Prom 32 - Audio Limiting (again) !

[Bryn]

In general, the Radio 3 engineers seem to work to an estimated 6dB headroom. A couple or so Proms this year have hit 2dB below clipping level. The engineers should indeed be able to judge the danger points from level check during rehearsals (and , of course, from experience). Upping the estimated headroom to, say. 9dB would be preferable, to my ears, to the use of hard dynamic limiting.

[Serial_Apologist]

In general, the Radio 3 engineers seem to work to an estimated 6dB headroom. A couple or so Proms this year have hit 2dB below clipping level. The engineers should indeed be able to judge the danger points from level check during rehearsals (and , of course, from experience). Upping the estimated headroom to, say. 9dB would be preferable, to my ears, to the use of hard dynamic limiting.

Which, I take it, would be closer to the dynamic range of most classical LP recordings in the 1960s, as was the case with R3 broadcasts at that time, as I remember it.

[rank_and_file]

johnb

Noted about the “thickness” and I have the same problem with Cool Edit Pro and, sure, the less time you show, the thinner the line/dots.

The lousy, lossy remark was supposed to be my attempt at humour.

The proms repeat was FM this afternoon. I was attempting to verify whether the Optimod is always on after your remark in post 18 that they can tweak (reduce or filter) its compression during the day.

[johnb]

The proms repeat was FM this afternoon. I was attempting to verify whether the Optimod is always on after your remark in post 18 that they can tweak (reduce or filter) its compression during the day.

Optimod would have been in operation. However, if you were recording from (analogue) FM the -3dB you mentioned would solely be due to how you set your recording levels.

[Bryn]

Which, I take it, would be closer to the dynamic range of most classical LP recordings in the 1960s, as was the case with R3 broadcasts at that time, as I remember it.

The theoretical signal to noise ratio of a 16 bit digital audio stream is 96dB. However, for practical purposes 90dB is the generally accepted working ratio. That's a good 25 to 30dB greater than that offered by vinyl. It is often recommended to enthusiasts (rather than professionals) that a headroom of around 12dB is allowed for when recording live, following an exploratory level check during a rehearsal. Even with that great a headroom, the S/N will still be at least 13dB greater than the best that vinyl has to offer.

[Serial_Apologist]

The theoretical signal to noise ratio of a 16 bit digital audio stream is 96dB. However, for practical purposes 90dB is the generally accepted working ratio. That's a good 25 to 30dB greater than that offered by vinyl. It is often recommended to enthusiasts (rather than professionals) that a headroom of around 12dB is allowed for when recording live, following an exploratory level check during a rehearsal. Even with that great a headroom, the S/N will still be at least 13dB greater than the best that vinyl has to offer.

Interesting that, and thanks, Bryn

[Resurrection Man]

Gosh, this brings back memories when I first started listening to Radio 3 all those years ago and long before Optimod. I didn't like the compressed dynamic range compared to a live performance but recognised the limitations of the transmission system. I had a multi-track tape recorder and as an experiment made an audio expander which was controlled by the signal from a spare track on the recorder. Two of the other tracks were left and right channels. I spent hours listening to the tape and as the music played applied my own control signal to the spare track so as to expand the dynamic range back to what I thought it should be. Sonically it worked...sort of...but took way too much time to prepare.

[johnb]

Out of curiosity I've just compared the iPlayer version of last night's VW4 (just the performance without the announcers, etc) with the Previn recording (analogue, remastered for CD).

After normalizing the Prom performance so that the highest peaks are at 0dB in order to match the Previn, the RMS levels of the two are:

Proms: -26dB
Previn: -20dB

The RMS is a measure of the average level, so the fact that the Proms is lower than the Previn indicates that that the Prom performance has a wider dynamic range.

By the way, the ambient noise levels of the hall/studio were -60dB RMS for the Prom and -67dB RMS and for the Previn. (Once again with the peaks for both @ 0dB.)

