Speech Recognition on ODROID

[joerg]

Try to install it from source code, as this:
pip3 install git+https://github.com/alphacep/vosk-api.git
If you see some missing modules, you can install them by hand pip3 install <missing_module> and try again.

[nshmyrev]

> is it possible to use Odroid N2?

I think yes

> odroid@odroid:~/vosk$ pip3 install vosk-0.3.16-cp38-cp38-linux_aarch64.whl

What is you python version? Maybe 3.7 then you need cp37, not cp38

[nshmyrev]

> Try to install it from source code, as this:

Sorry, it is not going to work like this as it requires many dependencies

[L67GS]

N2 has an odd mix of processors which may be an issue with this. Julius should run on it though.

[joerg]

I have received the ReSpeaker 4-Mic Array now. But when I try to compile the driver I see that on kernel 3.10.107 it will not build. Need to install Armbian with mainline kernel on C1. But I have doubts about the stability and it won't work on eMMC, also usb is not working well.
It should run on the C1 I have in the living room that do all the nrf24 traffic for my alarm sensors. I would need to do long time stability tests with mainline kernel, that I have done already a year or two ago and I installed old kernel again.
Old kernel runs stable: wlan, usb, emmc...
But don't have snd_soc_dapm and not gpiod for AC108 driver.

[L67GS]

I have received the ReSpeaker 4-Mic Array now. But when I try to compile the driver I see that on kernel 3.10.107 it will not build. Need to install Armbian with mainline kernel on C1. But I have doubts about the stability and it won't work on eMMC, also usb is not working well.
It should run on the C1 I have in the living room that do all the nrf24 traffic for my alarm sensors. I would need to do long time stability tests with mainline kernel, that I have done already a year or two ago and I installed old kernel again.
Old kernel runs stable: wlan, usb, emmc...
But don't have snd_soc_dapm and not gpiod for AC108 driver.

The source from Seeed isn't helpful?

Code: Select all

$ sudo apt-get update
$ sudo apt-get upgrade
$ git clone https://github.com/respeaker/seeed-voicecard.git
$ cd seeed-voicecard
$ sudo ./install.sh  
$ reboot

https://wiki.seeedstudio.com/ReSpeaker_ ... pberry_Pi/

[joerg]

Oh, I have tried that. It is coded for Raspberry Pi and gives a lot of errors compiling on C1 with kernel 3.10.107.
I was able to build it with mainline kernel after patching a lot.
But when I dig deeper into the source code and the datasheet of AC108 I see that it can't be used on C1!
It uses I2S. Somehow I was thinking it uses SPI, but not.
It makes also no sense to use a C1+, as the Data In for I2S is not there, only Data Out.
My alternatives: Send it back or use it with a Raspberry Pi.

[L67GS]

I thought it was overpriced to use in a personal project. The PlayStation eye camera is my current favorite but I'm still looking around, in a security system application the playstation camera may be a better choice than respeaker, just in case you get the itch to fiddle with facial recognition later in life.

[nshmyrev]

Looks like I need another Odroid to try everything in place. What do you think mates? Which one to get? Does Odroid C4 look reasonable?

[nshmyrev]

Btw, I pushed updated microphone code, should be much more reliable now:

https://github.com/alphacep/vosk-api/bl ... rophone.py

[L67GS]

Looks like I need another Odroid to try everything in place. What do you think mates? Which one to get? Does Odroid C4 look reasonable?

IMO, for R&D the C1+ would be the best choice, followed by XU4, here's my reasoning:
In the Odroid current ARM line-up 4 out of 7 are ARMv7 so ARMv7 is the logical choice to get things started.
Of course your C1 uses the same image so you already have something C1+ compatible.

C1+ -vs- XU4:
XU4: ARMv7 and a strong community favourite, but due to low voltage GPIO pins is less desirable for makers and more desirable for use as a server.
C1+: Backwards compatible with the old C1 & C0, both are discontinued but there are still quite a few still in use today. Also costs like 30% less than the XU4 making it more cost effective for R&D and small projects that one would typically use Speech Recognition with..

