tgt
stringlengths
1
2.76k
src
stringlengths
1
3.03k
lp
stringclasses
15 values
אז, התחלתי, אבל אני חושב שמעולם לא סיימתי לכתוב לך אימייל, כי הזכרת לי [אדם8] שמסנן ניבולי הפה עוד לא משולב.
So, I started, but I think I never finished an email to you, because you reminded me, [PERSON8], that your profanity filtering is not yet integrated.
en-he
ואני חושב שזו הודעה חשובה עבור [אדם5], ששוב נעלם מהשיחה.
And I think this is also an important message for [PERSON5], who has disappeared again from the call.
en-he
אז ההודעה החשובה היא שכן, טוב מאוד שאתה פועל כדי שהתוצאות שלך ישולבו וכולם צריכים לעשות את זה.
So, the important message is that yes, it's very good that you are actively pushing so that your results are integrated, and everybody should do so.
en-he
ובאותו עניין, אנחנו צריכים את ההגדרה, כדי שגם אתה תוכל לשלב ולבחון את זה בעצמך.
And at the same time, we need to have the setup, so that you can actually integrate and test it yourself.
en-he
אני קורא לזה, "שילוב של עשה-זאת-בעצמך".
So, I call it "do-it-yourself integration".
en-he
אז, [אדם8], כשעובדים עם [אדם5] וכשמתעדים את ההגדרות, צריך לוודא שהעמיתים מנסים את זה בצורה מספקת, כמו [אדם4] או אפילו [אדם9] לבדיקות של מודל השפה ועבור כולם.
So, [PERSON8], when working with [PERSON5] and when documenting what the setups are, make sure it is tested well enough by colleagues, such as [PERSON4] or even [PERSON9] for the language model checks and everybody else.
en-he
כך שבכל פעם שמישהו מפתח רכיב חדש ויעיל, צריך שתהיה לו גישה לצינור המלא בקלות יחסית, כך שהוא יוכל לנסות אותו בעצמו.
So, whenever someone develops a new useful component, the full pipeline should be accessible to him reasonably easily, so he can test it himself.
en-he
אז, השילוב של עשה-זאת-בעצמך הוא חשוב מפני שאחרת הכל נשאר רק אצלך, [אדם8] ואתה לא רוצה עומס יתר.
So, this do-it-yourself integration is important, because otherwise it will all remain on you, [PERSON8], and you don't want to be overloaded.
en-he
לכן, צריך לספק לאנשים האלה תשומות ותפוקות כגישה ראשונית לניסיון, מה שכבר נעשה.
So, you want to provide these people with inputs and outputs as the first testing approach, which has already been done.
en-he
[אדם4], נכון?
[PERSON4], right?
en-he
מסנן ניבולי הפה, האם הוא נבדק על לוגריתמים?
The profanity filtering, has it been tested on logs?
en-he
אני חושב שכן.
I think it was.
en-he
[אדם8], כן, אני חושב שהוא מבדק על לוגריתמים.
(PERSON8) Yeah, it was tested on logs, I think.
en-he
אז עכשיו הגיע הזמן לבחון אותו על צינורות חיים.
So, now it's time to test it on the live pipelines.
en-he
ושוב, אני חושב שיהיה יותר טוב אם [אדם8] יסביר ל-[אדם4] איך לעשות את זה, כך ש-[אדם4] יפעיל את זה עבור עצמו וכמה מהעובדים ויפעיל כמה מהקבצים הבעייתיים כמו שימוש ב-[פרויקט8] או משהו אחר.
And again, I think it's better if [PERSON8] explains to [PERSON4] how to do it, so that [PERSON4] runs it for himself some of the workers and live playing some of the problematic file like using [PROJECT8] or whatever.
en-he
פשוט להפעיל אותם.
Simply play them.
en-he
צריך לעקוב אחר פלט הצליל במכונה ולראות איך זה עובד.
Follow the sound output on your machine and see how that works.
en-he
כי רק כשמבצעים את ההגדרה האמיתית, יופיעו השגיאות האמיתיות.
