|---|
| 0:11:16 | derived from your data and we kept that same segmentation fixed across all the different |
|---|
| 0:11:23 | i |
|---|
| 0:11:27 | at each modality is so you don't |
|---|
| 0:11:30 | a better |
|---|
| 0:11:31 | we say |
|---|
| 0:11:32 | i |
|---|
| 0:11:33 | yeah |
|---|
| 0:11:38 | okay so that we haven't is are things that the basis |
|---|
| 0:11:42 | so for some systems the right later |
|---|
| 0:11:46 | so it's we have to recognise this so the first recognizer our baseline recognizer |
|---|
| 0:11:52 | is a is a conversational telephone speech recognizer |
|---|
| 0:11:56 | oh that's |
|---|
| 0:11:58 | has been used for the last two evaluations asr evaluation so that's why |
|---|
| 0:12:03 | it's based on telephone data only it has two stages |
|---|
| 0:12:07 | second stage |
|---|
| 0:12:09 | two hypotheses from the first one for |
|---|
| 0:12:11 | unsupervised adaptation |
|---|
| 0:12:13 | yeah |
|---|
| 0:12:14 | and |
|---|
| 0:12:15 | it has |
|---|
| 0:12:18 | we measure the word error rate on some assume you six microphone data that we |
|---|
| 0:12:23 | have transcribers cells |
|---|
| 0:12:25 | below thirty percent |
|---|
| 0:12:28 | yeah |
|---|
| 0:12:31 | but that of course since we now have the of the wideband version of this |
|---|
| 0:12:35 | data we actually have the opportunity to improve the recognition |
|---|
| 0:12:39 | by |
|---|
| 0:12:41 | oh we have a different system it actually have very similar structure results in terms |
|---|
| 0:12:46 | of the algorithms for acoustic modeling so for the type of language models of what |
|---|
| 0:12:50 | that's very simple compatible to the first baseline system |
|---|
| 0:12:53 | but it was trained harder on meeting data so that was trained wideband data and |
|---|
| 0:12:58 | furthermore |
|---|
| 0:12:59 | that meeting it includes a far-field for |
|---|
| 0:13:03 | which is important because some of the work |
|---|
| 0:13:05 | the majority of the speech in the i |
|---|
| 0:13:09 | you condition also far field microphone |
|---|
| 0:13:12 | so we found that this would be a reasonable match to the to the into |
|---|
| 0:13:17 | data |
|---|
| 0:13:18 | and it will be read this on a on the interview data from sre ten |
|---|
| 0:13:22 | we found that the output twenty one percent word tokens than the old |
|---|
| 0:13:28 | and because our recognizer tends to just delete words when it hasn't for acoustic match |
|---|
| 0:13:32 | that's a pretty good indication that is |
|---|
| 0:13:34 | substantially more at |
|---|
| 0:13:37 | and that we used |
|---|
| 0:13:38 | we didn't have any transcribed sre ten interview data |
|---|
| 0:13:42 | other cell is we simply matching compare the asr accuracy on meeting data rich |
|---|
| 0:13:49 | which result was similar character so we used a far-field meeting data from |
|---|
| 0:13:54 | i don't know which one it was one of the nist |
|---|
| 0:13:57 | rt evaluation sets |
|---|
| 0:13:59 | and we found that the original cts recognisers had a very high rate |
|---|
| 0:14:04 | and then the first stage of our meeting recognizer which |
|---|
| 0:14:09 | is important it still is eight khz models actually performs this kind of cross-adaptation between |
|---|
| 0:14:14 | different kinds of acoustic models of the first stage uses |
|---|
| 0:14:17 | no models were trained using data already have much better accuracy |
|---|
| 0:14:23 | over forty percent and then the second stage with sixteen khz models and unsupervised adaptation |
|---|
| 0:14:30 | i |
|---|
| 0:14:31 | percent error rate so clearly a big improvement in terms of |
|---|
| 0:14:34 | speech recognition accuracy and probably consistent with the observation |
|---|
