Apr 15, 2014

Speech recognition for translators: microphone tips

Guest post by Jim Wardell

Mark Myworts is heavy into a major procrastination project, pawing through moldy old copies of Popular Mechanics in Grandpa’s basement, when a small classified ad in the back pages catches his eye. He blows off the dust:
“Translators! Double your income overnight with amazing new technology! Works wonders for English, German, Spanish, French, Italian, Dutch. No obligation. Call ... Full confidentiality guaranteed.”
[Fade in “Twilight Zone” theme music.]

[Cut to Mark talking intently to his computer.] “... should be used instead of the diminutive terms ‘pud’ and ‘loser’” ...

[Fade to Mark and kiddies.] “Well daddy, can we? can we? Can we go to the circus tonight?” “Sure kids, I’m knocking off early today,” says Mark nonchalantly, getting a kiss and one of those sexy “well-what-about-after-the circus” looks from his admiring wife.

Science fiction? 1950s social mythology? Perhaps.

But the simple fact remains – at least now in 2014 – that translators into English, German, Spanish, French, Italian and Dutch really can double their productivity on average using some amazing, although not quite so new technology: speech recognition.


***

I’ve been using speech recognition to translate from German into English for nearly 20 years. But it was not until about seven or eight years ago that computing power and speech recognition software had improved to the point where serious productivity gains became possible. It was at that point that it became imperative for me to find a CAT tool that was totally compatible with Dragon NaturallySpeaking. At the time, only two products met this requirement: Déjà Vu and memoQ. For various reasons, which I won’t get into now, I decided to go with memoQ, a decision I have never once regretted.

Most any CAT tool can be made to work with Dragon by using the little “Dictation Box” text buffer that’s provided as a workaround in Dragon for software that is not truly Dragon compliant. The procedure needed to translate average moderately sophisticated technical documents in noncompliant CAT tools can often be cumbersome and inefficient: one copies the contents of the current source segment text into the Dictation Box so any strings that do not need to be translated can be left as is or moved around as desired and so that sections of text that need to be translated can be overwritten by marking them and then dictating the new translation “over the top”. Once the source segment has been duly massaged in the Dictation Box, the contents of the box are then transferred to the target box in the noncompliant CAT tool. Of course, various tags and formatting that might have been present in the source segment are often lost when source text pasted into the Dictation Box. So they need to be put back in again after the contents of the Dictation Box have been pasted into the CAT target box. If this sounds gruesome, it is.

So why not just dictate straight into the noncompliant target box and fix the messes as they occur? The answer is simple: there are often too many messes, and, worse still, any incorrect speech recognition that occurs cannot be corrected in a way that will ensure that correct and not erroneous data will be fed and saved in the Dragon speech recognition engine. Over time, this would degrade speech recognition accuracy! I spent some years trying to address this issue with publishers of CAT tools other than memoQ and Déjà Vu ... with zero success. So if you’re not already using Dragon and want to use it with a CAT tool, make sure that the CAT tool that you are thinking of using is really fully compatible with Dragon and that you can get your money back if it’s not. Do not trust and do verify.[1]

At one point before I switched memoQ, I was compelled to do a good bit of this acrobatics moving text into an out of the Dragon Dictation Box. I began to have the feeling that the cutting edge that I was working on in a number of well-known CAT tools was so dull that I might just as well have been typing in my translation in the old-fashioned way. So I collected some statistics discovered that that was indeed the case. My output was the same in noncompliant CAT tools and Dragon as with touch typing without Dragon.

All that changed with the memoQ’s full Dragon compatibility! Incidentally, memoQ has full Dragon compatibility throughout the interface and not just in the translation grid. So if you want to dictate notes or definitions in term base entries, you can, and you can still use all of the selection and correction features you are accustomed to using in Dragon. Want to write a longish note to a client in a memoQ Comment box? No problem. Dictate away.

