Audio amplifier power supply design - Part 3: Power-supply rail rejection

First, ensure there is a negligible ripple component in the noise output of the quiescent amplifier. This should be pretty simple, as the supply ripple will be minimal; any 50 Hz components are probably due to magnetic induction from the transformer, and must be removed first by attention to physical layout.

Second, the THD residual is examined under full drive; the ripple components here are obvious as they slide evilly along the distortion waveform (assuming that the scope is synchronized to the test signal). As another general rule, if an amplifier is made visually free of ripple-synchronous artefacts on the THD residual, then it will not suffer detectable distortion from the supply rails.

PSRR is usually best dealt with by RC filtering in a discrete-component power amplifier. This will, however, be ineffective against the sub-50 Hz VLF signals that result from short-term mains voltage variations being reflected in the HT rails. A design relying wholly on RC filtering might have low AC ripple figures, but would show irregular jumps and twitches of the THD residual, hence the use of constant-current sources in the input tail and VAS to establish operating conditions more firmly.

The standard op-amp definition of PSRR is the decibel loss between each supply rail and the effective differential signal at the inputs, giving a figure independent of closed-loop gain. However, here I use the decibel loss between rail and output, in the usual non-inverting configuration with a C/L gain of 26.4 dB. This is the gain of the amplifier circuit under consideration, and allows decibel figures to be directly related to test-gear readings.

Looking at Figure 9.7, we must assume that any connection to either HT rail is a possible entry point for ripple injection. The PSRR behavior for each rail is quite different, so the two rails are examined separately.

Positive Supply-Rail Rejection
The V+ rail injection points that must be eyed warily are the input-pair tail and the VAS collector load. There is little temptation to use a simple resistor tail for the input; the cost saving is negligible and the ripple performance inadequate, even with a decoupled mid-point. A practical value for such a tail resistor would be 22 k, which in SPICE simulation gives a low-frequency PSRR of -120 dB for an undegenerated differential pair with current-mirror.

Replacing this tail resistor with the usual current source improves this to -164 dB, assuming the source has a clean bias voltage. The improvement of 44 dB is directly attributable to the greater output impedance of a current source compared with a tail resistor; with the values shown this is 4.6 M, and 4.6M/22 k is 46 dB, which is a very reasonable agreement. Since the rail signal is unlikely to exceed +10 dBu, this would result in a maximum output ripple of -154 dBu.

The measured noise floor of a real amplifier (ripple excluded) was -94.2 dBu (EIN = -121.4 dBu), which is mostly Johnson noise from the emitter degeneration resistors and the global NFB network. The tail ripple contribution would be therefore 60 dB below the noise, where I think it is safe to neglect it.

However, the tail-source bias voltage in reality will not be perfect; it will be developed from V+, with ripple hopefully excluded. The classic method is a pair of silicon diodes; LED biasing provides excellent temperature compensation, but such accuracy in setting DC conditions is probably unnecessary. It may be desirable to bias the VAS collector current source from the same voltage, which rules out anything above a volt or two. A 10 V Zener might be appropriate for biasing the input pair tail source (given suitable precautions against Zener noise) but this would seriously curtail the positive VAS voltage swing.

The negative-feedback biasing system used in the design in Chapter 7 provides a better basic PSRR than diodes, at the cost of some beta dependence. It is not quite as good as an LED, but the lower voltage generated is more suitable for biasing a VAS source. These differences become academic if the bias chain mid-point is filtered with 47 µF to V+, as Table 9.1 shows; this is C11 in Figure 9.7.

Table 9.1: How decoupling improves hum rejection

As another example, the amplifier in Figure 9.7 with diode-biasing and no bias-chain filtering gives an output ripple of -74 dBu; with 47 µF filtering this improves to -92 dBu, and 220 µF drops the reading into limbo below the noise floor.

Figure 9.8 shows PSPICE simulation of Figure 9.7, with a 0 dB sine wave superimposed on V+ only. A large C_decouple (such as 100 µF) improves LF PSRR by about 20 dB, which should drop the residual ripple below the noise. However, there remains another frequency-insensitive mechanism at about -70 dB.

The study of PSRR greatly resembles the peeling of onions, because there is layer after layer, and often tears. There also remains an HF injection route, starting at about 100 kHz in Figure 9.9, which is quite unaffected by the bias-chain decoupling.

Rather than digging deeper into the precise mechanisms of the next layer, it is simplest to RC filter the V+ supply to the input pair only (it makes very little difference if the VAS source is decoupled or not) as a few volts lost here are of no consequence. Figure 9.9 shows the very beneficial effect of this at middle frequencies, where the ear is most sensitive to ripple components.