High dynamic range passive mic with JACK, LV2, and Faust
Note
Decibels used here, are re: voltage amplitude.
Microphones are noisy.
That's not a big secret. Very low output amplitude of most microphone transducers ("capsules") available, means high white noise level added to output signal by the following circuit, as well as high EMI noise on connecting cable.
Microphone noise gets way more frustrating if one tries to use some analog or digital AGC and/or digital compressors to somewhat rectify level variations: they both are considerably increase noise at silent parts of vocal (when noise already most noticeable). While for pure electronic music there is infinite SNR, so no any noise to increase by compressors, but, mic signals are opposite for that, sadly.
Large amount of ways to fight with both types of mentioned noise sources were invented, and they all can be divided to passive and active approaches (and with rocket priced setups, both are combined; we will omit this case here). Let us also omit here noise gate approaches like DNR from compact cassette epoch, as they and their limitations are well known already.
Active are "simple" (and cheap today), and they are "just" an LNA, placed near capsule, then outputs amplified signal in single-phase or differential form and with low or controlled output impedance. Indeed, there is a power supply required for that, but, most high quality microphone input standards are does not offer power output line (country dependent). There is another drawback: active circuit adds noise itself, so, there is a balance between added and rejected noise. Also, we do not talk here about added non-linearity distortions, dynamic range limitation, power line collected noise, as these can be eliminated more or less effectively today.
Let's take a look at passive one. Despite of its name, it is a way more expensive, due to it requires high quality wideband transformer. The key parameter of transformer is windings inductance. That means either very high magnetic permeability yet wide frequency band core (often impossible to obtain today); or, regular ferrite, but, to keep inductance, it will be very large (at least, way larger than you can find in some microphones). Largest possible size also unavoidable to keep both resistive and magnetic power losses (we have too little power to just lost it!).
Isn't "Larger size = lower losses" rule wrong and should be opposite?
No. Every electric, electronic, and radio frequency appliances, are suffered from this. Larger is better and more effective always. Don't worry if tiny power from membrane is not enough to drive palm sized transformer: it will.
Btw, however, there is not possible to make cavity resonators (= filters) for GHz and THz larger than wavelength allows, sadly (oh, sorry for off-topic).
Well, hopefully, there are two good news:
- Typical microphone (let's take regular SM58: it will be fake most probably, anyway we can't obtain correct one) have some space inside of its enclosure;
- Several transformers can be joined together with no losses, but only gains, to mimic larger one. And special winding is not required for that. So, we can use two easily available 600:600 Ohm with (1+1):(1+1) windings, of largest possible size, which fits the SM58 body.
Let's connect it as pictured:
How it looks using 26x20 mm sized core, 18 mm bobbin width 600:600 Ohm transformers:
Tip
Pins should be grinded out almost completely to fit the crank, but note that windings are extremely thin, this is most hard thing of our research.
As one can see, there is no more than 4x amplitude gain (when we use A & E pins). If one need more, then way more expensive custom transformers are need, and this setup will be even more sensitive to output cable capacitance (yes, this is some disadvantage of passive approach, and 150- or even 300-Ohms cables are recommended here, instead of regular 50/75-Ohms ones).
Now we can connect it to some ADC.
ADCs are noisy...
There is known approach how to reduce ADC noise at the expence of get only half possible sound channels 1. We will use same trick, but been even more smart: Mixing stereo to mono, we will not add, but subtract channels, which, giving us exactly same internal ADC noise power reduction, but, also creates perfect differential input.
Tip
Please read separately if you in doubt why and how differential signalling used and works.
Note
Musicians often calls it balanced (or symmetrical) wiring, it is same as differential one.
At other side, one may note that our schematic use exactly differential output. So it's time to connect it. We will use regular blue stereo line input, available on PC motherboards: while it is known they are perfect sound cadrs except ENOB, but it, looks like, enough for our microphone setup.
Note
Some motherboards are not exposes true stereo blue line input connector, but it often can be obtained via pink noisy mono mic input, using hdajackretask magic, please read separate research for it (TBD, work in progress).
Next step is tune up our line input to be less noisy yet non-scaling (24-bit ADC output is not altered by codec level scalers). Using alsamixer, we carefully tune both Line Boost and Capture sliders of Capture tab (F4) of alsamixer, when check using data flow statistics with jasmine-sa (F10 then F1) to get 24 bit data flow. Mine is Line Boost=100 and Capture=33. Line input itself can be unconnected when tune up bit depth using Capture control; but for set up best level and SNR using Line Boost control, it is worth to connect some clean sound to it and estimate SNR using our spectrum analyzer or other measurement.
Now it's time to add differential input plugin to our audio plugin host like Carla, and one may enjoy lowest possible THD+N with such a cheap components.
Low THD+N is important for reverbs, as they have some degrade of SNR. (Example: Simple echo have exactly 3 dB SNR degrade). So it's time to enjoy plenty of air with reverbs.
