Snowboy stop development

[pepebc]

KITT.AI announces: on December 31, 2020 stop the development of all their products
Have you thought of an alternative?

[Sispheor]

We are evaluating some alternatives.
Fell free to propose as well.

[corus87]

We don't have much options at the moment, there is still porcupine but its not open source.

With the next Kalliope release we are able to use mycroft precise trigger which is open source, I'm already played a bit with precise, but to create a good wake word, it need to take a lot of training but its not very user friendly.

We need at least a good Kalliope wake word, therefore we need to collect some voice samples from the community then it needs a lot of datasets (for example I have downloaded 50GB from the mozilla common voice datasets) to train my wake word.

I'm going to push a community based precise trigger after the next release of Kalliope.

[khimaros]

a few resources since i was just researching this:

• https://github.com/castorini/howl
• https://github.com/seasalt-ai/snowboy
• https://github.com/Picovoice/porcupine

you mention porcupine is not open source, do you mean the models?

[Miloune]

Maybe this one as well: https://github.com/alphacep/vosk-api ?

[Sispheor]

So, the best option is porcupine?
I have some time in the next coming days to work on Kalliope. I'll test some trigger engine myself.

[corus87]

I did take a look into vosk mentioned by @Miloune, it looks promising and at least in the german dataset the word "kalliope" is recognized. I don't have much time lately but a first try with a simple script for a single keyword worked pretty good.

[Sispheor]

Their is also the Snips project implementation. Used by Rhasspy.

[Sispheor]

So for porcupine we cannot generate model for RPI, and generated models need to be recreated every 30 days if I understand well.
So for me it looks like a no-go.

[khimaros]

if you are considering vosk, i'd recommend taking a look at mozilla deepspeech which i found to be both easier to configure and considerably more accurate. they also have examples for streaming transcription in python and bulk transcription in python

[Sispheor]

I thought they were more stt engines than wake word engines.

[khimaros]

@Sispheor -- that's right, including vosk. i'm not sure if either is a good fit for a snowboy replacement, but if you're going to go that direction i think deepspeech is a better target from my limited experiments.

[khimaros]

snips looks promising but appears to have been acquired by Sonos. i couldn't find the platform source https://github.com/snipsco?q=platform&type=&language=&sort= and their usage docs have been removed. the closest available is https://github.com/snipsco/snips-record-personal-hotword

[Sispheor]

Another similar project has created a list of wake work engine here.
They have created their own called Raven. Not yet a python lib but it seems to be full python. Need to take a look.
For me I think the best candidate is mycroft-prcise

[pepebc]

I have tried all and i prefer mycroft-precise, with good accuracy.
Raven is still a very young development, with many false positives.
And the mentioned vosk is very interesting

[Sispheor]

The problem with Mycroft is mainly that the only supported arm proc is armv7l.
I don't find a doc with a full map of all RPI but I think it's recent ones.

[corus87]

Vosk is working pretty good so far, but its slow, with a small dataset it takes about 1-1.5 seconds until the word is recognize. For an Stt it would be a good offline alternative but as a trigger, may not.

if you are considering vosk, i'd recommend taking a look at mozilla deepspeech which i found to be both easier to configure and considerably more accurate. they also have examples for streaming transcription in python and bulk transcription in python

@khimaros I already played around with deepspeech but for more than a year ago, I guess they did some improvements since then, so I guess I will take new look.
For Vosk its very easy to setup, just pip install vosk and download the language model.

The problem with Mycroft is mainly that the only supported arm proc is armv7l.
I don't find a doc with a full map of all RPI but I think it's recent ones.

@Sispheor
It should support the Rpi 2 + 3 and x86 at least those are the ones I tested, but Rpi 4 should also be supported.
Precise is pretty nice, its working good but only with a good trained wake word, and this is very hard to train.
I also read about raven (At least on the rhasspy page) and maybe it worth a shot.

