Suzu 12 — All Discrete Component Guitar Amplifier for 2023

In January & February 2023, I built 4 smaller size versions of the GAA -12 Practice Guitar Amp that we call Suzu. My design goals included fresh & unique circuitry, all discrete components, all split supply amplifiers plus a clean & simple signal path. I'll show my 4th and best version. Serving as my upstairs guitar practice amp, I specifically designed it for the T-style or Fender Telecaster ™ guitar and a 10 inch speaker.

 

The overall tone flavor of this amp harkens the Gibson GA-50. I avoided a mid range tone control and deep middle frequency scooping. If you boost the bass and treble controls, you do create some mid scooping but it's low Q and quite subtle compared to old black panel Fender guitar amps of lore.

Note this was originally published as an update on Dec 4, 2022.  I added much new content and then re-published it on Feb 20, 2023.

— C O N T E N T S —

1. Preamplifer 1

2. Preamplifier 2  + design spreadsheet to download

3. Power supply

4. Power Amp - PA -

5. Speaker

6. Miscellaneous Photos

Click here for my Guitar-Related Index

 

1. Preamplifer 1

Above — First preamplifier schematic. Preamp 1 and 2 connect directly to the main DC power supply with no voltage regulation to get the maximum possible rail to rail AC guitar signal. To subdue power supply ripple and to isolate the preamplifier from the PA supply, a ripple filter feeds the preamp stages DC. I employed further RC low pass filtration on each stage to enhance ripple & noise rejection in this single coil pickup purposed guitar amplifier.

 

The input 12K stopper resistor and capacitor form a low-pass filter to prevent AM radio detection. Eleven volt zener diodes clamp excessive signal amplitude from popping the input. This cold/dry Winter [coldest temperatures every recorded here in 2022] caused a lot of electrostatic buildup and discharge. Shocking. Sadly, empirically, I learned that static discharge can easily blow up front end circuitry & that all guitar amps need input protection.

A low-noise JFET with 1 megohm gate resistance provides a high input Z to the guitar pickup(s) and drives an emitter follower so the following stage tone circuit sees a low output impedance. The JFET voltage gain is set to about 3.3 with the 2K7 gain setting, source degeneration resistor. I normally set my maximal input stage voltage gain between 3 and 5. The JFET source current = 1.3 mA. The emitter follower collector current = 2.4 mA. When AC coupled to a 1K resistor load, the JFET + emitter follower can pass a 1 KHz signal with a magnitude of ~8.6 Vpp before it starts to clip.  Lovely.

I prefer to bias each preamp block with a signal generator and DSO running and temporary resistor load AC connected. I strove to run the lowest possible current for each stage along the signal path. I chose the FET drain resistor value by temporarily substituting in a 10K potentiometer while adjusting it to get the highest clean signal swing at my bias point and then swapped in the nearest standard 1% metal film resistor. Almost every resistor is a 1% metal film and I happily grew my metal film resistor collection this Winter.

 
2. Preamplifer 2

Preamplifer 2 functions as the heart of my amplifier.  I spent a month on this stage alone. Most of my discrete circuit designs resembled op-amps: For example, differential input, a voltage amp, plus a low Z output, however, but I found it wasn't necessary since I was not pursuing a ultra-linear preamp design. Some guitar amps built with op-amps and careful local + global feedback are said to sound sterile or too HiFi.  Perhaps this rings true?

I did not get hung up on an ultra-linear signal path, rather tried my best while avoiding the emitter-coupled pairs found in op-amps plus many other analog ICs. It's fun to bias discrete transistors, calculate & measure things like input impedance, or the feedback values needed to get a particular gain and so forth. I miss this stuff. Old school electronics for analog dinosaurs like me.

