Mark II Architecture Change - A Note from the CEO


#1

Originally published at: http://mycroft.ai/blog/mark-ii-architecture-change/

Architecture Change on Mycroft Mark II

Mycroft Community and Backers,

We set out in January 2018 to provide a smart speaker for the Mycroft ecosystem with a graphic display and substantial improvements to the audio. After taking our lumps during the Mark I manufacturing process, we decided the Mark II would go more smoothly if we brought in a vendor to handle the electrical engineering, audio processing, and firmware.

We decided to outsource this process because it would allow us to focus on software while a partner with more specific experience developed the Mark II hardware. Unfortunately our vendor hasn’t delivered. This doesn’t absolve us of responsibility because ultimately the product is our responsibility from start to finish, but it hopefully helps clarify what’s been going on behind the scenes here at Mycroft.

The last published delivery date for the Mark II was the end of Q1, but it looks like we’re going to slip significantly to the right. As we said in the January 2019 update, we won’t speculate on timelines anymore but will let you know as soon as we have clarity. We sincerely apologize for the delay.

When we selected the vendor for the Mark II, we felt confident that they would be able to deliver a great product. Our engineering and audio provider promised a “turn-key” solution that would give Mark II a lot of cutting-edge capabilities.

Unfortunately, the promised performance never materialized. Though the hardware itself seems to be capable, we have not been provided firmware, drivers, and software that work reliably. That means the screen crashes, the board overheats, and we’ve yet to see the claimed audio processing technologies working on the prototype hardware.

After fighting these problems for months at a cost of nearly $5,000 per day we’ve decided to cut our losses and move development to a proven off-the-shelf solution. The development platform we’ve adopted is running a Rockchip 3229. We’re deploying a four-step process to evaluate this platform as the basis for the production Mark II. We’re already in Phase 2 of validation.

Moving to a platform that has been the basis for other voice solutions in the past should accelerate our development effort. We’re now making rapid progress on the software and are optimistic given the progress of the new prototypes. We still need to produce and refine the plastics, pick a screen for the new systems, and refactor our setup software. However, since the developer boards are readily available, we are not waiting on a fabricator to print, populate, cook, test, and ship the prototype PCBs at a cost of thousands of dollars each.

One benefit of this approach is that outside developers will be able to participate in the effort without waiting for production hardware. Once we can confidently say we’ll be able to deliver using this solution, developers will be able to compile a prototype with off-the-shelf components and our supplied hardware schematics and software.

We already have a prototype of the new solution operational and will soon have the first run at the plastics printed. Chief of Design Derick Schweppe has an update with details that will be published this coming Friday. (Derick’s post is now live. Read it here.) We’ll keep you posted on further progress as we make this change.

As always, we appreciate everyone’s support as we work through these challenges. Hardware is hard and we’re glad to have so many cheerleaders supporting our efforts.

Sincerely,

Joshua Montgomery
CEO
Mycroft AI, Inc.


#2

Is Aaware still being used for the microphone array or is that out the window as well?


#3

Hi aussieW, we’re going completely new hardware and as Josh said, the Rockchip is already being used in other voice solutions which is a huge benefit. We have also ended the relationship with the existing vendors.


#4

That’s a bit of a bummer. I hope the new mic array solution gives a result as good as what was described in the Kickstarter campaign.

Improved Hearing!

The Mark II will be able to hear you even when you’re across the room, or your music is turned all the way up. When you need perfection, you partner with the best; based in Silicon Valley, Aaware, is a leader in designing microphone arrays that eliminate interfering noise and preserve speech with little to no distortion.

The microphone system consists of an array of 6 individual microphones that when combined with Acoustic Echo Cancellation, Noise Reduction and Beamforming are able to isolate the speaker even in a noisy environment.

Beamforming takes the input from the 6 microphones and locates the direction from which the user is speaking, giving it superior listening ability.


#5

It does indeed :smile:

There are more details in Friday’s post on the way forward


#6

(secretly hopes this means 2gb ram on board)


#7

Well, the SOC has 2 8 channel I2S interfaces, so hooking up the microphone array does not seem to be a problem.
In the worst case, you can do AEC, NR and beamforming in software, but it is going to use a nontrivial amount of CPU power to do that… I hope there is a DSP that can handle these task so the CPU can stay in its idle state most of the time.


#8

Any specific reason to choose the RK3229 and not the RK3399Pro?
http://rock-chips.com/a/en/products/RK33_Series/2018/0130/874.html

I mean, as the damage has already been done why not using the newer AI based SoC including a NPU?

Perhaps Precise (or mimic1/2, deepspeech, ehatever) could be updated to utilise this NPU, giving the CPU/GPU back to the other processes.

Just thinking out loud here, not based on any information. Just wondering, why 3229?


#9

I have a ReSpeaker CoreV2 which is RK3229-based. General performance is similar to a PiCroft on a RPI3B, the echo-cancel/beamforming algorithm slows it down noticably.


#10

Yeah, that might be the reason indeed. They went with seeedstudio as replacement board.

Considering the GUI and Wayland stuff, I do wonder however if 3229 is future proof as it is not a big step up compared to Mark-1 (cpu wise)


#11

The cost, I’m going to guess.
rk3288 is faster, but less popular and appears not to have as many integrations with sound-based boards. rk3399 is much more expensive, the pro even more yet. Allwinner parts are another option, but also not as much device integration that I see.
rk3328 seems like a more preferrred option, but i suspect there wasn’t an easy way to get that with the audio bits, either:


#12

Thanks for the update on the architecture revision. Looking at the info (RK3229) and the pictures I think it is safe to assume that the new development board is the Respeaker Core V2, correct?

As I already own a Core V2 I am happy to test the adapted Mycroft enclosure, skills and echo-cancelling software as soon as it is available.

And I am looking forward for the release of the 3d-printer-files for the “Mark IIr” from the second evaluation phase.


#13

I agree. That type of device could also be sold by Mycroft AI as something like a mark mini - something under $50…


#14

I hope you get compensation from the vendor for not delivering the expected hardware. Can you tell the name of the vendor? I’m just curious.


#15

Don’t remember the vendors name, but i am sure it is mentioned in one of the previous blog posts. There was also mentioned that some batches of component were already paid for which are non-refundable.