If the STL follows the audio processor, it must not distort the shape of the audio processor's output wave, which has been tightly peak-limited. To prevent tilt, overshoot, and ringing (which would cause such distortion), the STL's frequency response must be flat to ±0.1dB throughout the frequency range containing significant program energy. The group delay must be essentially constant throughout this range (deviation from linear phase < ±1 degree).

At low frequencies, by far the best way to achieve the specification is to extend the -3dB frequency of the STL to 0.16Hz or lower and to eliminate any peaking in the infrasonic frequency response prior to the rolloff. Poor AFC-loop design in STL transmitters (and FM exciters for that matter) is all too common, and this is the most likely cause of low-frequency response problems. Such problems can be corrected by applying prior to the STL transmitter equalization that is complementary to existing low-frequency rolloff, such that the overall system frequency response rolls off smoothly at 0.16Hz or below. This solution is much better than clipping the tilt-induced overshoots after the STL receiver because the clipping will introduce non-linear distortion, while the equalizer is distortion-free.

If the STL is passing the composite baseband stereo signal from a stereo encoder, the deviation from linear phase, 30-53,000Hz, must be less than 0.1 degree to ensure 60dB of stereo separation. The magnitude response must be ±0.1dB. Provided that the -3dB frequency has been extended to .16Hz or lower (as specified above), these specifications will simultaneously ensure low overshoot and high stereo separation.

Digital STL systems using lossy bit-rate-reduction schemes will not successfully pass peak-limited audio. The lossy compression adds large amounts of quantization noise to certain frequency bands, as determined by a psychoacoustic analysis of the program material. This added noise will cause the peak level of the peak-limited audio to increase substantially. For example, measurements have shown that APT-X™ at 256kps introduces as much as 3db of overshoot with processed audio and ISO/MPEG Layer II at 384kbps introduces as much as 1dB. Overshoots increase markedly as bit rate is reduced. While these overshoots can be clipped or limited, such processing will cause audible side-effects. On the other hand, if the audio processor is located at the transmitter, its input can be fed without difficulty from an STL using a lossy bit-rate-reduction scheme because it is unnecessary to preserve the waveshape of unprocessed input audio.

The requirements for peak control and spectrum control tend to conflict, which is why sophisticated non-linear filters are required to achieve highest performance. Applying a peak-controlled signal to a linear filter almost always causes the filter to overshoot and ring because of two mechanisms: spectrum truncation and time dispersion. One can build a square wave by summing its Fourier components together with correct amplitude and phase. Analysis shows that the fundamental of the square wave is approximately 2.1dB higher than the amplitude of the square wave itself. As each harmonic is added in turn to the fundamental, a given harmonic's phase is such that the peak amplitude of the resulting waveform decreases by the largest possible amount. Simultaneously, the rms value increases because of the addition of the power in each harmonic. This is the fundamental theoretical reason why simple clipping is such a powerful tool for improving the peak-to-average ratio of broadcast audio: clipping adds to the audio waveform spectral components whose phase and amplitude are precisely correct to minimize the waveform's peak level while simultaneously increasing the power in the waveform.

If a square wave (or clipped waveform) is applied to a low-pass filter with constant time delay at all frequencies, the higher harmonics that reduce the peak level will be removed, increasing the peak level and with it the peak-to-average ratio. Thus even a perfectly phase-linear low-pass filter will cause overshoot. There is no sharp-cutoff linear low-pass filter that is overshoot-free: overshoot-free spectral control to FCC or CCIR standards must be achieved with filters that are embedded within the processing, such that the non-linear peak-controlling elements in the processor can also control the overshoot.

If the sharp-cutoff filter is now allowed to be minimum-phase, it will exhibit a sharp peak in group delay around its cutoff frequency. Because the filter is no longer phase-linear, it will not only remove the higher harmonics required to minimize peak levels, but will also change the time relationship between the lower harmonics and the fundamental. They become delayed by different amounts of time, causing the shape of the waveform to change. This time dispersion will therefore further increase the peak level.

When a square wave is applied to a linear-phase filter, overshoot and ringing will appear symmetrically on the leading and trailing edge of the waveform. If the filter is minimum-phase, the overshoot will appear on the leading edge and will be about twice as large. In the first case, the "overshoot and ringing" are in fact caused by spectrum truncation which eliminates harmonics necessary to minimize the peak level of the wave at all times; in the second case, the overshoot and ringing are caused by spectrum truncation and by distortion of the time relationship between the remaining Fourier components in the wave.

It is essential that the measuring equipment have transient accuracy at least as good as the equipment being measured. Many popular modulation monitors introduce false tilt and overshoot into their readings. To avoid such inaccuracy, we used a Belar Electronics Wizard™ FM Modulation Analyzer - a known-accurate instrument - to plot peak modulation versus time on the output of an aggressive audio processing system and several contemporary STL systems and FM exciters. The results reveal the extent of the problem. The same program material and time segment was used for all plots. Fig. 1 shows the audio processor output peak modulation directly. Figs. 2 and 3 show the peak modulation of STL System 2 and FM Exciter 2 respectively. Note that both the STL System and FM Exciter suffer from overshoot causing over-modulation. To prevent the resulting over-modulation, the modulation level must be reduced over 13%, losing almost 1.5dB of loudness!

Fig. 1 AUDIO PROCESSOR OUTPUT PEAK MODULATION





Fig. 2 STL SYSTEM 2 PEAK MODULATION-Before Modification
Moseley Associates, Inc. PCL-6010/C





Fig. 3 FM EXCITER 2 PEAK MODULATION-Before Modification
Continental Electronics Mfg. Co., Inc. 802A

When the stereo encoder is driven by a pure right-only or left-only signal, "stereophonic separation" can be measured at the stereo decoder as the ratio between the desired and undesired signal levels, where the "desired" signal is the signal appearing in the decoder output channel corresponding to the channel driven at the encoder, and the "undesired" signal is the signal caused by the desired signal that appears at the remaining output.

Ideally, crosstalk is non-existent and stereophonic separation is infinite. In practice, both linear and non-linear errors cause these characteristics to deteriorate.

In the linear domain, separation and crosstalk are mathematically orthogonal. Phase and frequency errors that cause one to deteriorate will not affect the other. For example, phase or frequency errors in the composite signal channel will cause separation to deteriorate, but cannot affect crosstalk, since the stereo main and sub-channels are already separated in frequency and changes in phase or amplitude response in the composite channel cannot affect this frequency separation. Conversely, mismatches between the linear response of the left and right signal paths prior to the stereo encoder will cause crosstalk, but cannot affect separation.

Non-linearities in the composite channel can cause both separation and crosstalk to deteriorate because such errors cause harmonic and intermodulation distortion that introduce new frequencies into the baseband. These new frequencies are likely to inject power into a part of the baseband spectrum that will be decoded by the stereo decoder in spatial locations different than the locations of the original sound sources. Further, these new frequencies are perceived by the ear not as changes in spatial localization, but as highly offensive distortion.

This is somewhat analogous to "aliasing distortion" in a sample-data audio system. In such a system, any input frequencies greater than one-half of the sampling frequency (the "Nyquist frequency") are encoded with the wrong frequency: they "fold around" the Nyquist frequency and appear at the decoder as frequencies unrelated to the program material that produced them. The ear perceives this "aliasing" as offensive distortion.