Because only when doing the real setup will the true errors appear.
en-he
חשוב קודם לנפות את הבאגים באמצעות קובצי הלוגריתמים, ואז חשוב לנפות את הבאגים בצינור.
It is important to first debug it, using the log files, and then it's important to debug it in the pipeline.
en-he
ואם ניפוי הבאגים הזה יכול להיעשות על ידי מחבר הרכיב, כאן, במקרה זה [אדם4], זה יהיה הכי יעיל לכולנו.
And if this debugging can be done by the author of that component, here, in this case [PERSON4], it would be most efficient for all of us.
en-he
אז, [אדם8], אשר בבקשה שאתה מסכים עם הרעיון של שילוב עשה-זאת-בעצמך.
So, [PERSON8], please confirm that you agree with this idea of like do-it-yourself integration.
en-he
(אדם17) כן, לאנגלית במבטאים מרובים.
(PERSON17) Yeah, so for the multi-accent English.
en-he
אז אנחנו עכשיו [אדם14] מרכיבים את זה יחד לפתרון אחד טכני.
So, we are now [PERSON14] putting that together to just one technical solution.
en-he
הרעיון הנוכחי ש-[אדם14] עובד עליו הוא שהוא ייצור משפטים חדשים עם מילים שנאמרו במשפטים אחרים והוא יעשה את זה עם מספר דוברים.
The current idea that [PERSON14] is working on is that he will create new sentences with words that were spoken in other sentences, and he will do this across different speakers.
en-he
כך שיהיו משפטים של דוברים מרובים ואז אפשר יהיה לשפר את המבטאים השונים של הדוברים.
So, it will be really multi-speaker sentences, and therefore the robustness to the different accents of these speakers could be also improved.
en-he
זה ניסוי של אחד הרכיבים.
Ah, so that's one particle experiment.
en-he
ומאוחר יותר, נעשה עוד משהו בקשר למבטאים המרובים.
And later on, we may do something more about the multi-accent thing.
en-he
אז, המשפטים החדשים האלה בעצם ינסו לפתור את שתי הבעיות באמצעות ניסוי אחד.
So, these new sentences will actually try to solve two problems with one experiment.
en-he
בעיה אחת היא מודל השפה המובן.
One problem is the implicit language model.
en-he
כך שמערכת ה-ASR צריכה להיות את סט המשפטים הגדול ביותר האפשרי.
So, the ASR system has to see the largest possible set of sentences.
en-he
ואנחנו הולכים ליצור משפטים חדשים ממודל השפה של טקסט-בלבד על ידי הוספה של חלק הצליל לזה, כך שמודל השפה יהיה טוב יותר עבור ה-ASR, וההתאמה לדוברים שונים תהיה גם טובה יותר.
And we are going to create new sentences from the text-only language model by adding the sound part to that, so that the language model will be better for the ASR, and the robustness to different speakers would be also better.
en-he
ובשיחה אתמול, שמעתי עוד רעיון.
And in a talk yesterday, I heard another idea.
en-he
זה היה במהלך ההכשרה.
It was during the training.
en-he
הם הוציאו רצועות זמן ורצועות תדיר מהצליל.
They were dropping out time bands and frequency bands from the sound.
en-he
הם בעצם התאמנו על כניסות משובשות והם בעצם שיפרו את איכות המערכת.
So, they were training on disrupted inputs, and that also greatly improved the robustness of the system.
en-he
(אדם17) כל אחד אחר יכול לעבוד על העניין של המבטא שאינו שפת אם.
(PERSON17) Anyone else can work on the non-native accent thing.
en-he
אוקיי, אז עוד דבר שראיתי הוא – זה מסמך הבדיקה של יום שני.
Ok, then another thing that I spotted is - this is in the Monday test document.
en-he
הוא מודגש ארבע פעמים.
It is highlighted four times.
en-he
כאשר אחד הסשנים מתרחש, אנחנו צריכים שמות וטרמינולוגיה לסשן הזה.
So, when some session is happening, we need the names and terminology for that session.
en-he
ואנחנו צריכים לאסוף אותם, להכין אותם באופן ידני וליצור את זה איכשהו.