[Gordon]

The theoretical signal to noise ratio of a 16 bit digital audio stream is 96dB. However, for practical purposes 90dB is the generally accepted working ratio. That's a good 25 to 30dB greater than that offered by vinyl. It is often recommended to enthusiasts (rather than professionals) that a headroom of around 12dB is allowed for when recording live, following an exploratory level check during a rehearsal. Even with that great a headroom, the S/N will still be at least 13dB greater than the best that vinyl has to offer.

I would not take great issue with any of that. In a digital recording or transmission channel there is no "analogue " noise to worry about so that when there is no input - ie silence in a studio - the machine/channel contributes no "noise" such as that an analogue machine would. A quiet digital channel is truly silent. The only noise there is is very low level analogue noise from the microphones and their amplifiers at one end and the analogue circuits in the playback system. Typically this is way down around -120dB or better.

The equivalent digital audio impairment to noise is not strictly speaking “noise” in the old analogue sense of an independent, additive unwanted accretion. It is caused by the signal itself when it is quantised and so is not a constant background effect. It is called quantising “distortion” [QD] rather than noise because that is a better description of its origins. In extreme cases QD can be horrible to hear and the number of bits is often chosen to avoid this rather than meet a nominal theoretical noise specification that has equivalence with an analogue channel. IOW S/N or /QD is not the whole story.

One test the BBC used when developing digital audio was to take a loud chord played on a piano and let it decay in a quiet studio [you could build an electronic box to do the same thing of course]. As the sound dwindled into silence at low levels a harsh grating noise could be heard that indicated that the sound was now at a similar level to the quantisation step defined by the number of bits used and this of course meant that this low level sound was actually using 1 or 2 bits only. Add enough bits until that grating noise didn’t happen any more and you have a workable scheme.

Provided the wanted audio exercises a large number of quantisation steps – which are always assumed to be linear ie all the steps are the same size throughout the audio amplitude range – this distortion does have the statistical properties of random noise and so can be treated as such. The theoretical power of such random QD is [q squared]/12 where q is the step size, assumed constant – ie linear PCM. One other way to do the coding and minimise the number of bits is to use non-linear PCM codes so that the steps are smaller near small signal amplitudes and larger for large signals. This works well and is used for digital telephones but isn’t used much in audio for a number of reasons. It does mean that the QD varies with amplitude but can be hidden/buried for large enough signals yet inaudible for small ones.

For a peak to peak amplitude range [ie between upper and lower clipping levels] of 1 volt the step size for N bits is 1/2^N volts. So if you do the arithmetic the Signal to QD ratio is SQRT12 times 2^N. Turn this into logs to get dBs and the answer for 16 bits is 107 dB. So the S/QD is a direct logarithmic function of the value of N, it increases by 6dB for every extra coding bit. 24 bits gives 155dB which is starting to get silly given that the range between the thresholds of pain and hearing is a lot less. 24 bits AES and high sampling rates allows among other things, Noise Shaping [Sony's SBM for example] which is number crunching of the final target 16 bits for CD to make it sound and measure better than the 16 bits would imply. All it does is shift some of the QD out of the audible band so it sounds better - like 16 bits sounds like 18!!!

This 107 is greater than the usually quoted value of 96dB [as given in the quote above and found in many CD booklets] because of various factors one of which is to do with the linearity of the A/D converter, ie how equal are all the q steps throughout all 2^N? In the early years of CD and PCM they weren’t that linear and had other quirks too. Another reason is that decoders – D/A converters in consumer players - were not that linear either and of course it’s the combination that we listen to. Some early CDs didn’t sound well because of this. For N=14 [as used in NICAM] the value is 96dB. No equivalent to pre-emphasis is used in digital PCM [there is a flag in the CD system that allows it but no one much uses it, if at all]. One can apply subjective weighting to these numbers which generally improves apparent S/N but this is all based on analogue experience and so is not used in the digital case. To match an analogue S/N of say 65 dB theoretically only 9 bits are required!!

For my youthful sins I used from time to time to align analogue stereo Studer machines and following the handbook one could get low 60s dB unweighted out a good finely tuned one without Dolby A, 10dB or so more with it.