Above — Second preamplifier schematic. The 22 µF input capacitor gets driven by emitter follower Q4 from Preamp 1. Preamp 2 voltage gain = 17.7 

The Baxandall tone circuitry time constants reflect that T-style guitars generally sound bright.  For the classic 100 Hz / 10 KHz Bass + Treble 3 dB turnover tone section, you might wish to run 100 nF and 15 nF for the capacitors respectively. The 50K bass potentiometer works well since I tend to 'pump the bass' & this prevents the impedance from getting too low at the extreme wiper setting seen when when boosting hard.
The treble and bass are fairly independent and the boost / cut is just over 10 dB. Clearly op-amp tone controls boost and cut with more amplitude, but this work OK and proved very simple. The emitter of Q6 provided a convenient node for negative feedback into the tone circuit.

 

The two 100 µF coupling capacitors help boost the low end for bright T-style guitars.

 

Above — A DSO trace of the Q6-Q7 feedback amp probed at the 22K load resistor. I  measured 26 Vpp output clean signal voltage — at 26.1 Vpp, the lower half started to clip. This image shows a virtue of split DC supply for making amplifiers: better headroom.  Not nearly as good as an op-amp, but pretty good headroom all the same.

Feedback Amp Notes

Above — This is my favourite AF feedback amp in single DC supply.  In Suzu version 4, I employed this particular feedback amp for Preamp 2 with a split DC supply. Simplicity, wide bandwidth, stability —  and medium to higher voltage gain make this a favourite amp for me. It goes well after a follower since the input impedance is relatively high and won't load down a source or emitter follower.  I use a VCC from 3 to 28 volts DC in my single supply design work and whatever I can muster from my power supply in my split DC supplies. Of course, you have to watch the transistor collector to emitter breakdown voltage. I stock (hoard) high voltage BJTs knowing they are getting scarce and more expensive.

In late 2021, retired EE Ken Kuhn suggested that I learn to make every discrete amplifier in split DC supply. (Paraphrasing) Ken wrote ... "any circuit can be biased to operate on single or split supplies and split supplies do not have to be symmetrical (i.e. +5, -12).  All that matters is the total supply voltage."

To that end, I learned to make all the common configurations such as common emitter, emitter/source followers and differential amplifiers with both BJTs and JFETs at various total supply voltages. I struggled with some feedback amps as the calculations seemed tricky and I had no example circuits to inform my own designs. I sent Ken the above 19 volt single DC supply feedback amp requesting help to convert it to split DC supply.

To my delight, Ken made a spreadsheet that did all the calculations and allowed the user to change supply voltages with the ability to adjust the gain to a desired value (combination of RE1 and RF).
Big thanks Ken!  You may change parameters like VBE -- it might be best to measure VBE and input that value, however, if not, the spreadsheet gets you close and offers a great learning tool.

Spreadsheet taken down for re-location to another server. 

Above — A screen capture from the spreadsheet manipulated to fit this image file. This shows an example of using the tool to run the calculations for my single DC supply amp shown earlier. Note that the feedback resistor idealized value = 510 Ω, not 560.  I adjusted RF using standard resistor values so that the 2 values VC2 center and VC2 actual were as close together as possible -- in this case 0.16 volts.

 

Above —My actual single DC supply amp with RF = 560 Ω. The difference between VC2 center and VC2 actual is only 0.6 volts, so well within the +/- 2V specified by Ken's spreadsheet. Notice the unloaded voltage gain rose by .91 . In reality my measured voltage gain was 11.7 -- the spreadsheet gets you close. You can manipulate RF and RE1 within reason to target more or less gain. The spreadsheet has a split DC supply example design defaulted into it. Between that example and my single supply examples here, the spreadsheet should prove easy to use if you ever build this feedback amp.

Within Suzu, RB1 can be made from parallel and/or series values, although my collection of resistors over 100K seems quite limited. To provide the Baxandall tone circuit with a higher input Z, I increased RB2 to 10K and made RB1 from two parallel 120K 5% resistors placed in series with a 150K 1% metal film resistor. I measured 208K from this resistor block -- it worked perfectly.

You may also stick a temporary pot for RB1 [ I used a 250K potentiometer] to find the exact center for the Q1 bias on the test bench. With a 1 KHz signal generator and DSO probe on the 22K resistor, I drove the amp just into soft clipping and tweaked the pot to find the sweet spot for a perfect bias voltage. I removed the pot and measured just over 208K.  Do not leave a regular potentiometer or trimmer pot in the actual circuit as it may add noise and potential for oscillations. 