* * *

Anyone who has dealt with integrated technologies, any process in fact, knows that the old saying “A chain is only as strong as its weakest link” is totally true. So to get really great speech recognition results, not only does one’s CAT tool have to be compatible and outstandingly good, one’s computer needs to be sufficiently powerful, and one needs to use the best available microphones. I heartily recommend KnowBrainer.com as a source of top quality microphones for speech recognition. To my knowledge, KnowBrainer is the expert in the USA, probably the world, when it comes to speech recognition products. People who want to achieve maximum accuracy with speech recognition software should make their first stop KnowBrainer’s Microphone Comparison page.[2] For many years, I used KnowBrainer’s top-rated Samson Airline 77 microphone. This microphone was vastly superior to anything I had ever used in the past and came delightfully close to delivering 100% speech recognition accuracy. Earlier this year, however, I learned that the wireless channel used in my old Airline 77, which I bought while I was still located in the United States, was being shifted to use by mobile phones in Europe and would no longer be legal. So I checked out KnowBrainer again, and learned about a relatively new microphone being produced specifically for speech recognition by a Belgian company: SpeechWare. Upon consulting with KnowBrainer’s Lunis Orcutt (Mr. Speech Recognition in my book!), I ordered the SpeechWare 3-1 TableMike. This desktop mic is a great product and just as good as my old Airline 77. It’s the mic that Lunis himself uses.

However, after using it for a week or so, I realized that it was not for me because I had to keep my mouth relatively close to the microphone and couldn’t move around like I was used to in the past in order to relax my back muscles and stay fresh. I then ordered a FlexyMike headset mic from SpeechWare that basically uses the same technology but allows one to move around freely. SpeechWare has three models of the FlexyMike: the FlexyMike Basic (FMK01), the Single Ear (SE) and the Dual Ear. I chose the Dual Ear on the principle that distributing the weight of the mic over two ears would be more comfortable and stable for hours and hours of continuous use.

When I was still using the tabletop TableMike, I found that I had a tendency to move a little too far away from the microphone over time, which occasionally reduced speech recognition performance. The TableMike has two settings: a long-range setting, which allows one to have one’s mouth as far as 30 cm (12 inches) from the microphone, and a “normal and VoIP” setting (with a maximum distance of 15 cm / 6 inches). The greatest accuracy is achieved with the closer distance. Lunis says he likes the TableMike because he moves around in the office a lot and doesn’t need to fumble around with headset whenever he leaves his desk. For this reason, I would recommend the TableMike as the best choice for project managers and administrators who may frequently have to leave their desks and who mainly use Dragon at brief stretches to dictate e-mail messages or enter data in translation business management software. For hard-core translation work, the FlexyMike is the way to go. I find the accuracy with the FlexyMike to be perhaps a tad better than that of the Airline 77, which is saying a lot. I can use the FlexyMike while listening to the radio a moderate volume levels, so the noise cancellation is also quite good. KnowBrainer gives its noise cancellation a score of 9, which is better than that of the Sennheiser ME3 KB headset mic (gets an 8), which I have used with good success for years in automobiles, trains and airplanes! All the same, if anyone has to dictate in an extremely noisy environment, one might want to check out “theBoom v4 KB”, which gets a high accuracy rating and a 10 for noise cancellation (but only a 9 for comfort!) from KnowBrainer. My experience is that KnowBrainer is pretty fanatical about these evaluations and that they are quite reliable.

For the average translator, who works long hours in a relatively quiet environment, accuracy and comfort are the two most important factors, more important than noise cancellation. I don’t need speakers on my headset, which means that the headset can be as light as a feather and can be worn comfortably all day long. If need be, Skype calls or music can be played through normal computer speakers. On the other hand, if one is working in an open office setting with a number of other translators close by, one might want to have a headset with speakers covering both ears to block out distracting voices so one can concentrate better. In such cases, I’d consider the mono Umevoice “theBoom Pro-2 KB” or the stereo hi-fi equivalent “... 3 KB” if you want to block out room noise and also want to listen to music while you translate. (I translate very complicated, detailed stuff and usually extremely distracting to listen to music while translating, but not all material that gets translated requires extreme concentration. I could also easily imagine listening to a high-bandwidth feed from, say, jazzradio.com premium (unabashed plug) to make routine administrative work more pleasant.

Getting back to the FlexyMikes: SpeechWare was kind enough to also send me a single-ear model to test and evaluate, so I have used both versions extensively. Both the single-ear and the double-ear mics are extremely comfortable and both are very easy to adjust to get a custom fit that is secure and comfortable. The materials used in both mics are of exceptionally high quality and should provide many years of reliable service.

