Amplifier design project: the APS-2 speaker driver

This lesson brings the course together in one design: the APS-2's 4 W loudspeaker amplifier, built as a discrete three-stage power amplifier with global negative feedback. It uses nearly every idea from the earlier lessons: a degenerated differential pair with a current-mirror load at the input, a voltage-amplifier stage with Miller compensation, a class AB complementary output stage with a thermally tracking bias, a feedback network that sets the gain and suppresses distortion, a compensation design checked for phase margin, a heat sink sized for the worst case, and protection for both the amplifier and the speaker. The lesson follows the design the way Cedarline runs a project: specification, architecture, stage-by-stage design, verification by calculation and simulation, and bench testing, with a design review at the end.

The specification

Amara writes the specification with Daniel before choosing any part, because every later decision is checked against it. The numbers come from the system requirements: the loudness needed over traffic noise, the converter and attenuator levels from the microcontroller, the speaker's rating and the housing's environment. Accessible pedestrian signals in the United States follow the MUTCD and the road agency's own specification, which call for an audible locator tone near the push button and automatic volume adjustment to ambient noise; the APS-2 specification turns those requirements into electrical numbers.

The architecture

The architecture is the classic three-stage discrete power amplifier, the same plan as the op amp of lesson 4, but with a power output stage. The input stage is a differential pair, Q1 and Q2, with a 2 mA tail current from a current source and a current-mirror load; 100 Ω emitter resistors degenerate it to widen its linear range and to raise its slew rate for a given bandwidth (lesson 4). The voltage-amplifier stage (VAS) is a common-emitter transistor running at 10 mA from a current-source load, with the compensation capacitor C_c from its collector to its base. The output stage is a complementary class AB pair of power transistors with driver transistors, 0.22 Ω emitter resistors and a V_BE multiplier on the heat sink (lesson 10). Global feedback from the output through R_f = 10 kΩ to the inverting input, with R_g = 1 kΩ to ground through a 100 µF capacitor, sets the gain at 11 for audio but 1 at DC, so that the input stage's offset is not amplified.

Stage design

The input stage sets the amplifier's offset, noise, slew rate and unity-gain frequency. With 1 mA per side, each transistor has g_m = 38.7 mS; the 100 Ω emitter resistors reduce the stage's effective transconductance to g_m/(1 + g_m R_E) ≈ 7.95 mS and widen its linear input range from about ±50 mV to roughly ±200 mV. The current mirror converts the differential current into a single-ended output current with the full transconductance (lesson 3). The tail is a current source rather than a resistor, giving good rejection of the ±12 V rails' ripple, which appears as a common-mode disturbance.

The VAS converts that current into the large voltage swing needed to drive the output stage, about ±11 V, and the compensation capacitor across it creates the dominant pole by the Miller effect (lessons 4 and 8). Because the VAS must also supply the output stage's driver base currents and charge C_c, its current source is set at 10 mA, generous enough that the VAS never runs short of current at the highest frequencies and amplitudes the amplifier must handle.

The output stage uses complementary Darlington pairs, so the VAS sees the 8 Ω load multiplied by the product of two current gains, thousands of ohms, and the V_BE multiplier spans four base-emitter junctions. Amara sets the multiplier for a quiescent current of 30 mA, giving about 6.6 mV across each 0.22 Ω emitter resistor at rest, and mounts the multiplier transistor on the heat sink beside the output transistors. A current limiter on each side senses the emitter-resistor voltage and limits the output current to about 2.7 A, protecting the stage against a shorted speaker line.