 

The feedback amp also provides a soft start and silent power off for the guitar amp.

Output Filter

Preamp 2 contains a crude RC low-pass filter on the output. Some of my 10 inch speakers sound shrill -- and this switchable low-pass filter tames that down. Further, the added stopper resistor(s) changes the dynamics of the power amp. I like the 2nd or middle position switch a lot,  as it seems to make the guitar sound more “woody”.

I did make some active low-pass filter using FETs and BJTs and found they did not better,my tone. In the end, I preferred the RC filters since the added stopper resistors, plus the shunt caps provide me 2 additional practice tones to enjoy.

3. Power Supply

 Above — A basic power supply. The different green and orange LED resistors try to equalize their relative brightness on the front panel.  1 LED for each DC rail.

Above — For the first time ever, I'm using a commercial grade bridge rectifier and will also apply this part in my high powered amps. You may heat sink the GBUE2560 for high power amplifiers.

Above — Rectifier and 2 gorgeous reservoir caps for the DC power supply.

Above — The power supply transformer just sitting in the chassis prior to wire shortening and mounting.The Hammond 166L25 gives 12 watts out, while the166L20 gives about 8 watts clean output power. Further, if you regulate the op-amp DC supply with the 166L20, this means running +/-12 volts split as the unregulated DC voltage sags downs to less than 14 VDC on each rail when driven hard.

I also tested a larger transformer with 29 VDC unregulated on each rail & for awhile, Suzu was running at 27 Watts output power. The Hammond 166L25 and 166L20 have identical dimensions. In the end, I opted with the 166L25, since its higher output DC voltage allows running the preamplifier rails at 17-18 volts DC unregulated to get maximum headroom.

 Above — The power supply section mounted and tested. 

Above —  My downstairs Telecaster ™ with a Seymour Duncan Phat Cat single coil pickup in the neck slot and his Alnico 2 Pro™ in the bridge position. I added my newly designed, switchable treble bleed circuit in February 2023.  

 4. Power Amplifier    — P A —

Above — PA schematic. I chose different transistors for the input emitter coupled pair and also for the finals compared to the original GAA -12 Practice Guitar Amp. Further, I sank a little more current in the emitter coupled pair and the VAS/driver stack. At this point, I only plan to run voltage feedback in the global feedback loop, although, I can easily add current feedback if desired.

I measured a β of 540 for BC546B matched pair. The whole BC54-X- series seem an incredible BJT collection offering  low noise figure plus high β and, of course, is long obsolete. I've got 30 pieces of the über low NF BC549 in my parts bins for future 12 volt single-supply, discrete, low-noise AF amplifiers.

Above — Notice from Mouser. The day after I installed the power Darlington complimentary finals, I got this notice by email. Obsolescence might be the central story of my electronic hobbyist career ? Happily, I've got enough genuine power follower pairs -- both standard and Darlington style to last me for a long time.

 

Above —The finals mounted in their heat sinks. Once again a hack saw helped fashion DYI heat sinks.

 

Above — The finals and PA mounted in the "cake pan". The power transformer sat unmounted in this photo. Suzu with it smaller chassis and will go upstairs in our living room to serve as my main practice amp. The downstairs GAA -12 amp serves as my main transcription amplifier. I spend time downstairs  transcribing horn solos. I rarely listen to guitarists other than if a guitar happened to be on the song of the horn player whom I'm transcribing.

Above — Suzu's PA offers low distortion. I'm very happy with this PA stage. The matched input pair have obliterated the 2nd harmonic and I believe what's left are crossover + some intermodulation products from interactions with my outboard circuit, test leads, clips and probes. 

 5. Speaker

I chose the Eminence Legend 1058 speaker for my upstairs practice amp.

