Older blog entries for dhd (starting at number 102)

Well, I've just gotten Flite's unit selection up to roughly the same quality as Festival's, by filling in a few missing pieces in the join cost estimation and fixing the Viterbi decoder. The scoring, and the actual choice of units, is a bit different, but it sounds pretty good regardless.

It's a bit big though ... well, it's huge actually (so "small, fast" is not entirely accurate when it comes to unit selection voices :-) With 15,000 units and 150,000 frames of LPC plus residual data in a small unit selection database, one would expect it to be. This means that the really fun part of my task is just beginning.

Hanging out with friends in Ottawa, YAPC, B.C., Calgary, Ottawa, back home, sleep-deprived stupor, work, etc.

YAPC was pretty fun. As usual I didn't get enough sleep. My talk didn't go nearly as well as I'd hoped. I need to bite the bullet and actually practice these things ahead of time. I did, however, control my slides with the power of my voice!

I am working on really fun stuff now, namely converting cluster-based unit selection databases from Festival to work with Flite. I'm still managing to avoid writing any Scheme code, and have ended up using Perl to parse Scheme data and write out C. It works really well, so I don't care...

There's a lot of work to be done in shrinking and optimizing databases which should keep my busy for a while.

As a consequence, I'm finally really getting acquainted with the internals of the TTS engine. It's amazing how much more straightforward some things become when they're expressed as well-written code. Of course, I may not have been able to understand the code as well without the prior explanations.

It's also deceptive because some of the algorithms are very simple in implementation and "just work", but rely on a frightening amount of signal-processing-fu or statistics-fu to explain why exactly they just work. Unfortunately there's a tendency in some papers and books I've read to throw down the big scary math *first* and then eventually derive the practical implementation.

Well, I was hoping to ride at least 50 miles today, but I had to get some work done in the morning and by the time I had finished being lazy around the house, it was 4PM, so I had to settle for a slightly extended version of my daily 15-mile loop (10th -> jail trail -> greenfield -> beechwood -> fifth -> neville -> boundary -> jail trail -> 10th) ... this time taking a detour through Hazelwood and finding a more formidable hill to climb on the way out. The hills around here never cease to amaze (and exhaust) me; guess I'll just have to keep going up them until that's no longer true!

Having installed some voice-building machines and discovering the inherent suck-factor of all existing command-line audio recording tools, I just had to write yet another one, which I managed to do on the spot in about 30 minutes. Now if I could just get around to fixing POE::Component::Audio to do multiple channels and other trivial things, I'd feel better.

Stepping on the scale this morning I noted that, after giving myself an extra week of reduced eating to make sure that I was really actually under 150lbs, I was, well, really actually )well) under 150lbs, down from 180lbs three months ago and probably more before that... Hopefully being able to eat a bit more often will improve my demeanor and mental clarity... Now I just have to get some pants that fit.

Fun with Sphinx acoustic training. Boy, I wish I actually understood the math behind all this stuff. I guess that's what graduate school is for (most likely that's where I'll be in a couple years time).

For now, though, I'm going to enjoy some hacking on the implementation of it - the process (though it is inherently massively CPU-intensive) is taking a lot more time than it should and there are a lot of code cleanups to be done.

Computers are not really a hobby for me anymore. This is unquestionably a good thing in terms of my mental health, though I worry that I am not doing enough to maintain and improve my technical skills.

It was only a few years ago that I was a young liberal-arts student and self-taught hacker living off of student loans and struggling for meaningful employment, and I fear that, should things go wrong, my lack of formal education and credentials will come back to haunt me. I guess I can always go back to school.

In reality, things are actually going very well indeed for me; I work for a company with sensible management and finances, doing interesting work with real applications and customers. It's just that I have learned to survive by planning for the worst case scenario.

I'm worried about the perception that free software doesn't pay, and about the possibility that the demand for Linux/GNU/Perl/Apache expertise in the job market may be evaporating. I really hope that this is not the case, if not for my sake, then for the sake of people who might be in the same place I was a few years back; bright kids learning and using free software on their own time, who should have their efforts rewarded rather than repeatedly running into the "no BSCS degree, no job" barrier.

On the other hand, I am certainly glad that there will be no more bogus companies burning cash like Dubya, Cheney, and their oil buddies want us to burn oil, nor instant millionares and lazy people expecting huge payoffs. And speaking as a Canadian ex-pat, I'd be happy to see the outrageous salaries paid in the US be "corrected", so that perhaps the rest of the world will get a chance to keep its best and brightest.

