posted 28 March 2003 11:46            

srsencoding.kym gaincurve6db.aif

One is a reverb controlled version that uses gain compression on the S channel and gain expansion on the M channel with supression for artificial reverb. The other uses only gain compression on the S channel and no reverb control.

Both sounds have been carefully re-engineered:

1. The filterbanks have been properly tuned to present a flat passband over the entire spectrum when presented with a pink (1/F) spectral slope. This involves (a) attenuating each of the interior passbands with respect to the outer shelf filters on each end of the spectrum, and (b) using proper cutoff frequencies in the constituent Butterworth filters.

The gain balancing needed differs according to the number of filters in the filterbank, and on the spectral width of each passband. For the simple 6 band filterbank the gain adjustment is -3 dB on the interior passbands. For the 7 band filterbank it is -6 dB. These filterbanks were systematically checked for correctness using an external spectrum analyzer.

2. The HRTF correction on the M channel was derived from measurements made at the MIT Media Laboratory. These SRS encoders are both constructed to assume either headphone listening (my norm) or side mounted speakers. Hence, our ears are presented with side information for all sounds. In order to derive the frontal characteristics for the M channel, one must apply the HRTF for front placed sounds inverse filtered by the side listening HRTF. This was done by me, and an FIR filter was derived to produce apparent frontal sounds while listening to sound delivered from the side. (Is that clear as mud?)

While the gross characteristics of this filtering is similar to the crude model offered earlier with 3 notching filters, the actual correction spectrum is more complex, and exhibits a fair amount of peaking above 10 KHz. In other words, in order to make side presented sounds appear to our minds as though we hear them presented from in front of us, we need to notch out a broad passband from around 500 Hz to about 4 KHz, a deep notch at 8 KHz, and some peaking above 10 KHz.

[One result of this is an enhancement of sibillant sounds when heard from the front. I find that interesting... it suggests that our hearing is biologically adapted to be more sensitive to speech when presented from the front of us -- where we can also watch the person talking... Is that interesting? or what?!]

The FIR filter coefficients were derived from data sampled at 44.1 KHz. Hence this HRTF correction filter will only be correct when the Capy is running at 44.1 KHz sample rate. Changing to 48 KHz moves the key notch and peak frequecies lower by the ratio 44.1/48. (Not very large and probably unnoticeable).

3. Gain balancing on fixed point DSP's like the Motorola 56K's in the Capy is very important in order to avoid saturating any of the interior processing blocks. This has been carefully re-engineered for correct behavior.

4. Compressors and FIR filters impose delays in their processing. In order to maintain phase correctness in the recombined M and S channels, we need to insert compensating delays in the opposite channel. This has been done.

5. Kyma compressors are used in the S channel filterbanks. The maximum gain boost permitted is 12 dB. Hence, the thresholds have all been set to -30 dB, and the compression ratio is 5/3. Maximum gain boost occurs at and below -30 dBFS, and drops to unity gain at 0 dBFS.

Complementary to this gain compression in the S channel, the more sophisticated Sound uses a reverb supression system that gradually boosts the gain of the M channel in each subband, in proportion to the energy present in the corresponding S subband. To match the gain characteristics of compressors, the gain enhancement curve has been recomputed. That is the little .AIF file provide here...

6. Algebraic functional analysis of the overall processing showed many simplifications with the originally derived M and S channels. There is no need to recombine to intermediate L & R channels until the very final stage.

For my two cents, both Sounds (the expensive one and the cheapo one) sound pretty much the same to me, which suggests that the added complexity for artificial reverb supression is mostly unnecessary. And both Sounds do a remarkable job of stereo field enhancement. Control !K1 varies the amount of M channel present -- increasing it moves the soundstage more to the front. Control !K2 varies the side channel enhancement. Increasing it widens the field.

- DM