@Orbiter Thanks for the testing. Yes, large model is not going to work, it is mainly for the servers. You need to use small model.
We can add susi if critical, it is not very common words.
To recognize from 5m distance you need a special microphone (respeaker or something like that). What microphone did you use?
@Orbiter and you?
There are a lot of processes running when susi is startet, one java server, one node server and two python processes. I deliberately let them running while doing the first
test_microphone.py tests with the same German sentence as done before. vosk almost perfectly understood the sentence with one or two words wrong. I repeated the test several times and there were always at least one wrong word.
However when I stopped the susi processes and tried again, I got a perfect result with no wrong word at all. Thats very good!
@nshmyrev I did a
htop during the test and that showed that vosk is running on one single CPU only. Is there any way to run it on all four cores of the RPi? That would be a massive improvement!!
And it would mean it is fully functional on a RPi3 as well!
Isn’t this maybe as simple as passing some parameters to kaldi, I see some things here: https://kaldi-asr.org/doc/queue.html
–num-threads 2 where could we put such things in? Or maybe thats too naive..
It would be nice to have a presentation with SUSI on RPi that can be carried around with no need to attach it to a power socket. So that seems possible.
It can be easily intergated into a box with the RPi. Maybe I do a design.
There is also a 2-battery version, will try that next
I pushed support for vosk to susi_linux and susi_installer. If you have a desktop installation or rpi, git pull in the two dirs should be enough.
What remains is putting the models into a directory vosk-data/LL below the speech_recognition python lib, on the desktop this is in /usr/local/lib/python3.N/dist-packages/speech_recognition/vosk-data/en/
From the log:
...
LOG (VoskAPI:ReadDataFiles():model.cc:194) Decoding params beam=10 max-active=3000 lattice-beam=2
LOG (VoskAPI:ReadDataFiles():model.cc:197) Silence phones 1:2:3:4:5:6:7:8:9:10
LOG (VoskAPI:RemoveOrphanNodes():nnet-nnet.cc:948) Removed 0 orphan nodes.
LOG (VoskAPI:RemoveOrphanComponents():nnet-nnet.cc:847) Removing 0 orphan components.
LOG (VoskAPI:CompileLooped():nnet-compile-looped.cc:345) Spent 0.0242901 seconds in looped compilation.
LOG (VoskAPI:ReadDataFiles():model.cc:221) Loading i-vector extractor from /usr/local/lib/python3.9/dist-packages/speech_recognition/vosk-data/en/ivector/final.ie
LOG (VoskAPI:ComputeDerivedVars():ivector-extractor.cc:183) Computing derived variables for iVector extractor
LOG (VoskAPI:ComputeDerivedVars():ivector-extractor.cc:204) Done.
LOG (VoskAPI:ReadDataFiles():model.cc:251) Loading HCL and G from /usr/local/lib/python3.9/dist-packages/speech_recognition/vosk-data/en/graph/HCLr.fst /usr/local/lib/python3.9/dist-packages/speech_recognition/vosk-data/en/graph/Gr.fst
LOG (VoskAPI:ReadDataFiles():model.cc:273) Loading winfo /usr/local/lib/python3.9/dist-packages/speech_recognition/vosk-data/en/graph/phones/word_boundary.int
...
INFO: Keyword 1 detected at time: 2021-03-22 19:57:59
DEBUG: Entering hotword callback
DEBUG: We are idle, so work on it!
DEBUG: vlcplayer: starting to say something!
DEBUG: vlcplayer: finished saying
DEBUG: notify renderer for listening
DEBUG: listening to voice command
DEBUG: Converting audio to text
INFO: Trying to recognize audio with vosk in language: en
DEBUG: recognize_audio => what is the time
...
Dear @/all today we received the very sad news that our dear friend @iamareebjamal has passed away. We are very disturbed by Areebs sudden death. In the last few years Areeb became like the beating heart of FOSSASIA. He kept everything running and developed all kinds of amazing services and tools. We would not be where we are today without him. He was an amazing person and brilliant developer. I worked and communicated with Areeb nearly every day and I also had the chance to meet him in person in India. The last time we got together in Delhi when we organized the OpenTechSummit mini event in Delhi. He loved to share his knowledge with anyone and expected even nothing in return. He was a modest and humble person. He deserved to live a long life and make his dreams come true. Areeb had big plans and we had discussed to get him to Singapore to run the FOSSASIA development from there. An amazing future was in front of him. We will be sharing more about Areeb on our channels. Please post your thoughts and condolences on the web and provide support to each other in these difficulties times. Be safe and healthy!
You can find out a bit more about Areeb on a video we put together here https://www.youtube.com/watch?v=XbFCBp-LTXc
Hello! I cannot create an account for susi.ai . When I try to register on the web page, the button never becomes active; when I try to register on the app it tells me my email is already registered. When I try to log in, the website tells me my password is wrong and the app tells me there is no account. When I try to reset my password, the website sends the email with a link which appears to work, but nothing changes; the app tells me that there is no account.
hey, is the project frozen? the community doesn't seem to be very active, the e-mail of the site susi.ai doesn't seem to work, in this channel all are silent... I ask to get an idea of the current status O.o
Hi @rixty.dixet:matrix.org well it is not frozen completely, but we have higher priority things, and the only two working continuously on the project have become rather busy in real life. That said, we are happy about people having interest, and wanting to bring things forward. Ideas there are a lot.
1 reply
Also exciting http://p3r.one/gpt-j/ GPT-J: GPT-3 Democratized
Hello @/all, FOSSASIA and OPNTEC are running the Codeheat Coding contest again. The contest is run in memory of Areeb Jamal. Contribute to Free and Open Source Software and OpenH ardware FOSS projects. Our focus projects are Pocket Science Lab and Open Event eventyay. You can get a certificate and win prizes! More info on https://codeheat.org
Hi, Deutsche Telekom also provides a cloud voice service now. They have a couple repos online https://github.com/telekom/voice-skill-sdk and https://github.com/telekom/voice-cli
What are your thoughts @Orbiter
Hey all, Happy New Year.
Some news from Vosk project - we have just released public Japanese model for Vosk:
https://alphacephei.com/vosk/models/vosk-model-small-ja-0.22.zip
more languages are ongoing.
Is fossasia popular in Japan?
Interesting P2P search approach using bittorrent https://boredcaveman.xyz/post/0x2_static-torrent-website-p2p-queries.html
Interesting also:
SEPIA #OpenAssistant v2.6.1 😃 It comes with a lot of improvements and new features e.g. shared access permissions 📲📲, custom wake-words 🗣🤖, LED array interface 🚨, new TTS voices and much more. Check it out: https://github.com/SEPIA-Framework/sepia-installation-and-setup/releases