Well, that works fine!
But if mic is fake (output amplitude of capsule is low, or, capsule DC resistance is too hidh for that output), singer can note earlier on later that there is low dynamic range of our setup (or, which is same (*), still low SNR). Are we take all the best our setup offers, or time to think more?
(*) If we're not reached membrane resonance yet.
Best solution here is, exactly with loudspeakers world, is just to add extra microphone. But it is weird for singer to hangle this setup in hand, so it is "best" electrically only, and quite weird in practice. Another ideas?
When ADCs quantity is enough (in our case, if we have one more stereo ADC), i am use simple approach to extend dynamic a bit. Just feed a part (often a half of voltage) of input to separate ADC, and use this channel when main one is overloaded. This gives 3 dB extra overload margin. But also adds some amount of noise: 3 dB during the periods of overload of main channel, and 0 dB else. In terms of music, added noise are well masked with strong signal peaks at these moments. For sum of two sine waves, there will be ~0.6 dB total (long-term) noise added, so we won ~2.4 dB of measurable ("real") SNR or dynamic, or in simple words, have extra 3 dB for voice peaks while keep near same (for non-peaked voice part, is exactly same) noise.
The problem is resistive tap is rarely useful due to resistors itself adds noise. Recall these B and D transformer taps? Yes, you're got the idea.
Wut? 4-wire microphone? :-[ ]
It is 5-wire. But, yes, we will call it 4-wire here.
We will use 5(7)-pin DIN connector, some extra wiring, and one extra stereo ADC, for almost zero extra hardware price.
Note
Read here (TBD) to know if your motherboard (or pci-e sound card) have it. In short, ALC887 onboard codec, and ASUS Xonar card (ALC1220 codec) are have it, and are tested. While ALC897 or ALC283 are only one ADC, sadly.
Now we will need some gate to switch main and peak outputs. I've make it using Faust and faust2lv2 compiler. Comparator here is sharp and momentary (per-sample), but feel free to implement one with some transition band, if you find this approach useful but need to lower gating THD even further; while current is quite low.
Important
Tune Ratio parameter correctly.
It is run time knob, no need to change code. We have ratio estimate output meter for that, just give some non-overloading sound to mic. Then set this measured value (often near 0.5) with Ratio knob.
We offer also test signal source (using Faust, again) as sum of 420 and 440 Hz sine waves, with two fully differential outputs, main one and "taped" (reduced or peak) one; and peak output can be switched off by runtime button. Noise of equal levels are added to outputs, to emulate input noise of the following circuit (ADCs), which makes possible the measurements below.
Important
alsamixer should be used to set same settings for both ADCs.
Btw, one may note that float value storage is implemented in same way: It have momentary SNR of 25 bits (24 + sign) (for 32-bit floats), while way more long-time dynamic range. It is just full of these transition points each 3 dB. Each transition point changes noise by 3 dB, even if transition itself is small but still crosses the boundary. As all today's DAWs uses floats, this can be seen using output volume knobs offered by plugin host per each plugin, fine-tuning it near its 1.0 or 0.5 position, and see the output noise floor with sensitive VU meter or spectrum analyzer.
Here are test results with test signal source. One may recall that 3rd order IMD's extra frequency components (easily seen as extra blue peaks) are extremely sensitive to any overload as non-linearity. So one may exactly know if no IMD are introduced, when no extra peaks.
[Command]
jasmine-sa MightyMic:Out0 MightyMic:Thru MightyMic:ThruPeak -h 330,530 -d -60,0,6
Note
White ray traces on picture, are overlayed red and yellow.
One can see that we have extra SNR (lower noise floor) around, or more than, 2 dB at output of our plugin when it work at full its power (loud constant yelling), compared to using only one mic output with its max non-overloaded level. SNR will npt degraded but only better (up to 3 dB) when yelling reduces. These 2 dB are long-time value, and distributed like 0 dB at loudest yet non-overloaded peaks (where masked well, makes it all better) and 3 db between peaks.
Btw, why these 3 dB, if no peaks, just normal underloaded input? This is because we attenuate main channel by 3 dB (or more precisely, by ratio amount, like windings tap ratio) always, so noise are reduced by 3 dB along with the signal.
Such approach can be expanded by yet another extra stage for peaks, but we need 3rd stereo ADC and extra wiring (while DIN-7 connector and our transformers already allows it).
We have added some extra filters recently, together with bypass switch for each. As well as 2- or 4-wire mic selector was added, so now our plugin will work well with both regular and 4-wire microphones. It's worth to check .dsp source code for filters descriptions. Full instruction manual is WIP, due to final filters set is not settled yet.
Please let me know how you like new mains hum remover, especially if you have noisy (distorted) mains lines around, when single point notch filter not helps much. Note that we have run time fine-tuned value of mains frequency to reject. Base frequency value is changed at .dsp file.
