StuartIanNaylor

Is not the mdtc model recurrent? PS have you had a look at https://github.com/google-research/google-research/tree/master/kws_streaming as its also a great KWS resource?

Yeah I think it was that model that kickstarted much as Arm then did a repo. https://github.com/ARM-software/ML-KWS-for-MCU Which kick started the extensive Google Research https://github.com/google-research/google-research/tree/master/kws_streaming which is incredibly extensive even...

> I think they all are, which has worked out well for [openWakeWord](https://github.com/dscripka/openWakeWord) because it uses a pre-trained Google embedding model. [microWakeWord](https://github.com/kahrendt/microWakeWord/), on the other hand, performs much better with...

You can use synthetic data but additional, real highquality recordings with quality metadata is key. From Mlcommons spoken word the forced alignment causes huge bias in the dataset to the...

Its worth talking about some points as there are some datasets needed desparately, whilst other datasets are created on the back of missconception, myth and self interest. Its very simple...

Also here is the collection of piper clean samples before adding noise as I should of said I also run sherpa piper at 0.8 speed as also for me at...

:) No not when it will likely be a waste of resources and time. A word dataset needs to be more than just wakeword the user requests for different are...

Specifics about dataset used of the for the Kw samples that are still only 3,885 unique voices. Coqui ⓍTTSv2 870 Emotivoice 1932 Piper (Sherpa) 904 Kokorov1 53 Kokorov1_1 103 Kokoro_en...

> [@farooqkz](https://github.com/farooqkz) I'd love to continue the discussion, but this is not possible with Stuart present. He has received numerous warnings, and is no longer welcome in the Home Assistant...