Thursday 14 October 2004

Pic of the day: "Let's start with pronunciation". Screenshot from the anime DearS. (The indecent teacher is a not uncommon stereotype in Japanese anime, but this software is not Oriental despite its name and despite transcribing like an Oriental exchange student with hangover and homesickness. The student has improved steadily though, and is now almost sober.)

Dragon NaturallySpeaking 7.3

Today I downloaded the free upgrade for my speech recognition software. Last year, in time for NaNoWriMo, I had bought NaturallySpeaking 7.0 from ScanSoft.com. It was much better than version 5, which I had used the year before. But input still required quite a bit of editing. This fall I realized that a bad microphone was part of the problem, and got a Plantronics USB headset with a noise filtering microphone. Once again, things improved noticeably. But it was still not perfect. And when I read that someone would use version 7.3 for NaNoWriMo, I got curious and checked it out.

It turned out that there was a free upgrade from 7.0 to 7.3. It is basically a large bug fix. Most of the bugs were not relevant to me, but at least two were. One was a tendency to get small words wrong at the end of a dictated statement. (Actually the beginning and the end of the statement both seem to be vulnerable. I suspect this is because we humans pronounce words differently at the beginning and at the end of a statement. This is probably some kind of signal to our listeners, kind of like punctuation, to let them know that we are beginning to speak or that we are about to end a statement. But the computer has problems with it.) The other bug related to the acoustic optimizer.

One of the great things about modern speech recognition is that it learns from experience. If you correct a word that it has misheard, it is less likely to make the same mistake again. There are some words that are simply pronounced the same, and in that case the program needs to guess which is which based on the context. (This is why one should speak in complete statements and not one or two words at a time.) As you continue to use the program day in and day out, it accumulates information about what words you tend to use and how you tend to pronounce them. But this is a lot of information, and it just builds up on your hard disk. When you run the acoustic optimizer, the program takes the data it has gathered and uses it to redefine the vocabulary. In effect, it creates a special version of the program just for you. This is a pretty heavy job and could easily take half an hour or more.

In version 7, there was a bug or two in this process. The optimizer did not to delete the files it had used, so they just grew larger and larger over time. It also seems to have inserted deleted words back into the vocabulary. I didn't know this, but I did notice that every time I exited the program, it asked my leave to expand the size of the data files, even after optimizing.

***

The upgrade was easier than I was led to believe from their web site. They asked me to copy the user files to another folder of my choice, so I assumed that I would have to copy it back after the upgrade. This turned out not to be necessary. I downloaded and started the upgrade program, and it worked on its own for a while, then it was ready. I had to restart the computer: The install program offered me the option to exit without restarting the computer, but then the install program itself did not terminate and slowed down everything else, so I ended up rebooting anyway. After that, it has worked like a charm.

The program still makes mistakes. I am sure I don't catch them all. Some of them probably just look like bad English, using the wrong form of a verb or mixing up tiny words like "to" and "the". Some may be funny, and some not make sense at all. Perhaps I will find them next year when I read the year ago entry (if I'm still around). But it is a big improvement over the pain that I used to feel after writing an entry. Every night I can go to bed without the pain is a kind of victory.

And on that note, I'll stop now before my throat really starts to hurt...