Fortunately, many kind YouTube posters have uploaded head-to-head trials with various 10 inch guitar speakers for comparison. I tend to favour Alnico magnet 10 inch speakers, however, dislike their cost. My "non Alnico" preference seemed to the the Legend 1058 in several videos. So I bought one and found it well suited my purposes. — and the added bonus,  it's not expensive.

Above — The large dust cap makes the speaker look bigger than 10 inches in diameter. This speaker is a gem. Ferrite magnet and weighs 2 Kg.

Above — Transfer function of the Legend 1058 from Eminence. It directly connects to what I hear with actual playing tests. In a cube shaped cabinet with my preamp circuit, the bass is OK while lower middle response sounds a little scooped. There is 1 "sharp" peak at ~2700 Hz, but the treble response starts to fall down a cliff at around 5 KHz. Perhaps a good fit for a Fender Telecaster ™ through a 10 inch speaker?  I prefer scooped lower mids for rhythm, but stronger lower mids for lead playing. There is no 'ideal' speaker for me it seems. 

Above — My wife designed & built a prototype cabinet from a plank of 12 inch wide, 3/4 inch thick pine. The final specs are 12 inches depth x 12 inches height x 14 inches width  [ or 30.48 cm deep x 30.48 cm height X 35.56 cm width ]. I stuffed some fibreglass pink insulation in the cavity to dampen any reflecting waves. The back is partially open with a 2 inch gap across the top end. This keeps out cats (protects the speaker), keeps in the insulation and gives punchy bass tones with some room audio fill through he back of the speaker cabinet.

Above — I've got a Jensen Mod 10-35 in another identical cabinet at the moment. I like the strong mids for neck pickup solos better when compared to the 1058, however, it sound quite bright. It's best to listen to a speaker for many months before you write in in or off.

6. Miscellaneous Photos

 

Best regards!   Click here for my Guitar-Related Index

Above — 1 of the Preamp 2 designs I explored, but later discarded.

 

Another Look at the LM386

On a whim, I took another look at the LM386-4 one cold Sunday afternoon this November. My focus was to drive a loud speaker and not headphones. I won’t personally use the LM386 for a headphone amp as we enjoy so many better options. For example, an op-amp driving a pair of TO-92 followers, or perhaps placing 2 NE5532 op amps in parallel as the headphone PA stage.

Since the mid 1970’s the LM386 has enjoyed popularity amongst hobbyists for low to flea power AF power amplification. The NE612 mixer IC plus the LM386 have literally formed the basic building blocks of innumerable radio receivers amongst Hams and hobbyists for decades.

Although, imperfect like all other linear ICs, the LM386 design team delivered a simple, flexible, low power AF amp with reasonably low distortion.

This part is noisy though. The input noise density = ~ 50 nV/√(Hz)  — about  10X that of an NE5532 op amp. So. if you use this part in high gain mode [with a gain of 100 to 200] and drive it with a low-level audio signal, you’ll really hear the noise (hiss) in your speaker.

Others online have provided detailed analysis about each stage of the LM386, so I won’t bother. However, I will comment about why it might be noisy. Normally, in modern AF power amps. the differential input pair emitters get 50 - 100 ohms of degeneration to boost linearity at the expense of noise. Other than that, in the IC only current sources connect to the emitters (and usually active loads to the collectors -- i.e. no resistors), However, the LM386 input pair get multiple large value resistors connected to their emitters. This translates into lots of Johnson noise from thermal agitation within conductors, plus related high-level input current noise that all gets amplified by the NPN voltage amp and delivered to the output stage.

Above — My basic test setup superimposed on the internal schematic of the LM386. I RC low-pass filtered the 12.24 VDC and ran two 1 Watt resistors in parallel as my resistive load. The measured load resistance for my power calculations = 8.4 ohms. I AC coupled a 1 KHz, ultra low distortion signal generator with a gain control to the input and watched the output in a DSO containing a stalwart FFT with 12-bit sampling. 

Above — An early photo during my initial bread boarding for these experiments. The non-inverting input resistor was changed to 10K for my experiments.

