see the output of the semantic VAD with the server inference

[samuelOTOFA]

Is it possible to get the output of the semantic VAD with the server inference ?

[LaurentMazare]

Yes if you use this config, the server should actually send you some Step messages over the websocket. These messages should contain the VAD predictions.

[samuelOTOFA]

Great, I saw it, thank you.
I noticed that there can be several 'Step' data between a 'Word' and an 'EndWord' data.
Could you please give me a brief explanation of what each one means, and also of the 'prs' field found in the Steps?
I assume these are probabilities, but I'm not sure what each of the four values represents.
Thanks in advance, and congratulations — you're doing a great job!

[LaurentMazare]

The different messages are:

• Step message occurs on each inference step, so 12.5 times per second. prs contain the probability for an end of turn for the different time horizons (0.5s, 1s, 2s, and 3s).
• Word message is when a fully formed word has occured - if someone speaks super quickly this can actually contain multiple words. This includes the word start time.
• EndWord is the end of the word that started before. This includes the word end time.

[samuelOTOFA]

Great, thanks