xiphmont

Jean-Marc Valin has been applying deep learning frameworks to audio over the past few years. So far he's released RNNoise (a surprisingly good/fast denoising system) and LPCNet (a speech synthesis system along the lines of WaveNet, but fast enough to use realtime on commodity hardware).

Now he's built a codec out of LPCNet.

"A Real-Time Wideband Neural Vocoder at 1.6 kb/s Using LPCNet" presents a new wideband speech codec built out of the best parts of a brutally speed and space efficient vocoder paired with deep-learning analysis and excitation. It's alpha-grade research in a lot of ways, but decidedly not vapourware. You can download the source and play with it now, but first, go have a look at the demo page.

The end result of more playing around with the plastic LED collimaters: they're not going to work well. Dang.

The microscope wants a small diameter parallel light beam, and the plastic collimaters just won't do that. Even the tight spots overfocus and overmagnify, and there's no real room for adjustment.

Noq2's approach works because of:

a) brute force: 60W of LED light is the equivalent of around 350W of halogen light

b) a large die LED (7mm diameter!): that's more like a small COB

In his case it's a 'close enough' approximation to infinity focus because of the huge die and tons of output. There's enough nearly parallel light in there along with all the rest to make it work. Most of the light that makes it into the optical path (which is likely only 5-10%, the beamsplitter doesn't combine much) is just lighting up the inside of the microscope body, but the light that does make it all the way through is diffuse and even and nice. Brute force works!

Me, I want as much of the light to be usable as I can get.

So I'm probably back to using lenses. The stock Olympus setup uses two air-gapped elements with a total focal distance of 11-12mm placed on either side of a halogen bulb. The bulb filament intentionally sits just in front of the focal point to defocus the image slightly. (EDIT: Actually, I'm wrong here; it's just past the focal point in order to produce a focused image of the bulb filament at the point of the iris in the microscope body).

Olympus uses some really nice glass. The first condenser element is a partial shortpass to filter out some of the infrared. The second weaker element nearly touches the first and finishes the collimation job. Two 45 degree front-surface (!) mirrors direct the light into the beam combiner body.

The biggest constraint on the collimater design is the diameter of the optical path through the scope, which is a little under 15mm at multiple points. Opposing that, we want to collect as much light as possible from the LED into the condenser, which means putting the lenses as close to the emitter as possible, and so choosing the shortest practical focal length. Focal length trades off against beam width; the shorter the focal length, the more the 'image' of the LED die is magnified and the wider the final collimated light beam.

The original Olympus optical design expands the image of the halogen bulb filament into a beam of approximately 15mm diameter, matching the optical path. The Cree XP-L LEDs I'm using have a die almost the same major dimension as the bulb filament.

A more powerful LED with a bigger die probably isn't useful if we're going to use a single-stage collimator. And an XP-L is easily the highest-flux LED I can get with a die this small.

The XHP70.2 might put out 5x as much light, but if that's over 5x as much area, it's not a net gain given the constraints (I'm going to test it anyway, but I don't have high hopes). The big die of the XHP70.2 isn't a problem for noq2 because he's not using any optics that magnify its apparent size. He could usefully apply a 15mm COB.

This also means we're not going to improve on the original Olympus lens choices without going to a more complex beam reduction design that probably won't fit. (EDIT: Actually, I can probably fit a Keplerian design in there)

I'm going to try it Olympus's way. Given how dang nice those lenses are, I'm totally yoinking them. And since I'm messing so much with the physical layout, I want some adjustment ability. Which means at this point-- I'm most of the way back to my original design. Oh well. At least much of it is actually built and this has become an iterative process. A little at a time rather than one fell swooooooop.

So... The next step is making some new lens tubes.

And mounting the Olympus lenses in them (using nice, reversible, not messy O-rings).

This gives me the lenses at the proper separation in a durable, flexible package. Now I have to make a tube mount that fits inside the lighting enclosure.

Electronics assembly!

But does it, for the lack of a better word, chooch?

Yes, it chooches.

Adding the pot to test the adjustment, it turns out my BuckBlock is sourcing more than 500uA-- it's actually closer to 625uA. Also, it's firing up at ~1.5v rather than ~ 1.75v. That means I want a 2.2k low-end resistor and a lower pot resistance. So I modded another 10k linear pot into a 13k exp-ish pot, and continued assembly.

