First of all, congrats on your Technical Grammy Award this year! Daniel, you’ve once started DSP developments during the early days of digital audio. What was the challenge to that time?
Thank you very much, Herbert.
Yes, I started doing digital audio back in 1979 when I joined Studer-Revox. In that year Studer started their digital audio lab with a group of newly employed engineers. At that time there were no DSPs or CPUs with enough power to do audio signal processing. We used multiplier and adder chips from the 74 chip series and/or those large multiplier chips they used in military applications. The “distributed arithmetic” technique we applied. Very efficient, but compared to today’s processors very inflexible.
The main challenges regarding audio applications were:
- A/D and D/A converters had to be designed with audio in mind.
- Digital audio storage had to rely on video tape recorders with their problems.
- Signal processing was hardware coded, i.e. very inflexible.
- DAWs as we know them today have not been feasible due to the lack of speedy processors and the lack of large harddisks. (The size of the first harddisks started at about 10 MByte…).
- Lack of any standards. Sampling frequencies, wordlengths and interfaces have not been standardized back then.
Later the TMS32010 DSP from TI became available – a very compromised DSP, hardly useable for pro audio.
And a bit later I was able to use the DSP32 from AT&T, a floating point DSP which changed a lot for digital audio processing.
What makes such a converter design special in regards to audio and was the DSP math as we know it today already in place or was that also something rather emerging to that time?
The A/D and D/A converters back then had the problem that they either were not fast enough to do audio sampling frequencies (like 44.1 kHz) and/or their resolution was not high enough, i.e. not 14 Bits or higher.
There were some A/D and D/A modules available which were able to do digital audio conversion, but those were very expensive. One of the first (I think) audio specific D/A converters was the Philips TDA1540 which is a 14 bit converter but which has a linearity better than 14 bit. So we were able to enhance the TDA1540 by adding an 8 bit converter chip to generate two more bits for a total of about 16bits conversion quality.
The DSP math was the same as it is today – mathematics is still the same, right? And digital signal processing is applied mathematics using the binary numbering system. The implementation of adders and multipliers to some extent differed to today’s approaches, though. The “distributed arithmetic” I mentioned for instance worked with storage registers, shift registers, a lookup table in ROM and an adder / storage register to implement a complete FIR filter. The multiplication was done via the ROM content with the audio data being the addresses of the ROM and the output of the ROM being the result after the multiplication.
An explanation is given here: http://www.ee.iitm.ac.in/vlsi/_media/iep2010/da.pdf
Other variants to do DSP used standard multiplier and adder chips which have been cascaded for higher word-lengths. But the speed of those chips was rather compromised when comparing to today’s processors.
Was there still a need to workaround such word-length and sample rate issues when you designed and manufactured the very first digital audio equipment under your own brand? The DS1 compressor already introduced 96kHz internal processing right from the start, as far as I remember. What were the main reasons for 96kHz processing?
When I started at Studer the sampling frequencies have been all over the place. No standards yet. So we did a universal Sampling Frequency Converter (Studer SFC16) which also had custom built interfaces as those haven’t been standardized either. No AES/EBU for instance.
Later when I started Weiss Engineering the 44.1 and 48 kHz standards had already been established. We then also added 88.2 / 96kHz capabilities to the modular bw102 system, which was what we had before the EQ1, DS1 units. It somehow became fashionable to do high sampling frequencies. There are some advantages to that, such as a higher tolerance to non-linearly treated signals or less severe analog filtering in converters.
The mentioned devices were critically acclaimed not only by mastering engineers over the years. What makes them so special? Is it the transparency or some other distinct design principle? And how to achieve that?
There seems to be a special sound with our devices. I don’t know what exactly the reason is for that. Generally we try to do the units technically as good as possible. I.e. low noise, low distortion, etc.
It seems that this approach helps when it comes to sound quality….
And maybe our algorithms are a bit special. People sometimes think that digital audio is a no brainer – there is that cookbook algorithm I implement and that is it. But in fact digital offers as many variants as analog does. Digital is just a different representation of the signal.
Since distortion is such a delicate matter within the design of a dyncamic processor: Can you share some insights about managing distortion in such a (digital) device?
The dynamic processor is a level controller where the level is set by a signal which is generated out of the audio signal. So it is an amplitude modulator which means that sidebands are generated. The frequency and amplitude of the sidebands depend on the controlling signal and the audio signal. Thus in a worst case it can happen that a sideband frequency lies above half the sampling frequency (the Nyquist frequency) and thus gets mirrored at the Nyquist frequency. This is a bad form of distortion as it is not harmonically related to the audio signal.
This problem can be solved to some extent by rising the sampling frequency (e.g. doubling it) before the dynamic processing is applied, such that the Nyquist frequency is also doubled.
Another problem in dynamics processors is the peak detection. In high frequency peaks the actual peak can be positioned between two consecutive samples and thus get undetected because the processor only sees the actual samples. This problem can be solved to some extent by upsampling the sidechain (where the peak detection takes place) to e.g. 2 or 4 times the audio sampling frequency. This then allows to have kind of a “true peak” measurement.
Your recent move from DSP hardware right into the software plugin domain should not have been that much of a thing. Or was it?
Porting a digital unit to a plug-in version is somewhat simpler compared to the emulation of an analog unit.
But the porting of our EQ1 and DS1 units was still fairly demanding, though. The software of five DSPs and a host processor had to be ported to the computer platform. The Softube company did that for us.
Of course we tried to achieve a 1:1 porting, such that the hardware and the plugin would null perfectly. This is almost the case. There are differences in the floating point format between DSPs and computer, so it is not possible to get absolutely the same – unless one would use fixed point arithmetic; which we do not like to use for the applications at hand.
The plugin versions in addition have more features because the processing power of a computer CPU is much higher than the five (old) DSPs the hardware uses. E.g. the sampling frequency can go up to 192kHz (hardware: 96kHz) and the dynamics EQ can be dynamic in all seven bands (hardware: 4 bands maximum).
Looking into the future of dynamic processing: Do you see anything new on the horizon or just the continuation of recent trends?
We at Weiss Engineering haven’t looked into the dynamics processing world recently. Probably one could do some more intelligent approaches than the current dynamics processors use. Like e.g. look at a whole track and decide on that overview what to do with the levels over time. Also machine learning could help – I guess some people are working in that direction regarding dynamics processing.
From your point of view: Will the loudness race ever come to an end and can we expect a return of more fidelity back into the consumer audio formats?
The streaming platforms help in getting the loudness race to a more bearable level. Playlists across a whole streaming platform should have tracks in them with a similar loudness level for similar genres. If one track sticks out it does not help. Some platforms luckily take measures in that direction.
Daniel, do you use any analog audio equipment at all?
We may have a reputation in digital audio, but we do analog as well. A/D and D/A converters are mostly analog and our A1 preamp has an analog signal path. Plus more analog projects are in the pipeline…
Related Links
- Weiss Engineering Ltd.
- interview series (1) – Fabien from TDR
- interview series (2) – Nico from BigTone
- interview series (3) – Tony from Klanghelm
- interview series (4) – Bob Olhsson
- interview series (5) – Dave Hill
- interview series (6) – Christopher Dion
- interview series (7) – Dave Gamble
- interview series (8) – Sascha Eversmeier
- interview series (9) – D.W. Fearn
- interview series (10) – Vladislav Goncharov
- interview series (11) – Andreas Eschenwecker
Variety Of Sound 
[…] interview series (12) – Daniel Weiss […]