Here's the line-up (ARM):

C2:
1.5GHz ARM 64bit Cortex-A53 Quad-core CPU

XU4:
Samsung Exynos5422 Cortex™-A15 2Ghz and Cortex™-A7 Octa core CPUs

HC1:
Samsung Exynos5422 Cortex™-A15 2Ghz and Cortex™-A7 Octa core CPUs

N2+:
Amlogic S922X (4x Cortex-A73 @ 2.2GHz, 2x Cortex-A53 @ 1.9GHz)

HC2:
Samsung Exynos5422 Cortex-A15 2Ghz and Cortex-A7 Octa core CPUs

C4:
CPU: Amlogic S905X3 12nm Processor Quad-Core Cortex-A55 (2.016GHz) ARMv8-A architecture with Neon and
Crypto extensions

C1+:
Amlogic ARM® Cortex®-A5(ARMv7) 1.5Ghz quad core CPUs

[nshmyrev]

C1+ is out of stock unfortunately, I doubt it will be actively sold in the future.

What about XU4 vs C4? Does C4 also have low voltage?

[joerg]

No, the C4 has 3,3V IO pins, but power supply is 5.5V ... 15.5V.

[joerg]

Got it:

Bildschirmfoto vom 2021-01-10 15-53-05.png (108.42 KiB) Viewed 1268 times

[joerg]

Installed vosk-api without problems:

Code: Select all

sudo pip3 install vosk-0.3.17-cp38-cp38-linux_armv7l.whl
Processing ./vosk-0.3.17-cp38-cp38-linux_armv7l.whl
Installing collected packages: vosk
Successfully installed vosk-0.3.17

Code: Select all

python3 --version
Python 3.8.5

Code: Select all

Welcome to Ubuntu 20.04 LTS (GNU/Linux 3.10.107-20 armv7l)

[nshmyrev]

> Installed vosk-api without problems:

Works?

[joerg]

I have to learn how to use it. Let me a little bit of time. I am old.

[L67GS]

C1+ is out of stock unfortunately, I doubt it will be actively sold in the future.

What about XU4 vs C4? Does C4 also have low voltage?

C4 is ARMv8, a minority, the C1 you currently have is a better choice imo.

[L67GS]

Got it:
Bildschirmfoto vom 2021-01-10 15-53-05.png

Oh wow you got the PlayStation camera, it just plugs in and audio just works right?

[joerg]

I have to wait, got it right now in ebay auction. Need to wait until I receive the parcel.

[joerg]

I have tried the test_simple.py of vosk and of some spoken words and counting from one to ten it has recognized the counting correct but the words not correct.
I have recorded in German a sample with the microphone of a webcam. Then I reduced the noise with audacity, the recognition was the same. I used vosk-model-small-de-zamia-0.3.
It took 32s for a sample of 22s.
[edit]
With vosk-model-small-de-0.15. it took only 20s and the words recognition was better but not perfect.

[nshmyrev]

I have tried the test_simple.py of vosk and of some spoken words and counting from one to ten it has recognized the counting correct but the words not correct.
I have recorded in German a sample with the microphone of a webcam. Then I reduced the noise with audacity, the recognition was the same. I used vosk-model-small-de-zamia-0.3.
It took 32s for a sample of 22s.
[edit]
With vosk-model-small-de-0.15. it took only 20s and the words recognition was better but not perfect.

Great, thats something. At least it works. If you share the sample audio I can take a look. In general ASR should be more accurate if you speak closer to the microphone.

[joerg]

Oh, it's a terrible microphone. Let's wait until I get the PS3 Eye Cam and get it hopefully working.

[nshmyrev]

C4 is ARMv8, a minority, the C1 you currently have is a better choice imo.

I think armv8 will be a majority very soon, and even more will have GPU onboard like Jetson nano/BananaPi. Which is best for AI applications.