| 0:14:38 | that may be more talk spurt lattice |
|---|
| 0:14:45 | okay now to the systems |
|---|
| 0:14:47 | the |
|---|
| 0:14:49 | there were three systems over all these calls from a larger combination of systems that |
|---|
| 0:14:53 | were used in the official ester |
|---|
| 0:14:57 | twenty ten submission |
|---|
| 0:15:00 | so the first system is kind of our main state |
|---|
| 0:15:02 | cepstral system and use the |
|---|
| 0:15:05 | telephone band analysis of past twenty possible coefficients of the one K gaussians |
|---|
| 0:15:13 | and we didn't even bother to retrain the i-th channel i speakers for this |
|---|
| 0:15:19 | so we take those from the original |
|---|
| 0:15:22 | system |
|---|
| 0:15:23 | is based |
|---|
| 0:15:25 | data performs the t-norm |
|---|
| 0:15:29 | is a pretty run of the model system not using i-vectors but you know as |
|---|
| 0:15:33 | of twenty ten was a pretty standard state |
|---|
| 0:15:38 | okay that the two systems that to the asr the first one is or mllr |
|---|
| 0:15:42 | system uses a few days |
|---|
| 0:15:46 | model performs some |
|---|
| 0:15:48 | some |
|---|
| 0:15:49 | feature normalisation |
|---|
| 0:15:51 | oh yeah i a total of sixteen transforms |
|---|
| 0:15:56 | that come out by crossing rate for classes |
|---|
| 0:16:01 | and the two genders so we have made a specific reference models and female specific |
|---|
| 0:16:06 | right reference models with what they're always applied to both male data so yeah sixty |
|---|
| 0:16:11 | different transforms |
|---|
| 0:16:13 | okay the model that almost twenty five feature |
|---|
| 0:16:16 | features you the right now but then used as you know |
|---|
| 0:16:22 | i forgot to put in here that you don't perform now for |
|---|
| 0:16:26 | session |
|---|
| 0:16:30 | there is what i'm system |
|---|
| 0:16:31 | i've sounds very bad but that's give some |
|---|
| 0:16:37 | my brothers |
|---|
| 0:16:38 | consists of the relative frequency features are collected |
|---|
| 0:16:42 | the top thousand address and trigrams |
|---|
| 0:16:45 | you |
|---|
| 0:16:46 | the background data |
|---|
| 0:16:47 | and again using svm |
|---|
| 0:16:49 | or |
|---|
| 0:16:52 | okay so here for |
|---|
| 0:16:54 | the interesting |
|---|
| 0:16:55 | comparison |
|---|
| 0:16:56 | so we use these three different wait for conditions |
|---|
| 0:17:00 | and rank or cepstral system on the sre eight data and is short and long |
|---|
| 0:17:06 | data condition |
|---|
| 0:17:08 | and you can see clearly that the largest in about twelve percent relative |
|---|
| 0:17:13 | hums from |
|---|
| 0:17:14 | the dropping of this |
|---|
| 0:17:16 | impressive |
|---|
| 0:17:17 | coding as well |
|---|
| 0:17:21 | and that has a small additional gain a problem switching from eight to sixteen khz |
|---|
| 0:17:27 | sampling |
|---|
| 0:17:28 | and you might think well as a gmm front-end operates at eight |
|---|
| 0:17:34 | okay so what could possibly be improving by switching |
|---|
| 0:17:38 | sixteen |
|---|
| 0:17:39 | and the answer is that the noise filtering happens at the fullband |
|---|
| 0:17:44 | so the spec |
|---|
| 0:17:45 | subtraction |
|---|
| 0:17:47 | works better when you when you operate at |
|---|
| 0:17:52 | and then down sampling |
|---|
| 0:17:55 | right |
|---|
| 0:17:57 | so this was kind of an interesting result of us |
|---|
| 0:17:59 | is it |
|---|
| 0:18:01 | requires fairly minimal changes to the system and you know the gmms moments change so |
|---|
| 0:18:06 | that those |
|---|
| 0:18:08 | so |
|---|
| 0:18:10 | now we do the same system on sre ten data |
|---|
| 0:18:14 | and is a lot of numbers use of in and summarized |
|---|
| 0:18:17 | so basically you get pretty substantial gains |
|---|