Verifying the loop

The first verification is the loop gain and closed-loop response. The code models the forward path with the dominant pole set by C_c, the output stage's pole and a second, higher pole in the input stage, computes the phase margin and the loop gain at the audio frequencies of interest, and then the closed-loop gain across the specified band, including the input coupling capacitor and the feedback network's DC-blocking capacitor.

import cmath, math
fu, p2, p3 = 3.83e6, 2.0e6, 10e6          # unity-gain frequency of the forward path; non-dominant poles
A0 = 2e5                                  # low-frequency forward gain (sets the dominant pole)
fp1 = fu / A0
Rf, Rg, Cg = 10e3, 1e3, 100e-6            # feedback network with DC-blocking capacitor
Rin, Cin = 10e3, 220e-9                   # input coupling
def A(f):
    s = 1j * f
    return A0 / ((1 + s / fp1) * (1 + s / p2) * (1 + s / p3))
def beta(f):
    zg = Rg + 1 / (2j * math.pi * f * Cg)
    return zg / (zg + Rf)
def closed(f):
    hp = (2j * math.pi * f * Rin * Cin) / (1 + 2j * math.pi * f * Rin * Cin)
    return hp * A(f) / (1 + A(f) * beta(f))
lo, hi = 1e3, 1e8                         # crossover where |T| = 1
for _ in range(200):
    m = math.sqrt(lo * hi)
    if abs(A(m) * beta(m)) > 1: lo = m
    else: hi = m
pm = 180 + math.degrees(cmath.phase(A(m) * beta(m)))
print(f"crossover {m/1e3:.0f} kHz, phase margin {pm:.1f} deg")
for p in (1.0e6, 0.5e6):                  # sensitivity: what if the output-stage pole is lower?
    T = lambda f: A0 / ((1 + 1j*f/fp1) * (1 + 1j*f/p) * (1 + 1j*f/p3)) * beta(f)
    lo, hi = 1e3, 1e8
    for _ in range(200):
        x = math.sqrt(lo * hi)
        if abs(T(x)) > 1: lo = x
        else: hi = x
    print(f"  output pole {p/1e6:.1f} MHz: phase margin {180 + math.degrees(cmath.phase(T(x))):.1f} deg")
for f in (1e3, 10e3):
    print(f"loop gain at {f/1e3:4.0f} kHz: {abs(A(f)*beta(f)):6.1f} ({20*math.log10(abs(A(f)*beta(f))):.1f} dB)")
g1k = abs(closed(1e3))
for f in (100, 200, 1e3, 5e3, 10e3, 20e3):
    print(f"closed-loop gain at {f:7.0f} Hz: {abs(closed(f)):6.3f}  ({20*math.log10(abs(closed(f))/g1k):+.2f} dB re 1 kHz)")

crossover 343 kHz, phase margin 78.3 deg
  output pole 1.0 MHz: phase margin 69.8 deg
  output pole 0.5 MHz: phase margin 57.4 deg
loop gain at    1 kHz:  348.1 (50.8 dB)
loop gain at   10 kHz:   34.8 (30.8 dB)
closed-loop gain at     100 Hz:  8.911  (-1.81 dB re 1 kHz)
closed-loop gain at     200 Hz: 10.343  (-0.51 dB re 1 kHz)
closed-loop gain at    1000 Hz: 10.971  (+0.00 dB re 1 kHz)
closed-loop gain at    5000 Hz: 10.998  (+0.02 dB re 1 kHz)
closed-loop gain at   10000 Hz: 10.997  (+0.02 dB re 1 kHz)
closed-loop gain at   20000 Hz: 10.989  (+0.01 dB re 1 kHz)

The model gives a crossover near 350 kHz and a phase margin close to the hand estimate, and the margin stays above 60° with the output pole at 1 MHz, falling below it only if the pole drops to 0.5 MHz, which would need a far slower output transistor than the ones chosen. The loop gain is about 350, or 51 dB, at 1 kHz and about 35 at 10 kHz, so output-stage distortion is reduced by those factors. The closed-loop response is flat within a few hundredths of a decibel from 1 kHz to 20 kHz and only about 0.5 dB down at 200 Hz, inside the ±1 dB requirement. That small droop comes from the low corner set by the 220 nF input capacitor and the 10 kΩ input resistor, near 72 Hz, chosen to keep useless bass out of the small speaker while leaving the voice band intact.

