Modem architecture

The transmitter

Framer

The framer converts IP packets into a continuous stream of bits. AuDSL transmits IP packets encapsulated in HDLC-style frames. Currently, there is no frame checksum or packet header; the IP packets are simply bit-stuffed and separated with HDLC flag sequences. When the line is idle, a continuous stream of HDLC flags is transmitted.

Scrambler

The bit stream from the framer is scrambled using a self-synchronizing scrambler similar to the one defined in ITU-T recommendation v.29. The scrambling process spreads the transmit spectrum evenly over the transmit bandwidth, which helps reduce interference and aids in the convergence of the adaptive equalizer in the receiving modem.

Mapper

Consecutive groups of bits are mapped into transmit signal amplitudes. When transmitting at 96 kbps, each pair of transmitted bits is encoded as one of four distinct signal levels, resulting in a 4-level PAM encoded signal at a 48 kHz sample rate.

Partial response encoder

The PAM encoded signal from the mapper is filtered with a linear FIR filter to shape the transmitted signal spectrum. The filter has spectral nulls at DC and near the Nyquist frequency, concentrating the transmitted signal energy into the 4-20 kHz audio band. This filtering process can be viewed as introducing controlled intersymbol interference (ISI), turning the signal into a partial response signal.

The receiver

Sample rate converter

Because the crystal oscillators used to generate the sampling rate in sound cards have a finite accuracy, the sound cards at the two ends of the AuDSL line never run at exactly the same sample rate. Typically, the difference in sampling rate is on the order of +-100 parts per million (ppm). To compensate for this difference in sampling rates, the receving modem resamples the received signal using a high-quality sample rate converter based on a combination of polyphase FIR filtering and linear interpolation.

Adaptive equalizer

Next, the signal is passed through a linear equalizer to compensate for the non-ideal frequency and phase response of the transmission line. The equalizer adapts to the line response using a simple LMS algorithm during an initial training phase. During the data transmission phase, the LMS algorithm continues to run in decision directed mode, but only once for every 16 signal samples.

Partial response decoder

After equalization, the signal samples consist of the transmitted multilevel PAM signal, with some additional noise and a large amount of controlled ISI introduced by the partial response coder. This ISI is removed by using a decision feedback decoder. The decision feedback decoder works by constructing a replica of the transmitted ISI based on previously detected bits and subtracting that from the signal. The signal can then be detected using a simple threshold detector ("slicer").

Descrambler

The scrambling done in the transmitter is undone by a self-synchronizing descrambler as in ITU-T v.29.

Deframer

Finally, HDLC flags in the descrambled bit stream are detected, the bit stuffing is undone, and the resulting bits are assembled into an IP packet. Packets whose length is not a multiple of 8 bits are discarded.

Synchronizer

The receiver is kept synchronized with the remote modem transmitter by adjusting the sampling phase of the sample rate converter. This is done using a feedback loop based on the current equalizer coefficients. After the equalizer has been trained, its coefficient vector will have a peak corresponding to the maximum correlation between the received signal and the transmitted one. The feedback mechanism attempts to keep this peak at the center of the coefficient vector by continously adjusting the sampling phase.

The echo canceller

Even though the echo itself is of a fairly short duration, it may take a long time to arrive at the receiver because of latency introduced by the sound cards and its drivers. Additional delay is caused by the buffering needed to achieve real-time operation under an operating system with unpredictable scheduling latency. Therefore, the echo canceller filter actually consists of a long ajustable delay followed by a short FIR filter. The amount of delay and the FIR coefficients are determined by a training sequence performed at connection setup time.