| 0:18:20 | in order of ten percent relative |
|---|
| 0:18:22 | and the largest Z E R |
|---|
| 0:18:27 | for the vocal effort |
|---|
| 0:18:30 | very suggestive because i think that especially for low vocal effort affected by this |
|---|
| 0:18:36 | block coding but careful driving |
|---|
| 0:18:39 | datasets |
|---|
| 0:18:40 | very small |
|---|
| 0:18:43 | i should also point out as shown in the paper that the relative improvement on |
|---|
| 0:18:48 | a somewhat lower |
|---|
| 0:18:49 | the set of ten percent |
|---|
| 0:18:51 | i |
|---|
| 0:18:52 | so |
|---|
| 0:18:54 | page |
|---|
| 0:18:56 | but that i think that you get |
|---|
| 0:18:59 | vol |
|---|
| 0:19:03 | okay now more numbers |
|---|
| 0:19:05 | in our system |
|---|
| 0:19:07 | the benefits much more and here we have two different |
|---|
| 0:19:10 | contrast conditions we have E |
|---|
| 0:19:13 | so the mlr so the acoustic modeling always uses telephone speech so that we can |
|---|
| 0:19:19 | that you don't have to retrain anything the old telephone background data doesn't it doesn't |
|---|
| 0:19:25 | prevent us from using telephone data |
|---|
| 0:19:27 | the background model |
|---|
| 0:19:29 | however |
|---|
| 0:19:30 | the |
|---|
| 0:19:32 | the audio that you process before the final down sampling step okay the sixteen K |
|---|
| 0:19:38 | audio what you get the benefit from not having the lossy coding from doing that |
|---|
| 0:19:42 | the voice the filtering for that |
|---|
| 0:19:45 | and then you have two choices you can use the |
|---|
| 0:19:48 | first recognition step as your hypothesis of which comes from the eight khz models or |
|---|
| 0:19:53 | you can use the second stage comes from the sixteen khz models which as we |
|---|
| 0:19:57 | saw to sell better |
|---|
| 0:19:59 | and so we have both of these here and of course the second one |
|---|
| 0:20:02 | is consistently better |
|---|
| 0:20:04 | one very small |
|---|
| 0:20:06 | yeah |
|---|
| 0:20:07 | yeah |
|---|
| 0:20:09 | from a little data conditions |
|---|
| 0:20:11 | i |
|---|
| 0:20:12 | sorry |
|---|
| 0:20:13 | smart |
|---|
| 0:20:16 | seconds more accurate hypotheses |
|---|
| 0:20:19 | that's overall we see very substantial gains you only about twenty percent |
|---|
| 0:20:26 | hence |
|---|
| 0:20:26 | i six is a lot of numbers |
|---|
| 0:20:28 | a two-dimensional lots of these you know roughly |
|---|
| 0:20:31 | so you have one axis you see that the other axis |
|---|
| 0:20:35 | alright and you see that roughly two thirds of again come from the switch from |
|---|
| 0:20:41 | from eight khz to sixteen khz |
|---|
| 0:20:44 | still using the where |
|---|
| 0:20:46 | sorry then when you hear the asr |
|---|
| 0:20:48 | see |
|---|
| 0:20:49 | you get another |
|---|
| 0:20:51 | another time i |
|---|
| 0:20:53 | top |
|---|
| 0:20:54 | yeah |
|---|
| 0:20:55 | so this is pretty |
|---|
| 0:20:57 | cross |
|---|
| 0:20:58 | this condition |
|---|
| 0:21:00 | okay that just around a |
|---|
| 0:21:02 | the results |
|---|
| 0:21:03 | so the word n-gram system |
|---|
| 0:21:05 | voice operated much for |
|---|
| 0:21:09 | operating point |
|---|
| 0:21:10 | but the relative |
|---|
| 0:21:12 | these are much smaller |
|---|
| 0:21:14 | oh |
|---|
| 0:21:14 | and i would speculate one |
|---|
| 0:21:17 | whatever |
|---|
| 0:21:18 | just remark |
|---|
| 0:21:19 | okay |
|---|
| 0:21:21 | it's |
|---|
| 0:21:22 | prosody one second |
|---|
| 0:21:24 | overall |
|---|
| 0:21:26 | it does the same things |
|---|
| 0:21:29 | okay |
|---|
| 0:21:30 | so completely |
|---|
| 0:21:33 | recent changes and development data of course where is the question how to make |
|---|
| 0:21:38 | use of the of the full bandwidth the a priori that's a little to us |
