Before I forget part of the reason why usb of either 1/2 channel and you can 2 or more instances on either the ‘base’ or ‘satelite’ with many of the SoCs.
That is the 4th core ‘Pixel strip’ for a wired remote additional mic(s) with or without a ‘on-device’ one.
Also an endfire is a fixed distance static delay&subtract (one behind the other) and makes the mic more directional (cardioid), limited by channels and unidirectional mics being a bonus. The alg adds hardly any load as the broadside is 95% TDOA to beamform to get the delay.
Soundcards generally are geared up for a broadcast style distance (close talk) and lack the gain, even in near field, where the Max9814 is just a low noise preamp I like, adding a preamp extends into nearfield.
That can be expotentially better than digital gain of a low signal and filters from GitHub - SaneBow/PiDTLN: Apply machine learning model DTLN for noise suppression and acoustic echo cancellation on Raspberry Pi to simple hiss removal can be used also and artefacts are not a problem for a KWS if the dataset is mixed with noise and filtered and you train it in.
Also BSS that has no clue but seperates into x2 distinct streams where you select by running 2x KWS.
JorisCos/ConvTasNet_Libri2Mix_sepclean_16k · Hugging Face
Likely heavy but quantising whilst converting to Onnx could be possible, or just simple multiple mics.
1 Like