And we need to collect it, prepare it manually, create it somewhat.
en-he
וההכנה הידנית הזאת צריכה להיות נתמכת באמצעות כלים אוטומטיים עד כמה שניתן.
And this manual creation should be supported with automatic tools as much as possible.
en-he
כך שדרושים כישורים מסוג ב', שצריך לתרגל.
So, there is a certain skill b in that, which needs to be practiced.
en-he
אז אני יודע לערבב קובצי טקסט.
So, I'm quite skilled at shuffling text files.
en-he
וכל פעם שאני רואה מישהו מכם עושה את זה, תמיד יש לי טיפים איך לעשות את זה מהר יותר.
And whenever I see any of any of you doing that, then I like always have tips in my head of what could be done faster.
en-he
אולי זה לא מהר יותר בצד שלכם אבל אתם לפחות יכולים לשקול את זה.
Maybe it is not faster for your end, but at least you should consider it.
en-he
אז זה כישרון שאנחנו צריכים שיהיה לאנשים.
So there this skill is something that that we need people to have.
en-he
ואנחנו צריכים שמישהו יהיה אחראי על זה.
And we need someone to be responsible for that.
en-he
ואני חושש שהאדם היחיד לעניין הזה יכול להיות [אדם8].
And I'm afraid that the only person for this could be [PERSON8].
en-he
אם תמצאו מישהו אחר שיהיה מוכן לעזור עם האדפטציה המיידית של הדומיין, מחיקת הנתונים, יש לומר את זה.
If you find anyone else who would be ready to help with the immediate domain adaptation, the data crunching, please say so.
en-he
ואז אנחנו צריכים –
And then we need -
en-he
ברגע שנבטיח את מילון המונחים ואת הביטוי של המילים, נצטרך טכניקות להשמיש את המילונים האלה במערכות.
So, once we have secured the dictionary of terms and whatever the word pronunciations, we need techniques to put these dictionaries to use in the systems.
en-he
כך שהרושם שלי מהאדפטציה של הדומיין שמה ש-[אדם9] עשה בזהירות עבור כל הסשנים לא נראה ב-ASR ההיברידי.
So, my impression from the domain adaptation that [PERSON9] has been carefully doing for all the sessions was that it was not really visible in the hybrid ASR.
en-he
אז אחד הסשנים יתקיים ביום שני הזה.
So, one such session is again going to happen this Monday.
en-he
[אדם9] כבר התחיל את איסוף הנתונים, אבל אני רוצה להיות את היתרון של האדפטציה של הדומיין בהגדרה של [פרויקט5].
[PERSON9] is already starting the data collection, but I would like to see the benefit of that domain adaptation in in in the [PROJECT5] set up.
en-he
אז אולי [אדם9] ו-[אדם8], כי הוא לומד איך לעשות אדפטציה של דומיין ן-[אדם14], כי הוא לומד איך לעבוד עם [פרויקט5].
So, maybe [PERSON9] and [PERSON8], because he's learning how to do domain adaptation and [PERSON14], because he is doing how to work with [PROJECT5].
en-he
אם שלושתכם תוכלו להיפגש ולוודא מה [פרויקט5] עושה עם ה-
If you three could meet and double check what is [PROJECT5] doing with the -
en-he
זה לא [פרויקט5] בעצם, האדפטציה של הדומיין היא עבור סל הכלים שך [אדם18], כן.
It's not [PROJECT5], actually, the domain adaptation is for the [PERSON18] toolkit, right.
en-he
אז זה בעצם מבט לאחור, כדי לוודא שהגישה הקודמת עובדת.
So, this is kind of backward-looking, making sure that the old approach works.
en-he
כרגע זה לגמרי בלתי אפשרי לעשות אדפטציה לדומיין כלשהו ל-ASR שהוא לגמרי עצבי.
At the moment it is absolutely impossible to do any domain adaptation for the fully neural ASR.
en-he
אז אני שוקל איתור מילות מפתח עצמאי מצליל והליך מיזוג כלשהו.