Verifying distortion

The second verification is distortion. The output stage's transfer curve is computed from the exponential laws of its two transistors with their emitter resistors and bias voltage, for a class B bias (no quiescent current) and for the chosen 30 mA class AB bias. The code finds each stage's distortion on its own, driven to a small and a full output, and then inside the feedback loop using the loop gains just found at 1 kHz and 10 kHz.

import math
VT, IS, RE, RL = 0.02585, 1e-12, 0.22, 8.0
def vbias_for(IQ):
    return 2 * VT * math.log(IQ / IS) + 2 * IQ * RE if IQ > 0 else 0.0
def drive_for(vo, vb):
    """Base drive (relative to mid-bias) that produces output vo, for bias voltage vb."""
    iL = vo / RL
    lo, hi = max(iL, 0.0) + 1e-15, abs(iL) + 5.0     # npn current; pnp current = in - iL
    for _ in range(200):
        i_n = (lo + hi) / 2
        i_p = i_n - iL
        if i_p <= 0: lo = i_n; continue
        f = VT * math.log(i_n / IS) + VT * math.log(i_p / IS) + (i_n + i_p) * RE - vb
        if f > 0: hi = i_n
        else: lo = i_n
    return vo + VT * math.log(i_n / IS) + i_n * RE - vb / 2
grid = [-11 + 22 * k / 2200 for k in range(2201)]
def tables(IQ):
    vb = vbias_for(IQ)
    return [drive_for(v, vb) for v in grid]
def interp(xs, ys, x):
    lo, hi = 0, len(xs) - 1
    while hi - lo > 1:
        mid = (lo + hi) // 2
        if xs[mid] > x: hi = mid
        else: lo = mid
    t = (x - xs[lo]) / (xs[hi] - xs[lo])
    return ys[lo] + t * (ys[hi] - ys[lo])
N = 256
def thd(out):
    h = []
    for k in range(1, 10):
        re = sum(out[n] * math.cos(2 * math.pi * k * n / N) for n in range(N))
        im = sum(out[n] * math.sin(2 * math.pi * k * n / N) for n in range(N))
        h.append(math.hypot(re, im))
    return 100 * math.sqrt(sum(x * x for x in h[1:])) / h[0]
beta = 1 / 11
for name, IQ in (("class B ", 0.0), ("class AB", 0.030)):
    drv = tables(IQ)
    for Vpk in (0.5, 8.0):
        # open loop: drive the stage with a sine sized for the target output peak
        a = interp(grid, drv, Vpk)
        ol = [interp(drv, grid, a * math.sin(2 * math.pi * n / N)) for n in range(N)]
        res = [f"{name} {Vpk:3.1f} V peak: open loop {thd(ol):7.4f} %"]
        for f, T in (("1 kHz", 350.0), ("10 kHz", 35.0)):
            Af = T / beta                 # forward gain from error to output-stage drive
            xs = [d / Af + beta * v for d, v in zip(drv, grid)]   # input needed for each output
            amp = interp(grid, xs, Vpk)
            cl = [interp(xs, grid, amp * math.sin(2 * math.pi * n / N)) for n in range(N)]
            res.append(f"{f} {thd(cl):7.4f} %")
        print(", ".join(res))

class B  0.5 V peak: open loop 24.5089 %, 1 kHz  0.1204 %, 10 kHz  1.1598 %
class B  8.0 V peak: open loop  2.8604 %, 1 kHz  0.0085 %, 10 kHz  0.0830 %
class AB 0.5 V peak: open loop  0.0681 %, 1 kHz  0.0002 %, 10 kHz  0.0020 %
class AB 8.0 V peak: open loop  0.2404 %, 1 kHz  0.0007 %, 10 kHz  0.0069 %

The results show why both class AB bias and feedback are needed. Open loop, the class B stage produces gross distortion at 0.5 V peak, because a dead zone of more than a volt swallows most of the small signal; even feedback with a loop gain of 350 leaves measurable crossover distortion at low level, and at 10 kHz, with a loop gain of 35, it is far worse. The class AB stage starts with far lower open-loop distortion, below 0.1 percent at low level and a few tenths of a percent at full output, and feedback then brings it to about a thousandth of a percent or less at 1 kHz and below 0.01 percent at 10 kHz, at both the quiet and the rated output. The specification's 0.5 percent at 1 kHz and 1 percent at 10 kHz are met with wide margin. The model ignores the drivers, the VAS's own nonlinearity and the speaker's reactive load, so the bench result will be higher, but the ranking and the order of magnitude are what the design review needs.

