Closed timaschew closed 4 years ago
timaschew
commented
4 years ago ~~Just restarted the raspberry. Two minutes after boot, the docker consumed the whole memory.
I guess the high CPU is because it starts using the swap file.~~
// edit the problem was a wrong device for the microphone!
synesthesiam
commented
4 years ago Can you post your profiles? I'm not seeing this behavior at all. Thanks.
timaschew
commented
4 years ago Sure.
{
"intent": {
"remote": {
"url": "http://192.168.178.21:12101/api/text-to-intent"
},
"system": "remote"
},
"microphone": {
"arecord": {
"device": "hw:CARD=Device,DEV=0"
},
"pyaudio": {
"device": "2"
},
"system": "arecord"
},
"sounds": {
"system": "dummy"
},
"speech_to_text": {
"remote": {
"url": "http://192.168.178.21:12101/api/speech-to-text"
},
"system": "remote"
},
"text_to_speech": {
"system": "dummy"
},
"wake": {
"system": "snowboy"
}
}Maybe I know when there is a performance issue. I left rhasspy in the config of having the wrong device for the microphone: this was in the settings:
"microphone": {
"arecord": {
"device": "null"
},
"pyaudio": {
"device": "2"
},
"system": "arecord"
},Now I switched to the proper soundcard and the memory usage is almost constant at 38%.
synesthesiam
commented
4 years ago Ok, thank you. This must be related to the microphone recording a bunch of fake data very quickly when the device is wrong. I protect against empty data being returned, but maybe I need to rate limit it or something.
timaschew
commented
4 years ago Okay, yes so it's my fault, sorry for the confusion.
Then let's go back to the original error. I've just successfully recorded a wav in the docker container with arecord -D hw:1,0 -d 5 -f cd test.wav -c 1 and I can play it and it's correct.
but and then I did with this command: arecord -q -r 16000 -f S16_LE -c 1 -t raw -D hw:CARD=Device,DEV=0 test.wav -d 5and the result is this wav file which seems to be broken:
https://send.firefox.com/download/8ab4527d05737f61/#vLTPKDoKFgr4QmgQVYP_8A
I think it would be nice if the profile allows to control the whole record command (including all the args).
synesthesiam
commented
4 years ago The -t raw is what's making it seem broken. Wav files have a header normally, but Rhasspy already knows the format when recording. Change this to -t wav on the command line.
timaschew
commented
4 years ago You're right!
This works: arecord -r 16000 -f S16_LE -c 1 -t wav -d 3 -D hw:1,0 r16-fs16-twav.wav
I'm getting a warning that 16000 is not supported only 44100, but it still works.
Alright, so how this can be solved? Do you have a plan?
timaschew
commented
4 years ago Ah, I think you've missed my point that the recognition never worked via the microphone directly via rhasspy. It always recognize a totally different sentence. But when I'm uploading the wave files to the webinterface it works perfectly.
timaschew
commented
4 years ago Actually this time (I think I've tried it before with an untrained model) the recognized intent is an empty string. I tried it four times with different sentences but the response is always:
DEBUG:__main__:{"entities": [], "intent": {"confidence": 0, "name": ""}, "siteId": "default", "slots": {}, "speech_confidence": 1, "text": "", "time_sec": 0.0024182796478271484, "wakeId": ""}// edit I've disabled microphone, that's the reason why the intent is empty. Still getting wrong sentences.
This issue got a bit messy, so I've created a new one: #144
First of all, thank you for this great project!
I have this soundcard and this microphone. I was using snowbow, google and wit.ai last year but now stopped due to privacy concers. And now I want to switch to Rhasspy :). The microphone is attached on my raspberry 3B.
I spent 3 days to figure out that the german profile leads to a performance issues (high memory and CPU) in both docker and virtual-env setup (the raspberry becomes unusable after some minutes and then crashes). Somehow it doesn't happen wit the english profile, but that's another story (and issue which I will create later).
I followed the tutorial for the client/server setup.
However the microphone automatic selection doesn't work. For neither PyAudio nor ASLA. When I try to record something (triggered via the webinterface) the recording in the logs looks like this:
Choosing PyAudio leads to this error:
BTW: when running rhasspy via virtualenv it even showed me one device in PyAudio which could be autoselected via the Test button, but recording was still not possible. There was also a (error) suffix after clicking on Test. When doing a recording afterwards the logs showing this:
Using ALSA gives this feedback:
But when trying to record it doesn't record anything and I even cannot stop the recording via the webinterface. The logs is showing just this, even when I click on Tap to stop recording.
I know that the correct device is hw:CARD=Device,DEV=0 even if rhasspy shows (no sound) after clicking on Test. So I've just selected it and finally recording via rhasspy webinterface started to work, but:
I saw this in the rhasspy logs:
When running this command in the host or in docker:
arecord -D hw:1,0 -d 5 -f cd test.wav -c 1the file is fine and I was even able to upload it via the rhasspy webinterface and all my 4 different examples were recognized correctly. I almost gave it up with rhasspy but this proofs it's really working.But how can I fix it now?
Oh noo, I've just realized even with the english profile the raspberry becomes unusable after 2-4 hours. CPU is at 100% it shows me this on an attache monitor, but I cannot ssh or even do a ping.