ASR partial is painfully slow


#1

a version or two ago, when ASR Partial worked, it worked really great and fast…

the latest update… its painfully slow

[02:03:54] [Asr] was asked to listen on site zero
[02:03:55] [Asr] is capturing text: ""
[02:03:55] [Asr] is capturing text: ""
[02:03:56] [Asr] is capturing text: ""
[02:03:56] [Asr] is capturing text: ""
[02:03:57] [Asr] is capturing text: ""
[02:03:59] [Asr] is capturing text: ""
[02:04:00] [Asr] is capturing text: ""
[02:04:01] [Asr] is capturing text: "what"
[02:04:01] [Asr] is capturing text: "what"
[02:04:02] [Asr] is capturing text: "what is"
[02:04:02] [Asr] is capturing text: "what is"
[02:04:03] [Asr] is capturing text: "what is three"
[02:04:04] [Asr] is capturing text: "what is three"
[02:04:04] [Asr] captured text "what is three" in 10.0s

and using the default settings in tool

[snip-asr]
partial=true
partial_period_ms=250

#2

Hey ! Just to be sure, firstly, are you using the version 0.60.8 of the snips-platform ?
Secondly, like explained in the doc, this feature requires a lot more computational power than a simple inference so depending on the size of your assistant and your device it can be quite slow. You can start by increasing the partial_period_ms which will definitely speed up the process but you will have less partial decodings.


#3

yup running 0.60.8
RPi3B+ running small demo assistant with like 3 bundles in it

from the output i pasted you can see its like 1 sec between each output in the snips-watch window… i have finished talking before the for “what” is even returned which is returned 6 sec after the ASR begins

like i said, like 2 versions back the partial text was great, worked fast, no lag, same RPi device, same assistant bundles, just a different snips version

setting the toml to 350ms

[11:47:49] [Asr] is capturing text: ""
[11:47:49] [Asr] is capturing text: ""
[11:47:50] [Asr] is capturing text: ""
[11:47:51] [Asr] is capturing text: ""
[11:47:51] [Asr] is capturing text: "what"
[11:47:52] [Asr] is capturing text: "what is"
[11:47:53] [Asr] is capturing text: "what is three"
[11:47:53] [Asr] is capturing text: "what is three"
[11:47:54] [Asr] is capturing text: "what is three"
[11:47:54] [Asr] is capturing text: "what is three"
[11:47:54] [Asr] captured text "what is three" in 6.0s

set to 500ms

[11:48:47] [Asr] was asked to listen on site zero
[11:48:48] [Asr] is capturing text: ""
[11:48:48] [Asr] is capturing text: ""
[11:48:49] [Asr] is capturing text: ""
[11:48:49] [Asr] is capturing text: ""
[11:48:50] [Asr] is capturing text: "what"
[11:48:50] [Asr] is capturing text: "what is"
[11:48:51] [Asr] is capturing text: "what is three"
[11:48:52] [Asr] is capturing text: "what is three"
[11:48:52] [Asr] is capturing text: "what is three"
[11:48:53] [Asr] is capturing text: "what is three"
[11:48:53] [Asr] captured text "what is three" in 5.0s

slowwwwwwww assssssssssss :slight_smile:


#4

its now runs super fast on my Mac… but on a RPi3b+ its not worth it