Everything goes together as intended and looks nice.

But there's a snag; I knew there was a good chance the optics wouldn't play as well with the SZH as I hoped, and in fact, the focus behavior isn't working well once on the scope. The illuminator works, and alignment is spot-on, but it's focusing an image of the lens at working depth, which is suboptimal. Also due to the suboptimal focus, lots of light is getting wasted inside the scope as it scatters out of the parallel-light/infinity focus portion of the optical path.

I'm going to have to play more with it.

Earlier this year, when I was building mods for my Olympus SZH microscope, I planned a coaxial illuminator using LEDs. The design got... a little out of hand...

...and I put it aside.

Last week I stumbled upon noq2's blog, where he documented his LED mod for the same microscope. It's the polar opposite to what I was doing. His is simple, brute force, and gets the job done. Go have a read, it's nicely done (and his animated gifs of the SZH internals are great). And did I mention it's actually finished and working? :-)

This inspired me to reconsider what I was doing.

I mean, I know what I was thinking: a design worthy of the rest of the absurdly overbuilt Olympus. I mean, just look at the original collimator! They use front-surface mirrors in the light source!

Also I didn't want to sacrifice any original Olympus parts. Complete SZH coaxial illuminators still sell for over a grand. But that's mostly because they're usually missing accessories needed to operate them (the bulb holder and transformer) which I didn't need. Last summer I picked up a bare illuminator block for $90. That's cheap enough I'm willing to mod it permenently.

Then I saw noq2's build. It was brilliant. And even better, I can have everything I originally wanted with a dead-simple addition.

The whole 'tactical flashlight' craze has spawned a smorgasboard of cheap, canned collimation optics for LEDs. They're even better than lenses as they sit all around the emitter instead of a distance in front, so they catch more light. Most are 90%+ efficient.

So I'm taking what noq2 did and slapping a $2 spot-collimator on front. No muss, no fuss. Well, I have to fuss a bit, so I'll put the driver electronics inside the housing, and add an intensity adjustment too. Still not *much* fuss.

...beginning with hacking off a big ol' chunk of heatsink using a dull beaver...

and cleaning it up a bit on a sharper beaver.

Trim to fit the existing enclosure, with an eye toward reusing the existing mounting holes on the enclosure's rear boss. Also drill and tap holes for LED mounting, power jack, and a notch for a BuckBlock controller, and we end up with something like:

I want an intensity adjustment knob, so I stole the boss off the prototype I machined earlier this year. Along with bolts, a really nice 11-position-with-detents potentiometer, and the aforementioned LED optics (but not including wires) we finally end up with this exploded build:

Next up will be a little paint, a schematic, and build/test.

Performance-wise, 1.3 is the biggest Opus update so far. Several commenters have suggested it should have been named 2.0, but full interoperability is important, and '2.0' could suggest we broke compatibility somehow (we haven't).

Opus 1.3 adds full Ambisonics surround support, improves the built in speech/music detection, and greatly improves low bitrate performance. Wideband speech now goes down to 9kbps, narrowband to 5kpbs, and stereo performance improves especially in the 24-32kbps range.

Go forth and deploy!

tl;dr summary: many (most?) of the 12" HV121P01 SXGA screens with a 'bathtub ring' or 'retro-tv' effect around the edges are not defective, VR1 simply got knocked out of adjustment.

THE WHOLE STORY:

This is one of those occasions where I feel like a complete idiot and simultaneously wonder how nobody else noticed this before.

I experimented a while back with chemically stripping the front glass off HV121P01-101 screens, since these were plentiful (at the time) but you had to remove the glass and bonding adhesive to use them in an X61/X62. Mechanical stripping is labor intensive to put it mildly, and using xylene or alcohol seemed like a useful shortcut. It worked, but it also damaged the polarizer films in exactly the way illustrated in the picture.

At the same time, people were getting reclaimed screens from China showing the exact same effect. I got one or two of these myself, and the polarizer films were in fact damaged just like I saw on the screens I stripped.

And I assumed from there on out that chemical/heat stripping was the only explanation of the effect. Which turns out to be terribly wrong.

Most LCD screens have a variable resistor adjustment as part of a temperature/drive compensation circuit. I played with it on many screens in the past, and it was a way to alter either the absolute drive or overdrive speed of the entire screen, usually affecting gamma and contrast in some fashion.