Both FlexyMikes connect to a computer USB port across a “SpeechMatic MultiAdapter”, which has been especially configured for high accuracy with speech recognition. I am convinced that the special design of the MultiAdapter is one of the main reasons why the FlexyMikes work so well.[3] Be sure to buy this along with your FlexyMike. The same circuitry that’s in the MultiAdapter is integrated into the TableMike units. So if you already have a TableMike, you don’t need to buy a MultiAdapter, unless of course you want something really small and light to use with a notebook computer when traveling.

I did not test the basic version of the FlexyMike, from the pictures it didn’t look as comfortable as the other models.

KnowBrainer.com ships internationally. SpeechWare microphones are also available directly from SpeechWare in Europe.

[1] If you want to see what “fully compatible” means, have a look at http://kilgray.com/news/once-upon-time-there-was-dragon.
[2] http://www.knowbrainer.com/core/pages/miccompare.cfm
[3] So is KnowBrainer: See http://www.knowbrainer.com/NewStore/pc/viewPrd.asp?idproduct=464



***


Jim Wardell will be presenting optimized work methods for speech recognition once again at this year's memoQfest in Budapest, Hungary.



Apr 14, 2014

Legal English: Getting It Right!

That's not the title of the upcoming workshop series by attorney and linguistic specialist Stuart Bugg, but perhaps it ought to be. Like the recent successful business communication conference in Cambridge, these Stridonium courses for English Contracts, Legal Drafting and Commercial Law once again bring together a recognized subject authority with those involved in practice with the topics, both working attorneys and professional linguists. Mr. Bugg, author of Contracts in English: an introductory guide to understanding, using and developing 'Anglo-American' style contracts (C.H. Beck) and co-author of Langenscheidt Fachwörterbuch Kompakt Recht Englisch, has been associated with legal societies in four countries (NZ, AU, UK and DE) and has taught at universities and law school in New Zealand, Germany and the United States. He is one of the "go-to guys" in Europe training attorneys as well as government and industry professionals to understand legal English better and improve their use of it in international commercial practice. I cannot say how pleased I am that the Stridonium organizers have once more raised the bar and gone beyond the "usual suspects" of the translation circuit to connect more effectively with real experts who are at home in the world of our direct clients and offer first-class continuing education for serious professionals.

These workshops are an excellent opportunity for those who write and translate legal English to learn and discuss best practice, common pitfalls based on linguistic issues as well as differences in legal systems, and how to apply the principles of good, professional language for unambiguous, legally sound communication.

The relaxing venue in Holten, in the east of the Netherlands near the German border with excellent road and rail connections, is a perfect place for formal and informal discussions and exchange of ideas and is noted for its beautiful setting in the woods at the edge of town and its outstanding cuisine. Each workshop is scheduled on a Monday, and early registrants can usually secure a comfortable room at no charge as part of the event registration (subject to availability).

Attendees can register for individual workshops on the Stridonium events page; the fee for each full day of instruction is €350. But if you plan to attend all three workshops, a special rate can be obtained by inquiry to the event organizers at info (at) stridonium.com.

The past few decades have seen a dramatic increase in the share of European law contracts relying on English for their binding version. The result is ever-greater pressure on attorneys, their clients, translators and interpreters to understand and use legal English correctly and to understand how best to explain the legal principles of one country in the language of another.

The schedule of workshops is as follows:
English Contracts - April 28, 2014
    - Common Law vs. Civil Law
    - Cross-System Contracts
    - Legal English Terminology
    - Lost and Found in Translation

Legal Drafting - May 26, 2014
    - Basic Drafting Principles
    - Legal Terms
    - Principles of Drafting and Interpretation
    - Avoiding Ambiguity: Exercises in Drafting

Commercial Law - June 2, 2014
    - Overview
    - Legal Entities
    - Employment Law
    - Bankruptcy and Insolvency
There is a networking dinner each Sunday evening before the Monday workshop for early arrivals. Details of each workshop's topics and schedule as well as registration links are on the Stridonium events page.

Each workshop has been awarded 6 CPD points by the Dutch Bureau BTV and 6 ATA CE points.


How to get there (Google map link):

By train
- A 10-minute walk from the station (Beukenlaantje)
- OR let the organizers know when you arrive and either they or hotel staff will collect you!

By car (plenty of free parking!)