[nshmyrev]

With vosk-model-small-de-0.15. it took only 20s and the words recognition was better but not perfect.

On what device is that? C1+ ?

[L67GS]

C4 is ARMv8, a minority, the C1 you currently have is a better choice imo.

I think armv8 will be a majority very soon, and even more will have GPU onboard like Jetson nano/BananaPi. Which is best for AI applications.

I'm not so sure about that, the N2+ just came out, the XU4 is kind of the flagship SBC for Odroid, and for the others Hardkernal seems to keep stuff well beyond it's freshness date as long as the demand is there so it could be a while before the v7 is the Odroid minority.

As of now the C1+ is probably the logical choice for a maker board.

[L67GS]

With vosk-model-small-de-0.15. it took only 20s and the words recognition was better but not perfect.

On what device is that? C1+ ?

C1

[L67GS]

Installed vosk-api without problems:

Code: Select all

sudo pip3 install vosk-0.3.17-cp38-cp38-linux_armv7l.whl
Processing ./vosk-0.3.17-cp38-cp38-linux_armv7l.whl
Installing collected packages: vosk
Successfully installed vosk-0.3.17

Code: Select all

python3 --version
Python 3.8.5

Code: Select all

Welcome to Ubuntu 20.04 LTS (GNU/Linux 3.10.107-20 armv7l)

I do not understand why it doesn't work for me.

Code: Select all

root@odroid:~# pip3 install vosk-0.3.17-cp38-cp38-linux_armv7l.whl                                   
WARNING: Requirement 'vosk-0.3.17-cp38-cp38-linux_armv7l.whl' looks like a filename, but the file dot
Processing ./vosk-0.3.17-cp38-cp38-linux_armv7l.whl                                                  
ERROR: Could not install packages due to an EnvironmentError: [Errno 2] No such file or directory: ''

Code: Select all

root@odroid:~# pip3 install https://github.com/alphacep/vosk-api/releases/download/0.3.7/vosk-0.3.7-cp37-cp37m-linux_aarch64.whl
ERROR: vosk-0.3.7-cp37-cp37m-linux_aarch64.whl is not a supported wheel on this platform.

root@odroid:~# python3 --version
Python 3.8.5

[nshmyrev]

> I do not understand why it doesn't work for me.

Try

Code: Select all

pip3 install https://github.com/alphacep/vosk-api/releases/download/0.3.17/vosk-0.3.17-cp38-cp38-linux_armv7l.whl

[L67GS]

It worked, and I see it now, I had tried changing the package but failed to change the directory to 0.3.17

[L67GS]

Success! I can run test_simple.py now!

So I tried test_microphone.py

Code: Select all

root@odroid:/vosk-api/python/example# python3 ./test_microphone.py                                                                      
Traceback (most recent call last):                                                                                                      
  File "./test_microphone.py", line 6, in <module>                                                                                      
    import sounddevice as sd                                                                                                            
ModuleNotFoundError: No module named 'sounddevice'               

[joerg]

Try pip3 install sounddevice
Everytime you get a ModulNotFoundError, it is not installed on your device or it can't be found.

[L67GS]

Try pip3 install sounddevice
Everytime you get a ModulNotFoundError, it is not installed on your device or it can't be found.

Thank you, I'll try that tomorrow.
I found an older microphone comparison with the PS eye, it's an interesting read.
https://medium.com/snips-ai/benchmarkin ... 0de8876fda

[n2fan]