Index of this blog post

1. Gain = 20 Mode
2. Bass Boost and More
3. Gain = 50 Mode
4. Gain = 200 Mode -- plus lifting and AC bypassing Pin 2

I'm avoiding math in 2022, as data shows that my blog readers don't care for it.

Above —For comparison and contrast, here is the FFT of a discrete guitar PA pushing 3.4 Watts in this photo. The 2nd harmonic lies ~ 64 dB down.

1. Gain = 20 Mode

Above — The data sheet suggested amplifier with Gain = 20, using minimal parts. 

Above — The FFT of the above schematic with the fundamental plus 4 harmonic tones showing. I'll show many traces with a "standardized" output voltage of 5 Vpp or 372 mW into my 8.4 Ω load. This allows comparisons of various circuits. Note the 2nd harmonic is only 46 dB down  [ -46 dBc ].

Above —FFT with only 1 change -- I bypassed pin 7 with a 10 µF capacitor. The 2nd harmonic decreased by ~ 9 dB. Always bypass Pin 7.

 

 

Above — Apart from the 2nd harmonic, this FFT shows very low distortion at low signal levels (20 mW)  The clever 1/2 positive DC supply biasing scheme within the LM386 isn't perfect and the input pair are not perfectly balanced due to both variable AC and DC factors and this effects 2nd harmonic suppression. We also see some crossover distortion in the output. Don't get me wrong — I feel amazed by this design. Like you, I kind of like the humble LM386; plus have used it a lot over time.

 

Above — This is the cleanest sine wave I can drive before clipping using my eyeball. This is 725 mW output power.

Above —  I kept advancing the gain knob until clipping appeared on the top half.

Above —  Switch to FFT mode, although you can still see the yellow coloured sine wave. At this point, the tones are almost level and 45 - 47 dB down. You may easily hear distortion at this level.

Above — I kept advancing the signal generator gain knob until the tones look really strong. You can see the sine wave is now both bottom and top clipped. These FFT screen shots and my notes serve as the basis of what follows. I learned a lot doing these experiments and hope you like them.

2. Bass Boost and More

 Above — The data sheet suggested schematic for Amplifier with Bass Boost.

Above — Voltage gain versus frequency graph showing the 6 dB peak at ~ 90 Hertz. In the LM386, both AC and DC feedback runs back from the output to the non-inverting input emitter. Pins 1 and 5 allow us nodes on either side of this 15K resistor so we may provide additional AC feedback in parallel with it. The bass boost example provides us insight into possible feedback strategies: a 10K resistor + 33 nF cap form a network to boost the bass and also low-pass filters the frequencies out to ~ 5 kilohertz.

If you listen to this amplifier with music through a speaker, it sounds muffled and somewhat lacks the important mid frequencies for both voice and music. The hiss is definitely attenuated though. Further, the voltage gain goes from 20 without feedback down to around 8 with the 10K + .033 µF network added. There are other potential side effects to with such heavy feedback which I'll show soon.

Above — FFT with fundamental + 8 tones of the bass boost circuit driven to output 5 Vpp. The feedback has suppressed the 2nd plus all other harmonics effectively. Feedback certainly holds promise in reducing harmonic distortion in the LM386.

Above — The bass boost circuit may produce weird distortion when driven hard in some circuits. The bottom half of the sine wave clips initially.

Above — The FFT of the above DSO tracing when pushed a little harder. This looks and sounds terrible.I do not recommend people use the bass boost circuit, or if you do, please check for instability.

Let's adjust the feedback network and perhaps find something that works better.

In a classic PA with a differential transconductance input pair, 1 BJT base serves as the input stage while the other base receives the negative voltage feedback. The transconductance pair subtracts that negative feedback from the input and passes the difference voltage onto the voltage amplifier stage that follows. This doesn't happen in the LM386 -- negative feedback goes to the non-inverting BJT emitter which is often also the input side of the input emitter coupled pair.

Negative feedback also affects an amps gain, bandwidth, frequency response, plus its input and output impedance (although the output impedance is just fractions of an ohm). When we add AC feedback between pins 1 and 5, our network is in parallel with the 15K resistor and may be affected by other amplifier parameters including the gain and input impedance.