From Deventer(A1)
Take the A1 towards Hengelo/Enschede
Exit 26: Lochem/Holten
Turn left for Raalte, follow the signs for Holterberg
Go straight ahead over the roundabout, turn right after the viaduct and left at the T-junction
Turn left at the roundabout and after 50 m take a right turn for Holterberg
After approx 1 km turn right (at yellow building)
From Enschede/Hengelo (A1)
A1 towards Deventer/Apeldoorn/Amsterdam
Exit 27: Holten/Markelo
Continue through the center of Holten, take the Holterberg exit at the roundabout and after 50 m take a right turn for Holterberg
After approx 1 km turn right (at yellow building)




Apr 11, 2014

memoQ: stopwords for term extraction

The recent Kilgray blog post about the "terminology as a service" (TaaS) project reminded me of the considerable unfinished business with the term extraction extraction module introduced three years ago with memoQ 5.0. It's a very useful feature that I apply frequently to my projects, and my prediction years ago that it would not replace SDL's MultiTerm Extract in my workflows was wrong. Overall it proved to be more convenient, and after the shock of discovering that the defective logic of MultiTerm Extract created new German "words" that neither existed nor were in my text sources, I dumped that dodgy tool and stuck to memoQ's extractor. But sometimes its rough edges are irritating, and I wish Kilgray would finally pick up the ball that was dropped after a great start in the game.

One of the major weaknesses (aside from never remembering my changes to the options or my preferred settings for extractions) is the management of stopword lists.

Overall, Kilgray's approach to stopwords is reasonably sophisticated; some time ago I published a rather incomprehensible post in which I demonstrated how the "stopword codes" - those three digit binary numbers which appear stopwords - control whether a word can appear at the start, the end or the middle of a phrase even if it is excluded as a single word. These codes are quite useful in some cases. However, they also complicate the use of stopword data for most users.

memoQ includes only a few stopword lists for a few languages in its shipping configuration - German, English, French and Italian I think. Not even all the user interface languages are included. That's rather sad, because there are quite a few public domain stopword lists available on the Internet. However, most memoQ users have absolutely no idea how to incorporate these in memoQ, and Kilgray offers no features or information that I am aware of to facilitate this process.

I was reminded of this problem when I was asked to discuss terminology mining with masters students at my local university in Portugal. I thought it might be nice for the students to be able to make use of the stopword lists they could find on the Internet for their target languages (Spanish and Portuguese for that group). These lists are typically just text files of single words. When I build my own master stopword list for German a few years ago, I gathered half a dozen or more large, mostly redundant lists for a start. Then I carried out the following steps:
  1. Combine all the stopword lists for a language from various sources into one big text file.
  2. Open that text file in a spreadsheet program such as Microsoft Excel.
  3. Sort the list and use the integrated function to eliminate duplicates.
  4. Fill the number "111" in the second column of the spreadsheet. (This will completely exclude the term from phrases as well; if you want to make individual exceptions according to the scheme I described in an earlier blog post, you can do so now or at any time later after the list has been imported to memoQ.)
  5. Save the data as tab-delimited Unicode text.
  6. Open the text file and paste in this XML header, adapting the red parts to your particular list:

    <MemoQResource ResourceType="Stopwords" Version="1.0">
      <Resource>
        <Guid>dc7006ad-8db8-4724-b22d-7acfd600fd9f</Guid>
        <FileName>ger#KSL_stopwords-DE.mqres</FileName>
        <Name>KSL_stopwords-DE</Name>
        <Description>Combined lists from many sources</Description>
        <Language>ger</Language>
      </Resource>
    </MemoQResource>


  7. Save the file, change the file extension to MQRES, and import the file as a stopword list in the memoQ Resource Console.
Afterward, the stopword list you imported will be available  when you start a term mining session using Operations > Extract Terms...

During an extraction session, words can be added to the chosen stopword list (one at a time unfortunately - I've been asking for multiple addition as a productivity measure for 3 years now so far) by selecting a term and Clicking the Add as stopword command or pressing Ctrl+W.

You might be a little confused if you look for words you've added to a stopword list from the term extraction interface. They are not inserted in alphabetical order, but instead at the end of the words starting with a given letter. Thus, for example, the red box in the screen clipping below shows all the words I've added to my previously alphabetized list since I created it: