May 2006, Issue 190

ARM-Based Modern Answering Machine
Philips ARM Design Contest 2005 First Prize

FSK SECRETS

An FSK demodulator demodulates the caller ID information from the telephone line. A simple edge detector followed by a counter can be used to demodulate FSK even on the tiniest 8-bit microcontroller. These demodulators perform well when fed with a perfect signal, but they show limited performance in real (noisy) applications. I couldn’t run such a demodulator on my DSL-equipped telephone line. I needed a high-performance FSK detector.

Interestingly, it’s hard to find information about demodulating FSK without using an edge detector. One semiconductor company that shall remain nameless even turned the source code from an application note into data statements (.word xxxxh) in the heart of an FSK demodulator in an obvious effort to hide the secret of FSK demodulation. After searching the ’Net for some time, I finally found the technical note, “FSK: Signals and Demodulations,” in which author Bob Watson describes a filter-type approach to the demodulation of FSK signals.

I also found a practical implementation in the ham radio world: a data transmission system called packet radio can use FSK modulation. A radio link is generally a poor media for data transmission. Noise, distortion, and phase shift are common impairments that affect the signal. The receiver must be robust enough to take care of these impairments.

The FSK system used by ham radio operators is similar to the Bell 202 standard telephone companies use to carry caller ID information. The signaling speed is 1,200 bps. The frequency for a 0 is 2,200 Hz. The frequency for a 1 is 1,200 Hz.

There are also some differences. For example, packet radio is synchronous, and caller ID is asynchronous. That doesn’t really affect the signal processing though. A practical implementation of the secretive FSK demodulation involves two filters tuned to the 0 and 1 frequencies. 

The demodulator uses four tables. The coeffloi[] and coeffloq[] tables are initialized with the cosine and sine components of eight samples at low frequency (1,200 Hz). coeffhii[] and coeffhiq[] have eight samples at high frequency (2,220 Hz). Every time a sample is retrieved from the ADC, the low- and high-frequency filters are run with the last eight samples. This process is detailed in Listing 2.

At the end of the filter loop, outloi and outloq represent the phase and amplitude of the 1,200-Hz component of the input signal. outhii and outhiq represent the phase and amplitude of the 2,200-Hz component. I’m not interested in the phase information, so I can just calculate the total energy in each filter by taking the sum of the squared I and Q component (see Figure 3). I can then subtract the energy detected in the low filter from the energy detected in the high filter. If the result is positive, then a high frequency (2,200 Hz, bit 0) is assumed. If it’s negative, a low frequency (1,200 Hz, bit 1) is assumed.