Where are my files?

I’m looking for the audio file on my Raspberry that is created when I speak into the microphone with Rhasspy 2.5.9.
Where is this stored? How long is it stored?
In what format?
Can I reuse this for further tests?
Where is the json file stored for further use in node red?

It is not stored, it is 16000 16 bit mono per default

Does Raven not have a function to collect?

Do you know in which process I can add the save command?

Yes, but Rave only stores samples for the wakeword :slight_smile:

If you want to save the audio, you have to capture the audiostream from the hermes/audioServer/<siteId>/audioFrame topic :slight_smile:

thanks for your answer, I will read the manual
As a beginner, I have not found any indication of how to adapt the code.
Has anyone tried this yet?
I come back to my question as to whether it is possible to save the capture. I couldn’t find a directory with AudioFrame on my Raspberry. (Rhasspy is in docker). Can I install Rhasspy in a different way and then adapt it so that the waves are saved?
Can saved waves be interpreted again with the same quality?

The low accuracy quick start up KWS that allows you store usage data for latter more accurate NN KWS, but only stores KW where !KW and a mechanism that the audio frame correct are not implemented so you have to DiY with an API?

I will say no more but wow!

I have really no clue of what it is you are trying to say man :smiley:

Wasn’t the rationale for Raven to be able to collect for more accurate trained KWS models?

Its near pointless to collect custom KW without also the custom !kw so guess without the rationale apart from a quick but inaccurate KWS raven is essentially pointless.

The bit Raven was a good idea for to collect datasets easily is only being done partially and yeah its a wow to me!

Hmm, I don’t think so.
With Raven you can create a personal hotword, which is based on the Snips Personal Wakeword
It is not meant to get more accurate universal models
Raven is not pointless if you want a personal wakeword :slight_smile:

Just opinion for me with the WER (word error rate) and some others a personal wakeword needs to be more robust than Raven can produce.
I did think it did have one use as a makeshift KWS to collate datasets as its only use to make a robust custom KWS with precise.

But without it may be of use to some, for me its pointless then and just a bad KWS with an all time high WER.
Maybe its me and my memory and the Snowboy (another old and defunct KWS method) was being forwarded to collect datasets instead as I am aware of the accuracy of those 2 and switched off?

I did think it was Raven as when it comes to more samples and training apart from just sitting infront of a mic and recording its a catch22 to get the dataset from scratch.
It seemed like a really good idea to do so and actually thought it did.