That's not what it does on an HV121P01.

I was modding some of the screens I'd recently bought for LED backlight, screens which I'd tested carefully on receipt and found no defects. After modding, three turned up 'bathtub ring' defects when tested. The defect appeared spontaneously, and others had noted this happen after doing an LED mod. At the time, this was deeply disappointing, perplexing and expensive. I was not going to sell any defective screens no matter how subtle the defect.

Did the LED mod cause the fault? It had been near 100% humidity in NH all that week, did that do it? Was it a fault that was always there and only showed up with LEDs? Or was it always there and I had simply missed it?

I peeled and inspected the polarizers on one screen; this panel had a replacement polarizer that was glossy, so it wasn't going to be saleable anyway. And if chemical stripping had caused the defect, missed till now, why did the *replacement* films show the problem? They didn't, in fact--- after removal and testing, they were faultless, perfectly regular in every way.

I queued up a second panel for testing (I didn't want to burn my own replacement films on a potentially bad panel). Everything about it looked perfect until I was displaying low-brightness gray-to-gray stipple patterns, and that's when the bathtub pattern appeared. Could it be some sort of mismatched signal drive? I looked at VR1, which I'd never touched because I was sure I knew what it did.

So I tweaked it. And the problem got worse. I tweaked it the other direction and the problem disappeared.

I'm still testing in detail to make sure this isn't multiple unnoticeable problems stacking up into a noticeable one, but it sure looks to me right now that this is an adjustment to balance panel drive in the center versus the edges of the screen. It's normally fixed after adjustment at the factory with a little lacquer, but it's not the slightest bit surprising it might get knocked loose or dissolved during rebuilding or modding.

VR1 probably can't mitigate a genuinely frotzed polarizer but it's obviously worth trying it just to see if it was never the polarizer at all.

There's some good LCD news though-- I found a seller who's trickling out small numbers of refurbished HV121P01-100 screens, and unlike all the others I've sampled in the past several years, these have so far been excellent. Yes, they're rebuilt, but they appear to use entirely genuine, model-appropriate parts. I have no idea if these are parts from B-grade panels being reassembled into new screens or what, but the results are NOS visual quality.

There's a 'downside': he actually seals the panels together with black RTV silicone. If you want to open the panel up to do further surgery, you can't. Or rather, with a thin spudger, a ton of patience and very very steady hands you can, but slip once and you'll crack the matrix. I did open some up to have a detailed look--- Yup! All genuine inside!

If you *don't* have any reason to open the screen up, the black silicone is a good idea-- it keeps grit out, and prevents the dreaded 'white spots' from ever developing. I've considered building screens this way myself, so I actually approve.

I'm importing a few of these for conversion to LED. If you want one, contact me about it. If you want to order directly yourself, it's item #710816705 on AliExpress. I have no idea how many per month he can actually make, or if the quality is going to hold up, but so far, so good!

My second technology deep-dive article on the new AV1 codec is up, this time at Mozilla Hacks. It's a continuation of the video codec technology pages that I started with Daala.

"The Constrained Directional Enhancement Filter is the one we're interested in here; like the loop restoration filter, it removes ringing and basis noise around sharp edges, but unlike the loop restoration filter, it's directional. It can follow edges, as opposed to blindly filtering in all directions like most filters. This makes CDEF especially interesting; it's the first practical and useful directional filter applied in video coding."

My first technical writing regarding the new AV1 codec is up at Xiph.Org. We've been working on AV1, heads-down, for a long time and my writing took a hiatus for almost that entire period. Now that it's out, it's time to continue the technology pages that I started with Daala.

"AV1 is a new general-purpose video codec developed by the Alliance for Open Media. The alliance began development of this new codec using Google's VPX codecs, Cisco's Thor codec, and Mozilla's/Xiph.Org's Daala codec as starting point. AV1 leapfrogs the performance of VP9 and HEVC, making it a next-next-generation codec. The AV1 format is and will always be royalty-free with a permissive FOSS license."

I've modded nearly 300 X61 inverters to drive LEDs over the past four years, and I thought I'd seen all the possible FRUs. NOPE.

Not listed in the hardware reference or parts cross-reference: the Sumida FRU 41W1024. Undeniably an X61 inverter.

And here it is, before modification, for reference purposes.