From the data sheet, In low gain mode, we should strive to keep the amplifier's closed loop gain 10 or greater which happens with the 10K resistor in our R C network. It's quite easy to turn your LM386 into an oscillator with too much feedback. I've noticed that feedback networks that look OK in SPICE simulations may actually oscillate in real life bench work -- especially with higher gain and/or drive.

I took the bass boost circuit example, kept the 10K resistor, and tried different capacitor values. If you lower the 10K resistor, you'll have to watch for oscillations at input drive levels high enough to cause distortion. This also may also reduce the LM386 voltage gain considerably.

Above — Our base schematic to evaluate different values of C1 and view the resultant FFT and voltage gain.

Above — FFT at 5 Vpp where C1 = 0.01µF. Outstanding results! This turned out to be the best feedback capacitor of the few I tried. The LM386 voltage gain dropped to just under 11 with that particular capacitor value for C1.

Above — Driving it as little harder to give 623 mW output power. Still fairly clean compared to other tracings.

Above — FFT with the top just starting to clip. 3rd harmonic re-emerging. 735 mW output power.

Above — Pushed a little harder to 761 mW. Things are getting ugly. C1 still = 10 nF. Let's decrease the C1 value by a decade to 1 nF:

Above — FFT at 5 Vpp where C1 = 0.001µF. While not as impressive as when C1 = 10 nF, it's still quite good and the LM386 voltage gain is around 19.

Above — FFT at 7.14 Vpp or 759 mW output power, The second harmonic is somewhat better than the case where C1 = 0.01uF.

Above — FFT at 5 Vpp where C1 = 470 pF.  Another favourable reduction of harmonic distortion when the LM386 amp is running at reasonably high, unclipped power levels. I measured no loss in voltage gain with a 470 pF cap + 10K resistor.

I also tried a 220 pF cap - it worked somewhat, but the harmonic suppression started to fall off at this point. The overall best unclipped harmonic suppression occurred where C1 = 0.01 µF in my experiments, albeit with 45% voltage gain loss.
To decide on a C1 value, it's important to listen to it to with actual audio to ensure that any frequency peaks, or more importantly, the low pass effects caused by the network doesn't wreck the audio you listen to.

Above — Listening to monaural jazz from my CD player into the LM386 and then into my 8 inch lab speaker. Although C1 = 0.01 µF gives the best reduction in harmonic energy, it rolls off too much high frequency audio for my tastes. I usually start at 0.001 µF and work up in capacitance. To my tastes, a 0.0018 or 1.8 nF cap sounded best. We've entered subjective territory. Variables may include personal taste, your hearing + age, your AF signal source overall tone, speaker size -- and perhaps whether the speaker is mounted in a cabinet, etc.. You might consider choosing a feedback cap between 0.01 µF and 470 pF according to your needs and wants.

Above — Listening through my 6 inch lab speaker. My C1 preference = .0039 µF for this speaker.

3. Gain = 50 Mode

Above — Datasheet example: Amplifier with gain = 50

Above — FFT at 5 Vpp with a 1K2 and 10 µF cap between Pins 1 and 8. Although we see the 2nd harmonic at about 52 dB down, it's still OK for the LM386. Signal noise will appear louder compared to the "Bass Boost" feedback variants with the same output power.

Above — I pushed it hard to 7.39 Vpp or 813 mW output power. This is the nastiness you'll hear on loud signal peaks.

4. Gain = 200 Mode -- plus lifting and AC bypassing Pin 2

Last section. Here's the famous schematic used by millions to make simple DYI audio projects:

Connecting Pins 1 and 8 AC bypass a 1.35K emitter resistor in the input pair -- and unleash the hounds. We get a full menu of gain, noise, and potentially harsh sounding distortion.

Above —In this separate experiment to showcase the worst-case scenario, I've manipulated & then pushed this particular amp into raucous distortion. Note the strong 3rd and 5th harmonics relative to the 2 even harmonics. This is worst case fuzz box stuff. While this might sound bad with your ears, it's great fun to see it on a DSO.

