[Question] Fine tuning wake words

[VivantSenior]

Hi,

is there any possibility to fine tune the wake words?

Best,
V

[StuartIanNaylor]

Yeah its called ondevice training in tensorflow speak

There is an example here https://ai.google.dev/edge/litert/models/ondevice_training

It wouldn't be on device training but from a local server such as the Intel N100 recomendation to run locally where a small dataset of device capture would update a larger premade model.
The tensorflow training has shifted as it used to use a smaller model to bias the weights of a larger but now seems the initial model can be fine tuned.

So when idle which with a voice server is considerably it can be finetuning the model with actual usage data.
If the esp32-s3 only uses half the flash you can use the OTA (over the air) firmware update which works a bit like android switching beween 2 images.
Or some sort bespoke method to transfer just the model.
On the esp32-s3 you have a rolling window with enough margin that when the hit of a KW is detected that window is the capture of your KW.
That is uploaded to the local server to create a KW dataset.
Also command sentences can be captured in a similar way from ASR as due to the logic of a system to how further actions such as stop or repeat requests you can deduce what might be good and bad and your asr has the transcript.
You need meta data of the device so captures recorded on a device create a dataset for that device but yeah KWS and ASR could learn from use.