4 ms·

One Voice Detector to Rule Them All

61 points by sebg 5y ago

solarmist 5y ago
Oh nice! I was just thinking that I would need one soon. I'd like to break up audio into sentences/lyrics and that is exactly what I need for that.
larrik 5y ago
I wish this article existed a few years ago when my kid wanted to make a voice deactivated fire alarm. Most solutions require sending the audio data over the internet, which was unreasonable in a school gym. Most other solutions weren't great, and I was really looking for a "wakeword" type solution, rather than true "speech to text". I got it working alright at home, but loud environments triggered the wakeword constantly (so did simply blowing into the mic). It was good enough to demo, though.
- IshKebab 5y ago
  This doesn't do wake word detection. It's just "is someone speaking". But I agree, this is one of the building blocks that you need for fun voice projects that was hard to come by until quite recently. Good wake word and speech recognition libraries are still difficult to find.
- dekhn 5y ago
  A voice deactivated fire alarm sounds like something that would end up on the RISKS mailing list, where a bad implementation led to many deaths (for those who don't recall, the Nest Smoke Alarm had voice deactivation that didn't really work well. https://bgr.com/general/google-nest-smoke-alarm-video/ https://bgr.com/general/google-nest-smoke-alarm-video/)
  - larrik 5y ago
    On the other hand, if you are disabled and can't reach the fire alarm to silence it, you are at a disadvantage.
cushychicken 5y ago
Heh, I wrote a VAD from scratch when I was working at Sonos. Wasn't that different from any of these implementations. You can add some basic spectral analysis and make it even better, with minimal latency implications. "Cepstrum" and "spectral entropy" are some valuable search terms to apply to this.
- echelon 5y ago
  What was your time at Sonos like? Do they have a good engineering culture?

4 ms·

One Voice Detector to Rule Them All

61 points by sebg 5y ago

solarmist 5y ago
Oh nice! I was just thinking that I would need one soon. I'd like to break up audio into sentences/lyrics and that is exactly what I need for that.
larrik 5y ago
I wish this article existed a few years ago when my kid wanted to make a voice deactivated fire alarm. Most solutions require sending the audio data over the internet, which was unreasonable in a school gym. Most other solutions weren't great, and I was really looking for a "wakeword" type solution, rather than true "speech to text". I got it working alright at home, but loud environments triggered the wakeword constantly (so did simply blowing into the mic). It was good enough to demo, though.
- IshKebab 5y ago
  This doesn't do wake word detection. It's just "is someone speaking". But I agree, this is one of the building blocks that you need for fun voice projects that was hard to come by until quite recently. Good wake word and speech recognition libraries are still difficult to find.
- dekhn 5y ago
  A voice deactivated fire alarm sounds like something that would end up on the RISKS mailing list, where a bad implementation led to many deaths (for those who don't recall, the Nest Smoke Alarm had voice deactivation that didn't really work well. https://bgr.com/general/google-nest-smoke-alarm-video/ https://bgr.com/general/google-nest-smoke-alarm-video/)
  - larrik 5y ago
    On the other hand, if you are disabled and can't reach the fire alarm to silence it, you are at a disadvantage.
cushychicken 5y ago
Heh, I wrote a VAD from scratch when I was working at Sonos. Wasn't that different from any of these implementations. You can add some basic spectral analysis and make it even better, with minimal latency implications. "Cepstrum" and "spectral entropy" are some valuable search terms to apply to this.
- echelon 5y ago
  What was your time at Sonos like? Do they have a good engineering culture?