Odroid N2; thanks for help for the previous step; it was a symbolic link issue

odroid@odroid:~/vosk/vosk-api/python/example$ python3 --version
Python 3.8.0

pip3 install vosk-0.3.16-cp38-cp38-linux_aarch64.whl - worked fine

odroid@odroid:~/vosk/vosk-api/python/example$ python3 ./test_simple.py test.wav
LOG (VoskAPI:ReadDataFiles():vosk/model.cc:194) Decoding params beam=13 max-active=7000 lattice-beam=6
LOG (VoskAPI:ReadDataFiles():vosk/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 0.126654 seconds in looped compilation.
LOG (VoskAPI:ReadDataFiles():vosk/model.cc:221) Loading i-vector extractor from model/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():vosk/model.cc:246) Loading HCLG from model/graph/HCLG.fst
LOG (VoskAPI:ReadDataFiles():vosk/model.cc:265) Loading words from model/graph/words.txt
LOG (VoskAPI:ReadDataFiles():vosk/model.cc:273) Loading winfo model/graph/phones/word_boundary.int
LOG (VoskAPI:ReadDataFiles():vosk/model.cc:281) Loading CARPA model from model/rescore/G.carpa
terminate called after throwing an instance of 'std::bad_alloc'
what(): std::bad_alloc
Aborted (core dumped)

any idea what am I doing wrong now?

[L67GS]

It seems we may be able to do some beamforming and echo cancellation with the PS eye array in pulseaudio.

https://arunraghavan.net/2016/06/beamfo ... ulseaudio/

[nshmyrev]

LOG (VoskAPI:ReadDataFiles():vosk/model.cc:281) Loading CARPA model from model/rescore/G.carpa
terminate called after throwing an instance of 'std::bad_alloc'
what(): std::bad_alloc
Aborted (core dumped)

any idea what am I doing wrong now?

Hey, you tried too big model, it doesn't fit your memory. Try smaller one:

http://alphacephei.com/vosk/models/vosk ... s-0.15.zip

[n2fan]

many thanks! smaller model works fine.

###
odroid@odroid:~/vosk/vosk-api/python/example$ python3 ./test_simple.py test.wav
LOG (VoskAPI:ReadDataFiles():vosk/model.cc:194) Decoding params beam=10 max-active=3000 lattice-beam=2
LOG (VoskAPI:ReadDataFiles():vosk/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 0.0869389 seconds in looped compilation.
LOG (VoskAPI:ReadDataFiles():vosk/model.cc:221) Loading i-vector extractor from model/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():vosk/model.cc:251) Loading HCL and G from model/graph/HCLr.fst model/graph/Gr.fst
LOG (VoskAPI:ReadDataFiles():vosk/model.cc:273) Loading winfo model/graph/phones/word_boundary.int
{
"partial" : ""
}
{
"partial" : ""
}
{
"partial" : ""
}
{
"partial" : ""
}
{
"partial" : ""
}
{
"partial" : "one"
}
{
"partial" : "one zero"
}
{
"partial" : "one zero zero"

LOG (VoskAPI:ReadDataFiles():vosk/model.cc:281) Loading CARPA model from model/rescore/G.carpa
terminate called after throwing an instance of 'std::bad_alloc'
what(): std::bad_alloc
Aborted (core dumped)

any idea what am I doing wrong now?

Hey, you tried too big model, it doesn't fit your memory. Try smaller one:

http://alphacephei.com/vosk/models/vosk ... s-0.15.zip

[L67GS]

Of the users interested in this what Odroid devices do we have at our disposal to test vosk on? So far I see C1,C0, C1+ & N2. I also own an XU4 so I can test on XU4 as well but we're still a few devices away from the whole line-up.

[n2fan]

just one more question: I start the test program python3 ./test_simple.py test.wav and start htop in the second terminal. The processor load is 100% on one core but other cores are not loaded. It is intentionally or I overlook something?

[nshmyrev]

just one more question: I start the test program python3 ./test_simple.py test.wav and start htop in the second terminal. The processor load is 100% on one core but other cores are not loaded. It is intentionally or I overlook something?

Yes, we use just 1 core because we want to keep space for other tasks. In the future we might implement multicore system, not implemented yet.

[nshmyrev]

Of the users interested in this what Odroid devices do we have at our disposal to test vosk on? So far I see C1,C0, C1+ & N2. I also own an XU4 so I can test on XU4 as well but we're still a few devices away from the whole line-up.