Note that Notch control is three-position, and have scalepoints which describe current one. With Carla, value can be seen as tooltip during pressing Spacebar on mouse hover, or using mouse button. Or, use Set value... (E on hover).
Why not just use 1 at its full power, with regilar 2-wire mic and 4 (two stereo) ADCs?
It have 3 dB worse EMI noise due to lower signal levels, and like 0.9 to 1.5 dB worse SNR compared to our 4-wire mic plugin, but else it will work.
Same, yet use one of stereo ADCs at 1/2 level? Would give all the benefits, yet with 2-wire.
This one is closer. The problem is input capacitance of ADCs (and wiring). The 4-wire is best possible approach in terms of distributing our tiny power to system capacitances, but that does not mean others are not useful. It's worth to try, especially with our new spectral tilt correction (but amplifying high freq components also amplifies noise...).
Btw, system input capacitance is limiting factor for transformer gain: As it replicated backwards to the membrane capacitive load at same ratio as transformer winding ratio, so, the more ratio, the more spectral slope fall of high frequencies. More windings also can give high self capacitance (time to increase core size).
Happily, there are also good news. Most of mic transducer capsules, both branded and fakes, are increases efficiency at high freqs. Also, looks like, most of them are love, if not require, some capacitive load, like LP needle cartridges. So even with long cable, sound tend to be perfect and solid, and sometimes it even better to either add even more extra capacitance, or, add some negative slope with our frequency responce tilt correction knob.
Can this plugin be made even more strong, with same hardware?
We let our readers to answer it.
[Spoiler]
I think there is more than one way possible with same hardware.- When no 1st (main) stage overload, take not just it, but sum of both (main + peak). Shuold reduce noise ~1.5 times (~2 db) further.
- Use some trainable AI to "guess" lost peaks, when all stages are overloaded; but rarely possible with
Faust, i think.
I need even better SNR, maybe another passive hardware possible?
You will need to decrease windings resistance, as it is source of thermal noise. To keep inductance, this is possible either with very rare and fragile permalloy or similar core, of at least same size; or, larger transformer(s) need, will do not fit mic body any more. Pure silver wire winding also helps, but not so much.
Here is triple effect on SNR:
- Less thermal noise;
- Less power loss due to winding resistance;
- Less power loss due to core losses.
What if windings (and/or taps) are not perfectly balanced?
Our setup is highly immune to that. Some differential non-equivalency are easily rejected by definition of differential signals. Non-symmetric taps measured & agjusted using Ratio knob & scale.
Electric guitars?
Any analog input welcome! But if it is UNbalanced (Tip-Ring, or TS connector, like most electric guitars), you'll need to recompile, search for UNbalanced at code. Obviously, to take advantage of 4-wire magic, transformer is required, and it'll then provide balanced output as a bonus.
Disadvantages?
Both stereo ADCs should form a perfect tandem, in terms of timings (relative skew and wander of it), as well as already mentioned amplitude and preamp gain match. (However, it is true for every music work station). It is known that for JACK setup and ALCxxx codecs, and alsamixer, this is true; but your setup can differ, like "multi-channel soundcard out of el-cheapo consumer cards", and it is not easy to create test set for it to confirm its ideal behaviour. Hopefully, for voice, some skew is acceptable (think about compact cassette tape head misaligned a bit).
Amplitudes mismatch (imperfect Ratio value) can create some distortions at the points where gate flip flops. This is mostly fixed with included LPF. Probably, there is automatic calculation of Ratio possible, but not easy. It should be like step error detector, integrator, and feedback. It can rectify possible skew also. Ask me if you really need it.
Are these quite few dB SNR gain really counts?
If you found this plugin, you know how much every single dB costs, especially at the input of DAW, don't you?
- LNA: Low noise amplifier.
- SNR: Signal to Noise ratio.
- LPF: Low pass filter.
- EMI: Electromagnetic interference, cables pickup noise due to it.
- ADC: Analog to Digital converter, like input of PC or sound card.
- ENOB: Effective number of bits, "real" (heavy-loading) ADC/DAC SNR measure, unlike of just silence noise floor.
- AGC: Auto gain control.
- Mixing: Audio people uses it as (linear) adding but not multiplying.
- DNR: Dynamic Noise Reduction (Analog noise gate).
- THD: Total harmonic distortion, most often due to non-linearity. Sum of all IMDx.
- IMD: Intermodulation distortion. Most important is IMD3 part, because it falls to in-band.
- EIN: Equivalent input noise 2.
- DAW: Digital audio workstation, in our case it is realtime DSP chain.
- TBD: To be displayed.
- WIP: Work in progress.
This research text description, together with its inlined pictures, are licensed under Creative Commons Attribution 4.0. You are welcome to contribute to the description in order to improve it so long as your contributions are made available under this same license.
Included software is licensed as per LICENSE file.