Hacker News Reader

Comment by StavrosK | Hacker News Reader

Snips were acquired? That's bad news, they were the only ones doing offline voice recognition...

They were definitely not the only ones, just the ones with the most aggressively spammy marketing team. HN mods actually stepped in once or twice about it.

Rhasspy is a really cool project that is compatible with a bunch of fully offline speech frameworks. PicoVoice is a super neat company too.

Oh, thanks for that, I've been looking for a good offline speech framework. Unfortunately, the hardest part for me is the microphone array, hopefully I can find something good there.

I haven't had time to play with Rhasspy, but the fact that you can plug and play components from different speech platforms is really promising, particularly when I want to interface it with my own assistant at some point.

Wake word detection, speech to text, intent recognition, etc. is all split out and able to be plugged into separately.

Thanks for the tip! Both Rhasspy and PicoVoice look great. Anyone can recommend a microphone array that works with the raspberry pi?

Look for keystudio respeaker hat for raspberry. I got one for this exact purpose but still haven't got the chance to test it.

Very cool. I see that respeaker has even a 6 mic array for the pi: https://respeaker.io/6_mic_array/

Seconding both ReSpeaker and Rhasspy/voice2json, I've been having great luck with that combo. The docs and example code for ReSpeaker are great, very easy to work with.

I have the (older, discontinued) 6+1 USB mic array, but I'm in general wary of the ReSpeaker stuff. I have the original ReSpeaker Core v1 hardware as well, and it's incredibly unstable. It comes with a tiny amount of flash that you're supposed to augment with a SD card (which then gets overlayfs'd onto the root filesystem), but the hardware is so buggy that running things from the SD card causes random SIGSEGV and SIGILL all over the place (and yes, I've painstakingly verified that this is not a software issue).

They essentially abandoned development of the v1 line (but appear to still be selling it!) for the v2 line, and I have a hard time trusting them.

Admittedly, I haven't done that much with the mic array I have, mainly because I got burned out dumping tens of hours into trying to get ReSpeaker Core v1 working (which I ended up trashing; it was that bad). I'd like to use it with a Raspberry Pi; hopefully that works out.

What about mozilla? https://github.com/mozilla/DeepSpeech

That is one of the toolkits used by mycroft. It isn't everything needed for an assistant but if you want to make one it is probably the best starting point.

For n=1, I got DeepSpeech 0.6 working on a Raspberry Pi, and the recognition accuracy was atrocious. Haven't tried the new 0.7 branch, though.

Keep in mind deepspeech is an implement of speech recognition algorithm... the results will largely depend on model used which typically require lots of manual labor

Reply on news.ycombinator.com