Matrix Voice: Beamforming, AEC, dereverberation and noise cancellation

Greetings everyone!
My Matrix voice should arrive soon, so i am currently doing research on all components i need for my project.

Since the code of the FPGA is not open source yet, what is the state of the promoted features?

  • Beamforming
    -> The delay & sum algorithm is currently done on the CPU side, not FPGA. Only one beam supported.
  • Acoustic Echo Cancellation (AEC)
    -> No info found, is it implemented?
  • Dereverberation
    -> No info found, is it implemented?
  • Noise Cancellation
    -> No info found, is it implemented? Is the Autio Output already cancelled out?

There is code for receiving each microphones data as PCM, Everloop control and Acoustic Source Localization (also non-FPGA algorithm). I could not find code for Audio Output over the Hadphone Jack.

Also there seems to be no official pinout of the RPI Header or a proper board schematic. I only found some reverse engeneered(?) info. Are there some sort of official documents available somewhere?

Thanks for any help! :slight_smile: I hope this board gets a bit more open source, so one can actually use it efficiently to build stuff. Having to reverse engeneer stuff of an “open source” platform is quite annoying.

2 Likes

It arrived today, which is good, but nobody answered, which is bad.

I don’t use a RPi, so the most basic thing to get going is still missing: A board schematic!

1 Like

i do not have a matrix voice, but there was released an example yesterday, to use the audio jack on the voice:

@Draghi can you confirm whether the device has AEC?

AFAIK, no. But that knowledge is pretty old now. I didn’t touch it since February. The only thing the FPGA did back then was translating the analog levels of the MEMS microphones into PCM.

1 Like

Hi Matrix guys! I am the one from some people who have a big expereince in mic arrays. I am engineer with patents contains my own mic of arrays, the methods for localization and some scientific articles. I am also the buisness partner for some startups which creates smart speakers.
My attention is for this one: https://newatlas.com/matrix-voice-raspberry-pi-voice-recognition/53034/
This is a repost that your Matrix Voice board is:
It’s 3.14 inches in diameter (that’s Pi to two decimal places of course) and has far-field voice capture, beamforming, acoustic source isolation, noise suppression, de-reverberation and acoustic echo cancellation among its feature set, too.
So after introduction to github I see that you have not supplied your board with smth like you described. It is not a far field because you didn’t create the reverb code so the max distance is near 3 meters during the lack of noise. In noisy rooms it is not working at least from 1 to 3 meters.
You code provide the ALSA only raw PCM samples.
I am partner of company which aquired at least 6 boards which are creator or voice. But we are disappotinted with that you lied using your own sources of news. It is not good.
You use the incorrect way of public informing causing the sales to be growth but your customers are maybe disappoitend and then give your products back.
I totally agree with topic starter that now the Matrix projects is smth like a fool!
I have a big amount of another working products with dereverb, AEC, AGC, beamforming and localization algos. I made a test to compete between synaptics, intel, allwinner, microsemi and matrix. So after my partner check the results and then they give me a matrix boards to check why did they aquire the bad result. And it was a big news for me to understand that you. Matrix guys, just didn’t do nothing in acoustic algos. It is a simple data logger which is multichannel with pdm to pcm transform in FPGA. For this purpose the using of FPGA is a really funny decision. The PDM to PCM transfrom can be made in many ARM devices in software only. It use the DMA and it is enough to at least 80 Mhz of cpu like cortex M3. And it can be introduced in cpu like Cortex A using the DMA, the SPI interface and the correct C code in kernel module.
So my opinion is that you must aquire you boards back, return money and make a real product.
The another way is to give the community the proper code. If it will be not free than I agree it will be more good then nothing. The another incorrect way is that you didn’t suuport the hardware sources for the hardware product which now the Matrix is. So in such situation the Matrix as a product is just a waste of money and time.

1 Like

I 100% agree with you. The FPGA code has been published here 9 month ago. No development since then and I can’t find any of the promoted mic array features in there.

It’s a crowdfunding fraud as far as I’m concerned and I regret buying it.

The most diasppointing thing is to understand that after payment more than 100 $ we bought the device which natively act as 1 microphone instead of beam array of mics. The result to get the raw stream of several mic data getting from ALSA is so weak that I definitely understand what can I buy instead to save money because there are a lot of another products for example respeaker ones which is cheaper and has no dsp algos in decstiption. So I definitely understand that Matrix is just cheated me.
I also think that they have some algos. My opinion is based under fact that they used DDR2 memory in Voice with FPGA. But in Creator they didn’t. So I think that they make a decision and cheat the public which buy their product, but technically they made fome features to improve product but didn’t public ones.

Yepp. The Matrix Voice could be great with the right software for that FPGA. The propability that we will get that as open source anytime soon is low. So just forget about it and buy somewhere else. :slight_smile:

The new ReSpeaker Mic Array v2.0 has only 4 instead of 7 Microphpnes but they claim that it has a similar performance due to the more advanced XMOS XVF-3000. Did you test that one? Is it any good?