The moral of this boring tale is that there are many practical and engineering issues behind the numbers we bandy about in casual conversation about digital audio especially when we try and compare the two systems using the same measurement terms. In the case of digital headroom it isn’t just allowing room for error, avoiding clipping or for meters etc it is also for the properties of audio distortion when coding PCM. FM is another story entirely!

[PhilipT]

Gordon - I am a little surprised that the word 'dither' doesn't occur anywhere in such a long and detailed post. My understanding of what you call QD is that, without dithering, quantisation error, or quantisation noise, is signal-dependent, and any signal-dependent noise is perceived as distortion. Add in dithering, and although you now have more noise, it's not signal-dependent in the same way and so isn't perceived as distortion. Yes? No?

[Gordon]

Gordon - I am a little surprised that the word 'dither' doesn't occur anywhere in such a long and detailed post. My understanding of what you call QD is that, without dithering, quantisation error, or quantisation noise, is signal-dependent, and any signal-dependent noise is perceived as distortion. Add in dithering, and although you now have more noise, it's not signal-dependent in the same way and so isn't perceived as distortion. Yes? No?

In a word yes!! I did not want to write textbook and so left some things out as you point out. The idea is to make sure that low level signals don't dwell anywhere and so a small amount of noise [which is so low as to be almost imperceptible] is added to the audio and this has the effect of reducing the number of bits needed to get over the harsh QD at low amplitudes. Another way of looking at it is that the rich harmonic structure of the QD is modulated by this noise and is therefore "smeared" and made less obvious. It doesn't take the QD away, just alters it.

All in all it helps improve things. In some cases this noise comes naturally from microphones and desk electroncs so may be present anyway! One reason why it was explored in the early days [back in the late 60s/early 70s] was difficuties in building low cost accurate A/D and D/A converters having lots of bits. Dither alleviated that problem, it's not so critical with AES 24 bit samples. I remember measuring A/Ds with nominally 16 bits which when measured came in at less than 13. They weren;t very good!!

[OldTechie]

For my youthful sins I used from time to time to align analogue stereo Studer machines and following the handbook one could get low 60s dB unweighted out a good finely tuned one without Dolby A, 10dB or so more with it.

You could never get fantastic figures like that out of an EMI TR90 (or a BTR2 - but they were unbeatable for editing.) The one in Manchester BH Control room also had a habit of suddenly spooling off the tape mid-programme without anybody near it, which was more noticable than the noise level or flutter. The flutter figures were not up to Studer standards either. My father (a violinist in the BBC Northern Orchestra) could tell whether a programme was live, off disc or off tape within a couple of seconds because he could hear the flutter on every tape machine (there were not many Studer C37s around) and the wow on every LP. In my view the most important advatages of CD/digital over vinyl is that I don't sit there hearing the pitch going up and down at roughly 2 second intervals. I blame my father for making me too sensitive to this issue. I do have a couple of discs that don't suffer this - but they are original cuttings so the hole really is in the middle of the groove location and the metal backing plate keeps them flat.

Ah, but the topic was about levels and dynamic range compression. There is an interesting document at http://tech.ebu.ch/docs/techreview/t...ss_Camerer.pdf where the section 'The origin of the problem – the “loudness war”' describes some of the issues. I look forward to seeing a mechanical implementation of the meter with a bendy needle.

[Gordon]

#42: Well those machines were a bit older and mono [?] although I think there was a stereo version of TR90 and a portable one too. The BBC versions of the EMI machines were specials too - painted green - and TV had slightly different ones again. Remember that BBC stereo didn't really take hold until well into the 1960s and TV even later still. BTR/2 started life in about 1953 after BTR/1 which was a quickly cobbled version of the German Telefunken machine. It has a lot of problems - it had no HF bias - but got tape started at EMI in about 1948/9. BTR/2 came in as the mono workhorse and stayed for many years - built from recycled battleships! BTR/3 was the stereo version which started life in about 1958/9, recording before that had been done on modified BTR/2 chassis. I think there was at least one BTR/3 at Kingsway until well into the 60s - that and an old REDD 37 mixer, the flat topped one, possibly the prototype of '59!!

