The diagram of the reverberation network is as followed:

Input bandpass filter

There are several reasons to use bandwidth control here.

• Normal people can hear a frequency range from 20Hz to 20kHz.
• Virtually any natural environment has low-pass frequency response.
• In a digital system, highest signal frequency must be limited below half of the sampling frequency to avoid aliasing or frequency fold-over.
• Care has been taken to assure BIBO (bounded input bounded output) stability of the system, but for any kind of digital signal processing we must further avoid overflow, that is the value of the signal exceeds the limit of DSP hardware and software. For audio signal specifically, overflow means clipping which is very clear and unpleasant to ears. To avoid clipping, we must scale the audio signal to valid range, [-1,1] in VSS environment. Maximum signal amplitude should be scaled to 1 or -1. Thus, the DC component in audio is undesirable. It contributes little to the perception of sound, yet it raises the maximum signal amplitude, therefore demanding bigger scale factor and making some subtle aspects of the sound inaudible.

To address the above issues in a realtime process actor, we choose to use a bandpass filter with 0 gain at DC and half sampling frequency to preprocess the signal to be reverberated. Such frequency response satisfies all the requirements above. A 2nd order IIR filter is used for the purpose, its I/O function, transfer function, zero/pole distribution and frequency response is shown in the following graph.

As Z-plane zero/pole distribution of the filter tells us, the two zeros at 1 and -1 force the frequency response at 0 and fs/2 to be zero. The maximum frequency response is always 2 regardless of the position of the pole, however that position determines the frequency at which the response maximum occurs. It follows from straightforward derivation that the "center frequency" and the pole has the following relationship: cos(2*PI*fc/fs) = pole. Some pairs of fc and pole values are given below, with fs = 22,050Hz:

The closer the pole is to 1, the lower center frequency will be. When the pole is at 1, it cancels out with the zero, thus reducing the filter to 1st order FIR, which is not allowed. Similar discussion holds for negative pole position. When the pole is at origin, the IIR filter turns into a 2nd order FIR filter. Positive pole is preferable, because then the response peak is set in low frequency range to make the sound "warm". But pole > 0.5 leads to distorted output because of the unlinearality at low frequency is too sharp. Same thing happens when pole < -0.5. For some more information about filter, please consult Filter Actor documentation.

Early reflections

Reflections during about 20~80ms after sound is triggered are heard together with the direct sound as one single auditory event. An FIR filter is used to generate these early reflections.

Comb filters

Comb filters are used to imitate both time and frequency responses of enclosure reverberation. In time domain, comb filter produces exponentially decaying echoes of input. In frequency domain, the comb-shape frequency response resembles the natural modes of enclosure. The problem with comb filter is that the distances between adjacent echoes and frequency modes are decided by the single parameter of delay. Bigger delay leads to more scattered echoes and closer frequency modes, and vice versa, while the enclosure reverberation yields both dense echoes and close frequency modes. Because dense echoes can be achieved by all-pass filters in the next stage, and distant frequency modes yield unpleasant metallic sound, the comb filter delays should be set in the range of 50~80ms.

Output of one comb filter has very distinct periodicity which is called "flutter". This can been seen from its impulse response, or heard by setting the reverberator to have only one comb filter. To avoid the periodicity, several comb filters with delays in prime number can be paralleled. The frequency response of paralleled comb filters will be the sum of the individuals, imitating the overlap of natural modes of enclosure.

The reverberation time is decided by delay and gain. The standard measure of reverberation time is t60, the time at which reverberation level falls to -60dB or one thousands of the initial sound.

Another character of room and hall is that t60 is frequency-dependent. The t60-frequency curve is fairly complicated and distinct for individual enclosure, but the general shape is low-pass. The same 1st order IIR filter for bandwidth control can be added into the feedback loop of the comb filter to produce a smooth low-pass t60-frequency response. Given the ratio of t60(0) / t60(fs/2) (set by damp ratio), delay, loop gain and low-pass gain can be calculated.

All-pass filters

The name "all-pass" comes from its constant frequency response. But its phase response is quite complex, and the group delay is in comb shape. All-pass filter along can make pretty decent reverberatino because of the dispersive phase response. Here, they are used to increase echo density. The impulse response is similar to that of comb filter, so a small delay, e.g. 6ms can make very dense echoes, while no "color" is added into the sound as by the comb filter, because of the constant frequency response.

Output low-pass filter

A 1st order IIR filter is used to low-pass the reverberation to reduce the high-frequency ringing generated by comb filters. A single parameter BW is used to control the frequency response of the filter, thus the "brightness" of the reverberation. Frequency response at DC is always 1, and BW/(2-BW) at fs/2. Thus the bigger BW, the wider frequency band is passed through.

Reference

Salute! The algorithm implementation and program structure of STK reverberator was shameless copied to reverberator actor design.

Problems:

• Delay is in number of samples, not time, and isn't related to filter gain. So if D is big (e.g. 10000), significant periodicity can be heard. D = 2137 for all STK instruments.
• No early echo: For fast attack/decay sound like chime, there's a distinct dry part and wet part of the reverberated sound.