Speech Recognition on ODROID

[nshmyrev]

> is CUDA GPU is necessary to install TensorSpeech?

Not really, you can edit setup.py and change tensorflow-gpu to tensorflow. You can get tensorflow 2.3.1 wheel here:

https://github.com/bitsy-ai/tensorflow-arm-bin/

You also need to compile cpptflite, for that you just build tflite according to instructions on their website:

https://www.tensorflow.org/lite/guide/b ... _with_make

That almost works here, the only remaining problem is for tensorflow-addons for me, still looking on it.

[nshmyrev]

Ok, with the attached file you can get going without addons. They need to make separate python preprocessor though without tensorflow dependency.

You can get precompiled binary, the patch, instruction and audio sample here:

https://alphacephei.com/tts/tensorflow-tts-pi.zip

[nshmyrev]

On RPi3 it is 2 times slower than realtime though, not very practical still. Vocoder takes most of the time

Maybe it is worth to wait a bit more.

[n2fan]

many thanks; I will try and provide the results

On RPi3 it is 2 times slower than realtime though, not very practical still. Vocoder takes most of the time

Maybe it is worth to wait a bit more.

[joerg]

I have now a setup with rhasspy running, where my N1 with debian bullseye is base station and the C1 is slave station. This way the main work for dialog managment, intent handling, stt and tts is done by the N1.
On the C1 as can be seen in my thread I managed to output audio via spdif on pin 7 GPIOY_3. Now I need a small spdif to analog converter to feed speakers.
On the C1 is enabled the audio recording, the wake word with snowbird and the audio output. The communication is going with mqtt hermes protocol.
I was not able to install rhasspy in a venv environment on the N1 with bullseye, due to missing python 3.7 dependencies. But in a docker container it's running.

[L67GS]

I'm still having trouble with the microphone. Sometimes it just doesn't accept having modules module passed to it.
I wrote a little IF statement in my .bashrc to start pulse and load modules too. Even adding them command line doesn't always work.

[nshmyrev]

I'm still having trouble with the microphone. Sometimes it just doesn't accept having modules module passed to it.
I wrote a little IF statement in my .bashrc to start pulse and load modules too. Even adding them command line doesn't always work.

Pulseaudio is not really necessary, you can simply configure alsa, it should be enough.

[L67GS]

I'm still having trouble with the microphone. Sometimes it just doesn't accept having modules module passed to it.
I wrote a little IF statement in my .bashrc to start pulse and load modules too. Even adding them command line doesn't always work.

Pulseaudio is not really necessary, you can simply configure alsa, it should be enough.

Beamforming with ALSA? I haven't found anything about that anywhere.

[L67GS]

Oh, and ALSA hates my PS Eye, alsamixer crashes if I try to select it

[nshmyrev]

> Beamforming with ALSA? I haven't found anything about that anywhere.

Pulseaudio beamformer is just a name, it tracks fixed source, the implementation is trivial 10 lines of code:

https://cgit.freedesktop.org/~arun/puls ... beamformer

Real beamforming better done in a library, like webrtc:

https://chromium.googlesource.com/exter ... mformer.cc

[L67GS]

Julius works but isn't ideal for what I'm trying to do, Vosk works but it's not trainable.
I'm starting to work more seriously on DeepSpeech, the arm7 version wouldn't work before.
Worst case scenario I guess is Julius and a limited vocabulary for now.

[mad_ady]

Oh, how advanced does it have to be to understand "Kill all humans"?

[L67GS]

Oh, how advanced does it have to be to understand "Kill all humans"?

THAT can be done with Julius!
Even a simple phrase falls prey to misunderstanding on Vosk, pocketsphinx could do it given enough time, after a few years Kaldi might run to test it.

If DeepSpeech cannot be made to perform I'm falling back to Julius and doing "key phrases" (not sure if that's a thing or I just coined it) and limiting the conversational aspect of the project.

[nshmyrev]

Hey, some Vosk recent updates: we have moved to JNA/CFFI framework for all the bindings meaning you can distribute same library for all languages and versions. Now you can install Vosk with Maven/Nuget/NPM/Pypi with a simple commands and it should work independently of Node or Python version. Great progress!

https://www.nuget.org/packages/Vosk/
https://www.npmjs.com/package/vosk
https://alphacephei.com/maven/com/alphacephei/vosk/
https://pypi.org/project/vosk/

Android platform has been redesigned too, now it loads models much faster and follows recent Android practices too.