It should work more or less, the question is accuracy and features for practical applications. Like keyword spotting for activation.

[L67GS]

Of the users interested in this what Odroid devices do we have at our disposal to test vosk on? So far I see C1,C0, C1+ & N2. I also own an XU4 so I can test on XU4 as well but we're still a few devices away from the whole line-up.

It should work more or less, the question is accuracy and features for practical applications. Like keyword spotting for activation.

That wasn't on my radar because my Candice AI does it's own keywords, but I'll study how to test it.
Any recommended reading on vosk-api so I can incorporate into c++? I haven't really looked closely since I wanted to make sure it ran first.

[nshmyrev]

That wasn't on my radar because my Candice AI does it's own keywords, but I'll study how to test it.
Any recommended reading on vosk-api so I can incorporate into c++? I haven't really looked closely since I wanted to make sure it ran first.

For c++ you have to compile it yourself. We will probably release precompiled version soon, but no deadlines. You need to try python, its really flexible for ai applications.

[L67GS]

That wasn't on my radar because my Candice AI does it's own keywords, but I'll study how to test it.
Any recommended reading on vosk-api so I can incorporate into c++? I haven't really looked closely since I wanted to make sure it ran first.

For c++ you have to compile it yourself. We will probably release precompiled version soon, but no deadlines. You need to try python, its really flexible for ai applications.

I figured there was a C++ header file somewhere?
From https://github.com/alphacep/vosk-api:

"Speech recognition bindings implemented for various programming languages like Python, Java, Node.JS, C#, C++ and others."

Edit: yeah there's a header file, so I thought all I had to do was start vosk and then #include it to use it, darn.
https://github.com/alphacep/vosk-api/bl ... vosk_api.h

[nshmyrev]

Edit: yeah there's a header file, so I thought all I had to do was start vosk and then #include it to use it, darn.
https://github.com/alphacep/vosk-api/bl ... vosk_api.h

Yes, you can include that header.

For library file you can unpack wheel and use _vosk.so from it, just rename it to libvosk.so, link to it and it should work.

You can also check vosk-api/c folder for details.

[L67GS]

I found a deal on 2 PS Eye cameras and ordered. They'll be here in a few days and I'll have one for r&d, and one to tear apart for the Candice project.

[L67GS]

I've got vosk working with the microphone, it's not terribly accurate but it's working. I'm going to focus the next couple of days on trying to do C++ integration and see how it likes the PS Eye microphones when they arrive.

[n2fan]

Yesterday, I was able to start a bigger model on N2. It seems to work fine in the real-time with one core load 100%.
Next hurdle: if I start a second recognition process; the vosk server loads the model one more time and N2 will be unresponsbile.
I did following command:
python ./test_simple.py /tmp/fifo1.wav
python ./test_simple.py /tmp/fifo2.wav

how to reuse the model for several recognition processes?

just one more question: I start the test program python3 ./test_simple.py test.wav and start htop in the second terminal. The processor load is 100% on one core but other cores are not loaded. It is intentionally or I overlook something?

Yes, we use just 1 core because we want to keep space for other tasks. In the future we might implement multicore system, not implemented yet.

[nshmyrev]

> Yesterday, I was able to start a bigger model on N2

This model should be probably optimal for you, you can also tune beams in model/model.conf for best performance

https://alphacephei.com/vosk/models/vos ... lgraph.zip

> how to reuse the model for several recognition processes?

If you want to have multiple recognition tasks in parallel you can use vosk server:

https://github.com/alphacep/vosk-server

for simple API we have websocket variant, there is also mqtt variant compatible with home assistant:

https://github.com/alphacep/vosk-server ... aster/mqtt

[n2fan]

thanks, will try the server today. did you make some tests how many channels can be processed in parallel on Intel hardware or any other devices/processors?


vosk-server/client-samples/python/tts-test.py is it a text to speech engine included?