|---|
| 0:21:42 | and this applies able to capture systems answers |
|---|
| 0:21:45 | instead use asr |
|---|
| 0:21:47 | yeah |
|---|
| 0:21:48 | studying this on two datasets sre right sre ten |
|---|
| 0:21:52 | probably a few conclusions |
|---|
| 0:21:55 | so there is substantial gains to be happens conference on image fifteen found |
|---|
| 0:22:00 | so we can express |
|---|
| 0:22:02 | no losses encoding is it is it is a big plus that's probably the biggest |
|---|
| 0:22:07 | plus the cepstral system |
|---|
| 0:22:09 | but you also get a small additional gain by doing voice filtering at the at |
|---|
| 0:22:14 | the full bandwidth |
|---|
| 0:22:15 | and then that'll systems get a significant strong using better asr and that of course |
|---|
| 0:22:21 | you need to find a mask asr system and we were quite successful using a |
|---|
| 0:22:26 | meeting recognizer that was trained for what nist rt |
|---|
| 0:22:29 | evaluations using for a few data |
|---|
| 0:22:32 | and what we have get like units |
|---|
| 0:22:35 | we have not actually changed the analysis bandwidth all the acoustic models |
|---|
| 0:22:40 | selsa we still using the telephone |
|---|
| 0:22:43 | yeah |
|---|
| 0:22:44 | for both these cepstral |
|---|
| 0:22:46 | okay |
|---|
| 0:22:49 | and this process |
|---|
| 0:22:50 | future work is quite a few |
|---|
| 0:22:54 | so obviously |
|---|
| 0:22:56 | the three systems |
|---|
| 0:22:57 | are from |
|---|
| 0:22:58 | so to them questions |
|---|
| 0:23:00 | so we the next step was used |
|---|
| 0:23:03 | combining all here we haven't done that |
|---|
| 0:23:06 | a question how much can you |
|---|
| 0:23:08 | nation |
|---|
| 0:23:10 | what will require a |
|---|
| 0:23:12 | quite a bit of work is to also read that the prosodic sys |
|---|
| 0:23:16 | which is |
|---|
| 0:23:17 | the |
|---|
| 0:23:17 | very nicely complementary so |
|---|
| 0:23:19 | yeah is the right so |
|---|
| 0:23:21 | two covariance data |
|---|
| 0:23:24 | and then the questions of course can we do better by retraining or acoustic models |
|---|
| 0:23:31 | and then using wideband data or alternatively can come up with some clever ways |
|---|
| 0:23:37 | and wideband data you sequence |
|---|
| 0:23:39 | bandwidth extension methods |
|---|
| 0:23:41 | or a simply modeling bandwidth mismatch as a as one dimension |
|---|
| 0:23:46 | right |
|---|
| 0:23:50 | i |
|---|
| 0:24:13 | that's |
|---|
| 0:24:16 | after i |
|---|
| 0:24:21 | well that's then you have to you have to use the telephone |
|---|
| 0:24:27 | that is |
|---|
| 0:24:29 | i |
|---|
| 0:24:29 | so |
|---|
| 0:24:30 | i |
|---|
| 0:24:32 | well first of all |
|---|
| 0:24:34 | it was a bigger |
|---|
| 0:24:37 | well |
|---|
| 0:24:38 | yeah |
|---|
| 0:24:39 | we also felt that a large number of speakers |
|---|
| 0:24:43 | oh |
|---|
| 0:24:44 | as the results |
|---|
| 0:24:46 | i |
|---|
| 0:24:48 | well |
|---|
| 0:24:56 | oh |
|---|
| 0:25:02 | okay |
|---|
| 0:25:04 | we look at the at |
|---|
| 0:25:07 | did you don't shake spectral shaping slightly down sampled at three |
|---|
| 0:25:11 | here |
|---|
| 0:25:14 | oh |
|---|
| 0:25:17 | the reader must achievement |
|---|
| 0:25:20 | yeah |
|---|
| 0:25:24 | you do not but |
|---|
| 0:25:25 | i think after that we didn't change it |
|---|
| 0:25:31 | yeah |
|---|
| 0:25:32 | such as |
|---|
| 0:25:36 | substantially higher local optimum |
|---|
| 0:25:40 | yeah |
|---|
| 0:25:42 | and it |
|---|
| 0:25:43 | use the beginning |
|---|
| 0:25:44 | something like this |
|---|
| 0:25:47 | so my |
|---|
| 0:25:50 | the like |
|---|
| 0:25:51 | oh |
|---|
| 0:25:55 | course |
|---|
| 0:25:57 | it's try to date |
|---|
| 0:26:00 | also |
|---|