4 ms·

This would be pretty powerful if you combine it with speech recognition. If you put this array in the middle of a crowded room, can you better tease apart diffe

by bcherny 10y ago

This would be pretty powerful if you combine it with speech recognition. If you put this array in the middle of a crowded room, can you better tease apart different audio sources?

Artlav 10y ago
Ironically, this thing can't record sound as we know it - it only have enough buffer space for a few hundreds of samples per mic, 30 times per second. They are to be FFT-ed rather than continuously recorded. While such a contraption should, in theory, be able to distinguish between dozens of voices in a room, it would take much, much heavier signal processing to achieve.
jsprogrammer 10y ago
If you have more than one mic you can begin to triangulate sources (you might be able to use a single mic if you know the geometry of the room and/or could send out noises), so it would definitely be possible. Consider that most of us are carrying around multiple, networked mics. Some may already do it, but I doubt it will be too long before everyone can virtually tour real-life spaces and visualize the sounds there in real-time.
lucb1e 10y ago
Not this specific hardware, but in general, yes. Amazon Echo does pretty much this, except it focuses on one source instead of all possible sources (from what I know). PulseAudio also recently gained a feature to use beamforming (see a HN submission from either today or yesterday), but for now it needs the location of the object you're trying to record as an input. I guess you could kind of brute-force that and let it run many times over the same audio stream to find all possible audio sources.
dharma1 10y ago
You can use the phase/timing information between different mics to separate sources, yes. Interestingly it's also doable with a single mic - previously with algorithms like NMF and ICA, more recently with neural networks. Our brains do it very well so there is probably more to discover. Google blind source separation or cocktail party problem if you want to find out more.
dharma1 10y ago
You can use the phase/timing information between different mics to separate sources, yes. Interestingly it's also doable with a single mic - previously with algorithms like NMF and ICA, more recently with neural networks. Our brains do it very well so there is probably more to discover.

by bcherny 10y ago

This would be pretty powerful if you combine it with speech recognition. If you put this array in the middle of a crowded room, can you better tease apart different audio sources?

Artlav 10y ago

Ironically, this thing can't record sound as we know it - it only have enough buffer space for a few hundreds of samples per mic, 30 times per second. They are to be FFT-ed rather than continuously recorded. While such a contraption should, in theory, be able to distinguish between dozens of voices in a room, it would take much, much heavier signal processing to achieve.

jsprogrammer 10y ago

If you have more than one mic you can begin to triangulate sources (you might be able to use a single mic if you know the geometry of the room and/or could send out noises), so it would definitely be possible. Consider that most of us are carrying around multiple, networked mics. Some may already do it, but I doubt it will be too long before everyone can virtually tour real-life spaces and visualize the sounds there in real-time.

lucb1e 10y ago

Not this specific hardware, but in general, yes. Amazon Echo does pretty much this, except it focuses on one source instead of all possible sources (from what I know). PulseAudio also recently gained a feature to use beamforming (see a HN submission from either today or yesterday), but for now it needs the location of the object you're trying to record as an input. I guess you could kind of brute-force that and let it run many times over the same audio stream to find all possible audio sources.

dharma1 10y ago

You can use the phase/timing information between different mics to separate sources, yes. Interestingly it's also doable with a single mic - previously with algorithms like NMF and ICA, more recently with neural networks. Our brains do it very well so there is probably more to discover. Google blind source separation or cocktail party problem if you want to find out more.

dharma1 10y ago