Posted on May 17, 2011 at 14:15I use the Goertzel algorithm. My trick is that I precondition ADC inputs such that integer math works rather than the obvious floating point implementation. FWIW, more points makes a narrower filter at the various target frequencies. At 8KHz sample rate using 100 points (12.5 milliseconds of digitized analog samples) works well. My crappy $12.95 Fry’s handset often failed on the 2580 column using 200 points. In retrospect, (overlapping) 25 milliseconds of samples is overlong for good decodes anyway.
I also run the second harmonics of the standard DTMF tones. This is done to reject otherwise good decodes. The elimination of these frequencies would save ~40% of the calculation time.
Note for Janvi-
You wrote that your stm32 application decodes DTMF by hardware. It appears that you are using a timer and DMA channel and ultimately using a table of pulse lengths. Would you be so kind as to point me to a description of your technique? Are you looking at zero crossings detected via an stm32 analog port or external hardware?
I’m confused about your reuse of the timer. It makes it sound as if DTMF decode and its inverse (tone generation) are somehow coupled in some way other than the specific frequencies involved. My tone generation uses assembly code for a fast sin(x) where x is an unsigned integer where 0 to FFFFFFF = 0 to 2pi. That sin(x) is only good to 12 bits but it’s more than good enough for tone generation.
Email is in original post if you wish to use it.