Hopefully soon there will come a good alternative to snowboy or wake work training on precise will get much easier...

Edit: Not all x86 CPU's are supported by precise, AVX is required for the default tensorflow package, but most modern CPU's support AVX.

[nshmyrev]

Vosk is working pretty good so far, but its slow, with a small dataset it takes about 1-1.5 seconds until the word is recognize. For an Stt it would be a good offline alternative but as a trigger, may not.

@corus87 this might be an issue with setup or an old version. Which model/language are you using here? We have updated some of our models couple months ago: English, German. We have also updated the code for much faster latency of the answer. Please try to recheck with latest setup.

Hopefully soon there will come a good alternative to snowboy or wake work training on precise will get much easier...

There will be a keyword spotter in Vosk soon, couple weeks. Stay tuned.

[corus87]

@nshmyrev Thanks for your input! Those are great news you are working on a native keyword spotter, I'm looking forward to it!

I just started earlier this week with Vosk, so I'm using the latest pip version and model you provide here .
For my tests, I've made a small script using pyaudio, but even with your python example there is no difference, it takes about 1 second.

I'm running those tests on x86 with a ryzen 4800h, I guess there should be enough CPU power.

[nshmyrev]

I'm using the latest pip version and model you provide here .

Which model version exactly please? 0.15?

For my tests, I've made a small script using pyaudio, but even with your python example there is no difference, it takes about 1 second.

We do not recommend pyaudio exactly due to latency issues.

there is no difference, it takes about 1 second.

Ok, let me check too

[corus87]

Which model version exactly please? 0.15?

I tried vosk-model-de-0.6, vosk-model-small-de-0.15 and vosk-model-small-en-us-0.15.

Ok, let me check too

Thanks.

[HumanG33k]

hi i just try to install following website instruction.
The default setting file use snowboy, there is somewhere an alternative ?

[vkuehn]

that's closed since ages

[corus87]

The guys from rhasspy built a docker container to run a local web server where you can easily create your own pmdl wake word for snowboy.

https://github.com/rhasspy/snowboy-seasalt

Just install docker and run the command in the readme, it will download the container and starts the web server. Then you can access the interface on http://localhost:8000 and create a wake word in 2 minutes.

[R-Jurado]

Sorry for refloating this old issue, but since I haven't found any recent related issue or article on the matter, may I suggest to support TeachableMachine? It's opensource and easy to use. Don't know if there is an official alternative to Snowboy, I haven't found any.

[1001111github]

I have been using Precise quite happily for years.

However, I currently have an openwakeword trigger, hacked from the precise code, that is sort of functional. Unfortunately, it does not seem to release the microphone after a key word is spotted. In addition to that, there seems to be a mis-match between the precise buffer system and openwakeword giving weird, but reliable, detection values. Sigh, there are advantages to analog when dealing with sound.

YMMV
Dave

[1001111github]

I have a fully functional openwakeword trigger for Kalliope. It is much, much better than my precise trigger. All the default wake words are easily recognized, using any voice. Prototype code can be accessed at https://github.com/1001111github/kalliope-trigger-openww. In addition Home Assistant offers a huge number of word models.

Local STT using the openai-whisper model and code at https://github.com/1001111github/kalliope-stt-whisperer. The only major drawback in Kalliope, was the dependence on "foreign" STT. CMUSphinx was never really an option. Whisper just works. I did have to remove punctuation. This module is not suitable for mobile, lol, as it does require some CPU and memory. Again prototype code, very little testing of edge cases or review

And finally, another TTS, again fully open source using piper. https://github.com/1001111github/kalliope-tts-piper. The biggest advantage is there is no need to execute a system level program anymore. An added bonus is a lot of modern voices.

While working with piper I found that most of my "Device not available OS Error 9985" was caused by pulseaudio and could be prevented by running pavucontrol in the background. I guess my next search is "python pulseaudio control"

All feedback and ideas welcome

HTH
Dave