Offset, thermal design and protection

The DC offset at the output comes from the input stage alone, because the 100 µF capacitor makes the DC gain 1. The input pair's bias current is about 1 mA/(β + 1) ≈ 10 µA for β = 100; if it flowed through unequal resistances at the two bases it would create a large offset, so Amara makes the input resistor equal to R_f, 10 kΩ, leaving only the offset current, perhaps 1 µA, times 10 kΩ, about 10 mV, plus the pair's own V_OS of a few millivolts. The total, below 20 mV worst case, meets the 50 mV requirement without trimming.

Protection completes the design. The current limiters hold the output current below about 2.7 A into a short. A Zobel network of 10 Ω and 100 nF across the output keeps the load resistive at high frequencies, where the speaker and its cable would otherwise look inductive and capacitive and erode the phase margin, and a small air-core inductor with a parallel damping resistor isolates the loop from long speaker cables. A DC-detection circuit compares the low-pass-filtered output with ±1 V and opens a relay in series with the speaker if a fault drives the output toward a rail, so a failed transistor cannot burn the voice coil. At switch-on, the relay closes only after the rails and the feedback capacitor have settled, avoiding a loud thump from the speaker in a quiet residential street at night.

Bench testing and the design review

Leo builds three prototypes and tests them against every line of the specification, at room temperature and in the environmental chamber at −40 °C and +74 °C. The test plan includes the DC operating points of every stage against Amara's calculations; the quiescent current at 25 °C and with the heat sink at 110 °C; the gain and frequency response from 100 Hz to 50 kHz; THD at 0.5 W and 4 W at 1 kHz and 10 kHz with an audio analyzer; a small-signal square-wave test into the real speaker and 15 m of cable, with and without the Zobel network, to read overshoot as phase margin (lesson 8); slew rate with a large square wave; a short-circuit test at full drive; the DC fault test with a deliberately shorted output transistor on a spare board; and a continuous worst-case tone in the 74 °C chamber, with the heat sink and transistor case temperatures logged until they stabilize.

The results mostly agree with the calculations. The quiescent current is 32 mA at room temperature and 37 mA with the heat sink at 110 °C, confirming that the thermal tracking works, though not perfectly. THD at 4 W and 1 kHz is 0.015 percent, higher than the simulation's thousandth of a percent because of the drivers and VAS, but thirty times below the requirement. The square-wave test into the speaker and cable shows about 6 percent overshoot without the Zobel network and less than 2 percent with it. In the 74 °C chamber the heat sink stabilizes at 108 °C with the worst-case tone. One result surprises everyone: at −40 °C the amplifier's quiescent current falls to 18 mA and the crossover distortion at 0.5 W rises to 0.08 percent, still within specification but four times the room-temperature value. Amara traces it to the thermal tracking working in reverse and adjusts the multiplier ratio slightly, accepting a few more milliamps at the hot end.

At the design review, Daniel walks the team through each specification line, the calculation that predicted it, the simulation that checked it and the measurement that confirmed it, and records every discrepancy with its explanation. He also asks for the class D comparison from lesson 11 to be kept in the design file, so the next housing revision can revisit the choice. When the review closes, Daniel signs the design package as the professional engineer responsible, and the APS-2 moves to compliance testing and a field trial at two Madison intersections. For Amara, who will sit the PE exam once she has her required experience, the review record is also evidence of exactly the kind of engineering work the license requires.