So, what I'm considering is to have an independent keyword spotting from sound and some merging procedure.
en-he
כך, יכולנו שיהיו לנו שני ASR שיפעלו באותו זמן.
So, we could have two ASRs running at the same time.
en-he
ASR מקצה לקצה, שזה טוב יותר באופן כללי.
End-to-end ASR, which is better in general.
en-he
ואז הגדרה של [פרויקט5] מותאם-דומיין, שמשמש רק לאיתור מילות מפתח.
And then domain-adapted [PROJECT5] setup, which is used only to spot the keywords.
en-he
וכשנראה מילת מפתח בגרסה מותאמת-דומיין, נשתמש במשפט הזה מ-[פרויקט5], שבאופן כללי הוא גרוע יותר אבל מכיל את המונחים הנכונים.
And when we see a keyword in the domain-adapted version then we would then we would use that sentence from [PROJECT5], which is in general worse, but contains the right terms.
en-he
זו ההצעה שלי לגבי מה שאנחנו יכולים לעשות.
So that's my suggestion of what we could do.
en-he
והצעה נוספת היא שכדאי שיהיה לנו ASR עצבי מלא ושנעשה ניסויים על כוונון וכו'.
And another suggestion is that we really should have our own fully neural ASR and do various experiments on finetuning and all that.
en-he
אז דיברנו על זה עם [אדם14].
So, we have discussed this with [PERSON14].
en-he
ו-[אדם14], האם יש עדכונים מעמית פוטנציאלי או מחבר שלך?
And, [PERSON14], is there any update from the potential colleague or friend of yours?
en-he
(אדם17) כן, אז אם יש מישהו אחר שסקרן לגבי העניין, הודיעו לי או צרו קשר.
(PERSON17) Yeah, so if there is anyone else who'd be curious about this please let me know or get in touch.
en-he
אז זה יתקבל בצורה טובה באופן כללי כמסמך כי אנשים עדיין לא עושים את זה.
So, it's this is something which is which would really be accepted well in generally as a as a paper, because people don't do that yet.
en-he
וזו הבעיה הדחופה ביותר בימים אלה.
And that's the most urgent problem these days.
en-he
אנחנו באמת יכולים להשפיע כאן.
So, we really could make an impact there.
en-he
(אדם6) כן, אני חשבתי אולי על איזה סוג נתונים אנחנו כרגע משתמשים עבור זה, כי, לדוגמה, אם תבדקו את מודלים של [ארגון4] על [ארגון5], הם כבר די טובים בדברים האלה -
(PERSON6) Yeah, I was maybe just thinking about like what kinds of data we currently use for this, because, for example, like if you check like [ORGANIZATION4]'s models on [ORGANIZATION5] that they are like already pretty good at these things -
en-he
(אדם6) ואני חושב שאם אנחנו רק -
(PERSON6) And I think that if we just like -
en-he
אנחנו יכולים רק לקבל את הנתונים מהם, כי יש לך סט גדול של סרטונים למעשה עם דומיינים שונים ושפות אם שונות של דוברים ב-[ארגון5].
We could probably just get the data from them, because you have like a such a large set of videos basically with different domains and different speaker native languages on [ORGANIZATION5].
en-he
והיה לי רעיון שנוכל להשתמש בכלים כלשהם כדי להוריד למעשה סוגים של סרטונים מסוננים מ-[ארגון5] ולהכין סט הדרכה מהם.
And I had this idea that we could just like use some tool to download basically some kinds of filtered videos from [ORGANIZATION5] and make training tests a training set out of them.
en-he
אז אני מתכנן לעשות את זה בעצמי.
So, I plan to do it myself.
en-he
אני מתכנן ללכת לשם אבל אנחנו צריכים מישהו לגיבוי.
I plan to go there, but we need a backup person.
en-he
וגם, עבור האירוע של [פרויקט210], אנחנו צריכים מישהו שיטפל בהקלטה,
And also, for the [PROJEC210] event we need someone to take care of the recording,
en-he
כך שאיש הארכיון, כך שלא נאבד את הסשן שוב.