Above — Back to the main experiments using the schematic shown above... The FFT at 5 Vpp or 372 mW output power. The 2nd harmonic lies at -46 dBc.The various tones do not go down much at lower input signal levels.

Above — Increasing the signal generator output to push the amp into clipping. The 3rd harmonic looks ready to break open.

Above — A slight increase from 7.01 to 7.09 Vpp. The 3rd harmonic is about 41 dB down.When in full gain mode, the LM386 tends to offers more odd harmonics

Above — Top and bottom sine wave clipping translates into wretched distortion.

My question -- will feedback similar to what we used in the Bass Boost variants lower the harmonic distortion?

Above — At our standard of 5 Vpp, the affects of feedback leap out at us. [ 10K plus 0.01 µF cap ] Feedback is our friend?  However, the gain dropped to around 100 and you will hear lots of high frequency roll off.

Above — Vpp = 5 with a feedback network consisting of a 10K + 0.001 µF cap. The response is lack luster compared to 10K + 0.01 µF capacitor. The drop in voltage gain was only 2%. The 2nd harmonic is maybe 1-2 dB better than without the feedback network?

Above — Vpp = 5 with a feedback network consisting of a 10K + 220 pF capacitor. Interesting FFT ! The 2nd harmonic is now -50 dBc where without the network it measured -46 dBc.. Some of the tones dropped around 4 or 5 dB too. No change in voltage gain by adding this network.

Above — There's an old trick left to try. Normally, most builders will ground Pin 2 like I did throughout this blog post. What if we AC couple Pin 2 to ground through capacitor C2?. This may help to better DC balance the input pair bases ( may reduce DC offset ) and perhaps even bypass some portion of the distortion to ground.

Does this work?

Above — The FFT tracing shown above is with C2 in place showing that C2 does decrease distortion in certain cases. I could not superimpose 2 FFTs, so I made a red line above each of the 4 harmonic tones.

The bottom of the red line is the exact peak of the each tone with Pin 2 shunted to ground. Above the red line is the measured improvement for that particular tone caused by C2. I installed a switch across the C2 capacitor to make comparisons. . C2 = 0.01 µF in this particular experiment.

In my experiments with a gain of >=150 and no feedback network, when the LM386 is pushed into harmonic distortion, C2 lowered the harmonic tones by 4 to 8 dB. I tried C2 values of 0.01 to 0.27 µF
and changing the value of C2 within that range seemed to make no significant difference. Replacing C2 with a resistor of any value did not work to lower distortion.

C2 seemed to have less of an effect when the LM386 gain was lower than 150. At Gain = 20 with no feedback, I observed a maximum 2-3 dB maximal improvement in any 1 tone. With feedback, the effect diminished a little further, however, results were inconsistent. C2 does not appear to lower the harmonic distortion when the audio signal is unclipped -- rather, it seems to reduce distortion due to clipping when it happens. 

I performed other experiments such as bringing the feedback to Pin 2 with Pin 2 connected to ground via a resistor or resistor + capacitor (like what you do with an op-amp or discrete AF amplifier). I also tried lowering the feedback 10K resistor value at various gain levels. Often enough, the result was that the LM386 would go into a writhing spasm when pushed into distortion. See below.

Above —Fancy feedback experiments often resulted in the above tracing. It seemed better to explore  simpler ways to lower distortion.

Conclusion

Wow, this was a lot of work, but proved fun. I encourage you to perform your own experiments with the LM386. While no panacea, and a little long in the tooth, the LM386 reflects a simpler, mostly analog time for many of us home builders.

I suggest you consider using the LM386 with lower gain and build up your audio signal voltage with a low noise preamp using an op-amp like the NE5532. 

Further, consider adding feedback [ 10K plus some value of C1 ] from Pins 5 to 1 and also AC coupling Pin 2 to ground. I did both of these tricks in my 2 photographed bench CD player listening tests shown in Section 2. 

とてもいい