Anyway, as I was saying, I've been splitting my spare time between my old and neglected interests of cooking, cycling, and homebrewing, and if you'll excuse me, I've got some bread to bake.

I have volunteered myself for a million little projects with unspecified goals and deadlines, and am consequently getting very little actual work done on any of them. I miss my old project manager...

On the bright side I'm finally fiddling with the voice building tools in greater depth, since we're actually, like, building voices now. I went into the studio to record some unit selection prompts and I'm slowly working on a French diphone voice (if I ever get the diphone list generation and letter-to-sound stuff done...)

Arr. Once again I am in computer telephony hell. In the hopes of achieving reliable echo cancellation and full-duplex performance, we acquired a very expensive Dialogic PCI telephony card and the package of (stupid, proprietary, gross) Linux software support. Well, it appears that the software support ... doesn't. And the documentation ... doesn't.

And as far as I can tell, Dialogic doesn't actually want to support people who try to use it. Their customer support e-mail directs you to a web forum or a pay-per-incident service. After paying $BIGNUM for this piece of crap board and proprietary crap software, the least they could do is write some bloody documentation for it that is good for more than toilet paper and give at least some complimentary e-mail support. Fuck you, Dialogic.

Surveying the landscape of computer telephony fills me with a deep feeling of helplessness and despair. All the software is broken, all the hardware is obscure as hell, it all costs unbelievable amounts of money, and it never seems to work correctly. And when it looks like I find a board that is actually useful for my applications, it turns out that it only works on Windows NT. (come on, people, you could at least support Solaris or some kind of semi-reasonable proprietary OS).

It looks like the Quicknet/IXJ stuff, as quirky as it may be, is the best thing there is for Linux and open source telephony at the moment. I hereby take back all the bad things I said about them :-)


Oh yeah! That's right! We released some speech stuff a few days ago. Please beat on it, etc. I'm mostly working on higher-level things at the moment (i.e. actual applications that use these modules) so these should be stable for a while.

In other news, Sphinx2 0.3 was put out with little fanfare. So you don't actually need to use the CVS version to compile Speech::Recognizer::SPX. Which reminds me, I should probably update the README.

I've already been asked if this stuff works on Windows, I wonder when someone will ask if it will be rewritten in Python...

I'm playing at rewriting the silence filtering library. Handling all the audio buffering and framing safely and efficiently is unfortunately trickier than it looks. Pointer arithmetic is HARD, let's go SHOPPING!

Playing with the iPaq. Wrote some nice feature extraction code for FPU-less machines. Somewhat unsurprisingly, it's three times slower on my P3 laptop. (but 80 times faster on the iPaq)

Discovered a nasty bug in the iPaq audio driver, leading to a kernel oops. Predictably, select(2) was broken. Sent off patches. No reply. You'd think that a patch that changes a grand total of 4 lines, and fixes an oops, would get a bit of priority in people's mailboxes. You'd think wrong.

Oh the PLE133 is really lovely. So, to reduce costs, it seems that VIA decided it would be clever for it to only support PC133 memory, and not even all PC133 memory (it has to be CAS3).

I discovered this since I finally broke down and got a better motherboard for my home box and thought it prudent to reuse the old parts for a firewall box at work ... several sticks of SDRAM later and I finally have a working machine built around this thing which seems to be way too overpowered to serve as a firewall. <sigh>

What else ... well, various stuff. Having more fun with audio. I was thinking of presenting on speech and Perl at YAPC but I might just do audio and Perl as there is more than enough there to take up 45 minutes.

It really seems as if you cannot win with Linux audio drivers. If they don't support setting the fragment size, then you throttle with SNDCTL_DSP_GETOPTR (which is possibly a better idea anyway). But then you discover that lots of drivers don't support that either. Does anyone actually use this stuff for anything besides playing MP3s and Quake?

93 older entries...

New Advogato Features

New HTML Parser: The long-awaited libxml2 based HTML parser code is live. It needs further work but already handles most markup better than the original parser.

Keep up with the latest Advogato features by reading the Advogato status blog.

If you're a C programmer with some spare time, take a look at the mod_virgule project page and help us with one of the tasks on the ToDo list!