Hacker News Reader

Comment by wl | Hacker News Reader

> they made an audio filter using techniques and understanding of sound in space that only became available in last 10 years.

This is an old technique. The idea of convolving the impulse response of a system with an input to yield the output is the basis of LTI systems theory. And while convolution isn't the most computationally light operation in the world, computers have been up to the task of practically performing such operations on audio signals for at least 20 years.

There's been some more recent work about more efficiently and accurately measuring the impulse responses of a space, but given they're using balloon pops, they're not using that work.

View on HN

Offline audio-rate convolution (particularly with shorter impulse responses) has been plausible for a while, but real-time convolution reverb has definitely come into its own over the last ten or fifteen years. Overlap-adding blocks of audio that have been processed in the frequency domain with a low enough latency so as to feel instantaneous is a more recent capability that likely had an influence on this work.

I think the reason they used balloon pops was because that was all they were allowed to use: many heritage or archaeological sites can be nervous about researchers bringing in large loudspeakers and amplifiers, whereas a portable recorder and a bag of balloons can feel a little more harmless. Most acoustics researchers are aware of and use those new techniques whenever it's possible!

I wonder why one would need an impulse response recording. The dimensions and materials of the chamber are known. Seems like the effect could be synthesized from that.

As I metioned earlier, the tricky part is measuring the materials. The rest can be done with pretty high accuracy. There are commercial packages for that like Odeon or CATT acoustics.

These days, some of the simulation algorithms are slowly trickling into game engines, too, with the push for realistic VR. Unfortunately, these still require some larger quality/performance tradeoffs because they have to share the processor with all the other work the engine has to perform. But it is still quite convincing. So with this tech, you actually get to walk around in a space and hear the reverb adjusting in real time, even when the geometry of the space changes.

Impulse responses (or an equivalent like transfer functions) are the ultimate output of such a model. Here, they're measuring them directly.

Reply on news.ycombinator.com