Performance evaluation of XVF3610

[CorentinL]

Hello,

I am currently evaluating the performance of the XVF3610 for wakeword detection. I am using the UA firmware in version 5.7.3.
Using the eval kit L71, the kit is connected by USB to a Linux computer. The sound is played on a TV using HDMI, but is also provided as reference to the XMOS by using ALSA multi plugin to duplicate the sound.
In this configuration the sound played on the TV is not very well reduced and we are afraid to have false positive wakeword if the wakeword is pronouced in the TV stream.

To better understand, what is going on inside the chip, we have used the packed output, and also running a script to retrieve some parameters during the recording.
Here are the parameters used :
SET_IO_MAP 2 7
SET_IO_MAP 3 8
SET_IO_MAP 0 16
SET_IO_MAP 1 16
SET_ALT_ARCH_ENABLED 0
SET_ADEC_ENABLED 1

Here are the polled parameters :
GET_ERLE_CH0_AEC
GET_ERLE_CH1_AEC
GET_ADEC_MODE

Here are the results :

Capture d'écran 2025-01-10 201937.png

As you can see, the ERLE is around 10 dB. I don't know if it can be better...

I also give the corresponding audio file with the 6 channels :

unpacked_audio_tracks.zip

Can you tell me if the results are within the expected performances ? Or is there something we can improve in by the tuning of the parameters ?

Thank you for your help