So, the archiver person, so that we don't lose that session again.
en-he
ונושא חשוב יותר שקיים עדיין ושצריך לעבוד עליו הוא הטיפוס בסולם.
And the more important and still persisting topic to work on is ladder climbing.
en-he
והצעתי כאן בפסקה עבור [אדם2] שנוכל להשתמש בטבלאות html פשוטות שיווצרו באופן אוטומטי וזה יגדל.
And I've suggested here in the paragraph for [PERSON2] that we could use simple html tables that would be autogenerated and that that would grow.
en-he
אני חושב שזה יהיה כנראה הכי קל פשוט לשים את הטבלאות בפורמט של html אבל זה יכול להיות קשה כשנרצה לראות את ההבדלים, כי מערכות חדשות יעברו הערכה אחרי כל קמפיין הערכה חדש.
I think that that would be probably the easiest to just put up the tables as html, but it could get difficult when we want to see the differences, as new systems will be evaluated after every new evaluation campaign.
en-he
יכול להיות שיהיה קשה להכין את הטבלאות האלה.
It may be difficult to prepare these tables.
en-he
אז ביקשתי מ-[אדם7] ן-[אדם7] לצערי לא יהיה זמין במהלך השיחה הזו.
So, I asked [PERSON7] and [PERSON7] unfortunately won't be available during this call.
en-he
להגיד לנו אם [ארגון242], שהוא פתרון מבוסס-אינטרנט.
To tell us whether [ORGANISATION242], which is a web-based solution.
en-he
חברת סטארט-אפ כלשהי שמציעה חשבונות למטרות אקדמיות בחינם.
Some start-up company which offers accounts for academic purposes for free.
en-he
אם זה יעזור לנו או לא.
If that would help us or not.
en-he
אוקיי, אז [ארגון242] לא מומלץ על ידי [אדם7].
Okay, so [ORGANISATION242] is not recommended by [PERSON7].
en-he
אז html פשוט זו כנראה הדרך.
So, simple html is probably the way to go.
en-he
אז, ברגע ש[אדם2] יסיים את ההערכות האלה.
So, as soon as [PERSON2] has these evaluations m cleaned up.
en-he
זה אומר להעביר כמה ספקריפטים למקום שאליו הם שייכים מסט הבחינה של [פרויקט3] ואיזהרץ אוטומטי.
So, that means moving the few scripts to where they belong from [PROJECT3] test set and some automatic runner.
en-he
אז צריך גם ליצור המרה מהטבלאות ל-html בדומה להמרה הטכנית ואז לשים את התוצאות האלה באופן קבוע בדף האינטרנט.
Then also please create a conversion from the tables into html similar to the tech conversion and then put these results regularly to your webpage.
en-he
או מה שאתה יכול לעשות הוא לעשות בדיקת git שממוקמת ישירות ב-html הציבורי או החיווי מה-html הציבורי הזה.
Or what you could even do is to have a full git check out directly located in this public html or singling from that public html.
en-he
כך שזה יהיה נגיש מיד באמצעות האינטרנט וכל אחד יוכל להסתכל על הניקוד הקיים ולחקור אותו.
So that it would be immediately accessible through the web, and anybody could have a look at the current scorings and explore that.
en-he
אז זה יהיה שימושי, במיוחד עבור אנשים מחוץ ל-[ארגון2] כמו, [ארגון15].
So, this would be useful, especially for m people outside of [ORGANISATION2] such as, [ORGANISATION15].
en-he
אנחנו יודעים שיש להם -
We know they have like -
en-he
הביצוע היה רע עבור כמה מהקבצים, והם יכלו לעיין בהם ישירות באינטרנט.
The performance was bad for some of the files, and they could directly m browse it on the web m.
en-he
הם יכלו למעשה לעיין בהם בעצם ב-[ארגון101].
They could also browse it actually in on [ORGANISATION101].
en-he
זה תלוי בך להגדיר את זה כך שזה יהיה ידידותי למשתמש וקל לעקוב.
This is up to you to set it up so that it's user-friendly and easy to follow.
en-he