Whistled Platform Upgraded With Word Recognition 30
An anonymous reader writes "A few weeks ago, Slashdot featured a cheap platform performing 80FFTs per second to recognize whistles. The platform is open hardware/open source and is aimed for sound processing projects. To this goal, the creator (limpkin) just implemented a simple proof of concept algorithm that will control your lighting once the platform listens to a particular word. A small video has been made to explain the basic concepts of sound recognition to encourage hobbyist to make their own."
Not A Criticism, But... So What? (Score:5, Interesting)
The author of the book, a high-school student, built (with the help of his engineer father, I don't want to downplay that) a small robot that had obstacle sensors, light sensors, and some basic mapping capability so it could find its own charging station (not so different from a Roomba today).
But my main point is: it also had effective voice recognition, for simple commands. And the implementation was pretty simple: the audio input was amplified, then sent through 3 notch filters to separate high, medium and low audio frequencies. Each of the 3 frequencies was digitally sampled at about 40kHz. Repeated samples were averaged and saved in a table in memory.
The CPU (and remember, this was a 1970s-era CPU, if I recall an 8080a or a Z-80 or similar) constantly sampled incoming sounds, and when one sufficiently matched one of the stored templates it meant "command received".
It was a simple scheme, and it worked fine. I don't want to detract from this inventor, but in essence he is doing a similar thing. Except instead of using notch filters, he's using FFTs to do the frequency analysis and build (and then compare to) the templates. The ideas aren't all that different.
But personally, I think I'd prefer the old method, as it demonstrably worked at least as well as this, used only a few $ in hardware in addition to the CPU, and was pretty definitely less compute-intensive to achieve.
Keep in mind: that was 40 years ago. Maybe this newer approach has more potential; I don't know. But it certainly doesn't look much different at this time.