The Studer A80 2 track was the one I played with around 1980 was it, not sure, so it's no surprise they measured better. We forget that all those classic recordings at Abbey Road and Kingsway Hall were made on machines that today would not be considered very good!!

Thanks for reminder about that loudness article, I had forgotten about it!! I used to read all those EBU TRs in those days!!

[OldTechie]

I have listened to the programme repeat and looked at the compression on the iPlayer version. It is, I'm sure, the same as the one johnb originally reported. By my measurements the signal is limited at 10db below digital full scale. So, I'll explain why that is the level I would expect to be used for a limiter - merely for general interest.

Many years ago the BBC invented the peak programme meter (PPM) to measure the sound level for broadcasting. It has become the standard for all british broadcasters. It does not have a techical scale marked in dB, but simply the numbers 1 to 7. Between each number there is an interval of 4db (except for the old valve ones which had 6db between 1 and 2.) Number 6 is the maximum allowed programme level. The meter can read higher than that so if the sigmal is too large, the operator has some chance to turn it down by the correct amount. Number 4 was originally at the vertical position, but it moved a bit off when the first transistorised meters were introduced because they could not get the logarithmic convertor quite right. Number 4 is the line-up level. So line up tone should read 4. That means the maximum signal is 8db above line up and the readable range is only 20 dB. Modern mixers will have addtional meters with other scales, but the PPM is still the reference standard.

The PPM is not a true peak programme meter - it has an attack time of 10 milliseconds. So very short peaks will not register on it.

Across Europe, different brodcasters use meters similar to the BBC PPM but with small differences. It is common to set the peak level as 9db above line up level.

The EBU has specified a line-up tone level for digital links and transmission as 18db below digital full scale. That means that the BBC peak level is 10dB below digital full scale. If there is no limiter and a BBC transmission does not read more than 6 on the PPM, there will typically be peaks that are higher but do not read on the PPM because they are too short. The paper I mentioned in post42 suggests that the EBU allowed 4.6 dB for this effect. Of course there can always be accidental sending of levels above PPM 6 for a moment, mainly on speech.

They definitely put a hard limiter on this concert - and set it to limit at 6 on the PPM. The limiter would not be built to have the PPM effect of missing the short peaks.

My memory of the Havergal Brian concert was that it sounded more like a speech compressor in circuit, with a bit of gain to compress to a target level of PPM 4 to 5 and an absolute limit at PPM 6. I thought it had a fairly short recovery time typical for speech, because the pumping effect was very obvious.

This concert sounds to me more like a straight limiter, doing nothing until the level is within a dB or so of PPM 6. It is set with a very long recovery time of 10 to 15 seconds, so pumping is not so objectionable.

Listening at the section around 01:22:00 on johnb's waveforms we have a tremendous crescendo. It has multiple sections as more and more instruments and choirs join in. As each new sound is added, the limiter reduces all the existing ones to keep the level absolutely constant, thus fairly firmly destroying the music. This is followed by a solo voice in grey below.




You can see the level of the voice increasing when the orchestra and choirs stop. The first section of the voice comes in loud to compete with the orchestra, and it seems he gains the ability to compete with the level of the whole orchestra within one breath!. Actually it is the limiter slowly restoring the gain.

I estimate that the limiter was applying in excess of 8dB of hard limiting at the end of the crescendo (if you can call it that when it is all at the same level.)

I think this is even worse than the Havergal Brian event last year.

[Stunsworth]

Listening to the Mahler tonight the dynamic range seemed fine - I only heard the first half of the concert.

The software I use to play the HD internet stream shows the dynamic range of the signal. Unfortunately this isn't what I normally think of as dynamic range, which is the difference between the quietest part of the recording and the loudest. Instead it's the difference between the RMS value of the entire recording and the peak value. In the Mahler this was around 34dB. A 'normal' concert is usually somewhere between high twenties and very low thirties - typically 31-32.

The peaks seemed to be coming in at around -5dB, which I think is very good.

Perhaps the problem is with the engineers erring on the side of caution with unfamiliar forces. Regrettable though that may be.