Ways to improve the voice quality with xCORE Smart Microphone and XVSM-2000

Voice related projects and technical discussions
Posts: 8
Joined: Mon Jun 26, 2017 1:08 pm

Ways to improve the voice quality with xCORE Smart Microphone and XVSM-2000

Post by florian.ott »

we selected the xCORE Smart Microphone and XVSM-2000 board for use in our project because of its noise suppression and the functionality to get the direction of arrival.
We want to do speech to text recognition using the signal we receive via USB. Therefor the qualtity of the signal must be as good as possible.
Tests with the unchanged sw_usb_audio implementation on the device now have shown the following:
- with no noise the speech to text works good even at different distances of a speaker to the microphones
- if there is constant noise directly next to the board and the speaker is 0.5m away, the speech to text fails even at low noise levels
Directly selecting the microphones to use (Button C) did not improve the result much. There still seem to be too many distortions on the signal output to detect the speech correctly. Or there might be some information on the voice signal missing. I am not sure about that.
For comparison we repeated the test with:
- smart phone
- laptop with integrated microphone (two way array, also with noise suppression)
- simple analogue microphone
The results with the smart phone and laptop were pretty impressive because the speech to text worked very well. With noise directly next to the microphones the distortions of the noise were nearly completely removed when recorded with audacity.
With the simple analogue microphone the results with speech to text were comparable to the xCORE board. A recorded signal with the analogue microphone showed really loud noise, which was a lot lower with the xCORE board, but the speech to text results were nearly the same.

That leads me to the question, what can be done to improve the efficiency of the noise suppression? We expect the result of the xCORE board to be be at least as good as the rather cheap laptop microphone.

Thank you in advance.
Posts: 8
Joined: Mon Jun 26, 2017 1:08 pm

Post by florian.ott »

I would really like to get a feedback on the above test results. What are your experiences with the board? I am not sure, if the xCORE smart microphone is not the right choice for our purposes, if the reference implementation is not sufficient and should be adapted somehow or if this is just a configuraton issue. I hope to hear from you soon.
User avatar
XCore Legend
Posts: 1913
Joined: Thu Jun 10, 2010 11:43 am

Post by mon2 »

Hi. Did you ever receive a reply from XMOS on your question? Perhaps the response went to private? Curious on your results. Did the XMOS microphone array work for you correctly?

We are experimenting with a mix of alternate options as $1500 for the XMOS kit is silly to evaluate the tool.

So far, have the SEEED version of the older XMOS array which offers a directional LED index; Conexant and Cirrus. Each is plagued with some issue. You could consider to bring in the SEEED PCB for under $80 USD from Digikey but don't hold your breath for any support. This is true for most silicon vendors these days. The Cirrus tool offers a GUI interface that offers a checkbox for with and without their IP performing the clean up of the picked up audio in realtime. Nice feature.
User avatar
XCore Expert
Posts: 1006
Joined: Thu Dec 10, 2009 9:20 pm
Location: Bristol, UK

Post by Ross »

It should handle stationary noise suppression quite well is my understanding - there are some parameters to that can be modified in this area; ns_on, ns_attlimit_dB. Most of our attention is now on the VocalFusion range, this has much improved and entirely new DSP.
Technical Director @ XMOS. Opinions expressed are my own.