Rhasspy offline voice assistant toolkit

Share your Projects! Custom Integrations
446

2.07 works with me!

I have first used EN, downloading went ok, after that I selected NL. Also worked correct!
Then I did an extra language DE, which downloaded fine as well :slight_smile:

1 Like
447

Sorry about that! I stopped storing the compiled web interface files in git, but I forgot to add the appropriate build step to create-venv.sh. It’s update in the git repo now, so please try again.

448

Based on the recommended settings, it looks like you should check “Apply Frontend” for computer.umdl and set the sensitivity to 0.6. You can download any of the other models too, and see how they do. The jarvis and neoya models probably won’t work, since the seem to require setting multiple sensitivity values (I need to figure out how to do this inside of Rhasspy).

449

Shouldn’t it be to 0.4, since I need the wake word using snowboy being detected a bit easier?

Edit:
It is working better with 0.6 now. Thanks.

1 Like
450

@S_n_Nguy_n, I’m happy to announce that the latest version of Rhasspy is my first attempt at supporting Vietnamese! Pick the vi profile, save, and restart Rhasspy to get started. After it downloads, make sure to re-train before trying it out.

Thanks again for the WAV files and transcriptions. I used them to create the sample sentences and test the acoustic model. Hopefully I got it working well enough to do simple voice commands. Let me know if it works; I could try the same approach for some a bunch of other languages.

451

If there is something that I can do to help, I would like to see Portuguese (BR) implemented.

452

I will try and respond to you to further improve, thank you

453

This is stop_words for vietnam profile, can you add it?

454

Got it. I’ll add this for the next update. Thanks!

455

I’ve got a start on a Portuguese profile, but I don’t have any WAV files for testing. If you would be willing to record yourself saying a few commands (“turn on the living room lamp”, etc.), that would be great. A few WAV files and the text that Rhasspy should detect are all I need for testing.

456

That’s great! I will record and send to you in a few days. I am out of the city right now. I presume 16bit 44.1kHz mono should do the trick.

457

That would be perfect. My initial tests with phrases “spoken” by Google Translate have not been very promising. If testing with real audio from you doesn’t work either, I have a backup method. Someone awesome put together a list of free Portuguese speech data sets, so I can just train my own model.

458

I’m happy to report that the latest version of Rhasspy adds support for Portuguese! Thanks to @FunkyBoT for supplying test data, and a big thanks to the people who created the Montreal Forced Aligner for supplying pre-trained acoustic models in many languages where it’s almost impossible to get good quality, free speech training data.

If anyone can supply some example WAV files with transcriptions in any of these languages, I will do my best to add it to Rhasspy.

459

Got an error downloading the PT profile.

sshot-1
sshot-11859×146 24.7 KB

460

Sorry, had the wrong download link in the script. Should be fixed now. You may need to use the red Re-Download button on the Settings page, or just delete the profiles/pt/download directory.

461

Updated to 2.10 but still getting an error:

sshot-1
sshot-11784×148 22.4 KB

462

I’d recommend deleting the download directory in your pt profile on your Hass.IO share. It pulled down the wrong file (portuguese_g2p.zip), and won’t re-download it unless it’s gone.

463

I had pressed the “red download button”. Will try deleting the files in the pt profile.

464

I’ve double-checked that everything got pushed up to Dockerhub correctly. The problem is this file, which should be about 5MB. I had accidentally used the URL for the Vietnamese file, which is about 3MB.

465

Training in PT are taking too long. “R” spinning forever

There is a small misspelling in “light_state = (acenda | apague) {state}”