Small title bar4

  Home -- article index                                                                    

Updated 4-28-2015

About Total Harmonic Distortion Analyzers

THD analyzers are important tools in the design and construction of high performance audio equipment, as well as in the design and construction of oscillators and other test equipment. As Paul Klipsch used to say, "You can't make what you can't measure because you don't know when you've got it made."

How THD analyzers work
Total Harmonic Distortion is measured by using a rejection filter system to remove the fundamental frequency (the first harmonic) of a waveform, usually a sine-wave, and then measuring whatever is left over — second and higher harmonics, and noise.

Naturally, the measuring system has to have wider bandwidth than the fundamental frequency being measured. For very low distortion signals, say under 0.05% THD, the significant harmonic content will be in the 2nd thru 5th harmonics — above the 5th, noise tends to dominate. So a measuring bandwidth of 5 to 10 times the fundamental will be needed, with 10 times usually preferred.

Most analog analyzers, like the HP 334A and the Heath IM-5258 for example, basically take the input signal, buffer it from the outside world, adjust its voltage to a convenient level, split it into two paths, sometimes inverting one path to get both non-inverted and inverted signals, and then pass one of the two resulting signals through a narrow-bandwidth filter of some kind — sometimes a peak (band-pass), sometimes a notch (band-reject). The filter is essential in order to eliminate, as much as possible, the fundamental, while the remaining harmonics and noise are processed for measurement (there are other circuit topologies for measuring THD which I'll discuss below). Then the two signals are summed (if one was inverted) or they are differenced (if they both have the same polarity).

In many distortion analyzers, the fundamental and an inverted exact copy are precisely matched for phase (either precisely in-phase or precisely out) and for level, at which point the summing or differencing eliminates the fundamental and leaves everything else. Feedback is used to sharpen the notch, raising the filter's Q. Depending on the circuit used, the feedback can be negative or positive. Given sufficient feedback, the fundamental will be eliminated but the harmonics will all be passed through with little or no attenuation -- a requirement for accuracy. The notch filter is not infinitely sharp, so some noise signals are attenuated along with the fundamental, but this usually is not meaningful. And if the cancellation is not complete, some residual fundamental can be passed through the filter, setting a floor on the measured minimum signal and maximum resolution. Ultimately, the filter used serves as a notch (band-reject) filter once all the processing is done.

Filter selectivity or "Q" is very important, because most filters have fairly broad peaks or notches, and will ultimately attenuate the 2nd and higher harmonics along with the fundamental, giving inaccurate measurements. A passive Twin-T filter, for example, will attenuate the 2nd harmonic by about 9dB and the 3rd harmonic by about 5dB, with higher harmonics being attenuated proportionately less. This attenuation represents a severe loss of accuracy unless mitigated in some way.. In active systems, feedback around the filter system sharpens the notch, preventing the harmonic attenuation and yielding trustworthy measurements. General practice is to keep 2nd harmonic attenuation to less than 1dB, and preferably less than 0.5dB, thus having little effect on the measurements of all the harmonics.

The output of this rejection system is then passed to a high-gain, wide-band voltmeter for measurement. Various filters can be used following the rejection system in order to eliminate hum and/or other spurious signals, such as wide-band noise that obscures the actual harmonics of interest. The voltmeter needs sensitivity in the microvolt range when the input signals are below 3VRMS -- a 3V signal with 0.001% THD will have distortion and noise components with an RMS sum of 30 microvolts; obviously, low noise and high gain are essential to the meter.

As mentioned, in a cancellation system, tuning the filter to cancel the fundamental consists of adjusting the amplitude and phase of the filtered signal to exactly match those of the fundamental component of the wideband signal. This means controlling the overall gain of the filtered fundamental and adjusting the phase of the filtered signal. Gain tuning is often done by adjusting the value of an attenuator resistor or a feedback resistor. Phase tuning is usually done by adjusting a filter component, usually a resistor. Automatic tuning is almost essential for measuring very low levels of distortion -- the null is very sharp and hard to achieve and stay in manually.

The natural phase transfer function of filtering the band-limited fundamental makes it possible to use a phase detector to measure and amplify the phase difference between the unfiltered fundamental of the wide-band signal and the filtered fundamental, allowing automatic control of the tuning element. Gain differences between the two fundamental signals are somewhat easier to measure, amplify, and use to adjust the level of the filtered signal. Properly done, an automatic system easily tracks even small differences in amplitude and phase, allowing the system to acquire and hold a deep null of the fundamental. In my experience, auto-tuning can hold nulls of from -100 to -120dB. It is important to realize that the level of any residual fundamental can set the bottom limit of measurable distortion. This means that a fundamental null of -100dB will limit the measured distortion to 0.001%, regardless of how low the actual harmonics are or how low the noise is.

The system described above is how the HP 330 thru 334, all the Heathkits, the Cordell State-Variable, the Tektronix AA 501A, the Sound Technology 1700, and many other analyzers work. There is another way, which is simply to use a traditional notch filter like a Twin-T or a Bridged-T and use positive feedback to sharpen the notch and deepen the null. Of course automatic tuning still offers the needed stability for deep nulls. This is the system used in the ShibaSoku 725 series and the HP 339A, among others. These analyzers often use a Bridged-T filter (with four components) for ease of tuning, while the HP 4333A used a Twin-T, which has six components, which complicates the tuning and switching.

The design issues in THD analyzers are:
1) the fact that optimally tuning them means changes of microvolts in amplitude and changes of milli- or micro-hertz in frequency.
2) measuring signals in the microvolt range while maintaining wide bandwidth.
3) noise.
4) intrinsic analyzer self distortion.

Many forms of filters and amplifiers have been used, each with its own strengths and weaknesses -- here's a summary:

Wien Bridge Filter
The Wien Bridge peak filter (band-pass) is one of the simplest and easiest to tune, which is why it has been used in so many THD analyzers, such as the HP 330 through 334 series, in all of the Heath THD analyzers, several Leader brand analyzers, and others too numerous to mention.

The Wien Bridge uses two resistors and two capacitors -- one resistor and one capacitor in a series leg, and one resistor and one capacitor in a parallel leg, to make a peak filter with a very broad peak. Usually, the two resistors have equal values, as do the two capacitors, which is intrinsically convenient.

The heart of the system (and of most analog analyzers) is that the phase of the filtered fundamental is zero, compared to the input signal, only at the precise center frequency of the filter. The Wien Bridge has an insertion loss of exactly 2/3, or roughly -10dB (-9.54dB), so the wide-band signal also has to be attenuated by the same amount before being combined with the filtered signal.

HP 331-4 notch filter

Advantages    With phase fine-tuning via small variable resistors fitted as series segments of the total bridge resistances (R, above), and just one variable attenuator resistor needed to fine-tune amplitude (variable part of Ra, above), tuning the Wien Bridge is dead easy from a circuit point of view. The Wien Bridge is easy to auto-tune as well, because both phase and amplitude can be adjusted by voltage-dependent variable resistors, usually photoresistors (CdS cells, also called LDRs) driven by lamps or LEDS, or by varying the channel resistances of JFETs.

The broad peak of the Wien Bridge is significantly sharpened by overall negative feedback around the filter. This sharpening gives a deep null of the fundamental. The null depth essentially sets the measurement floor of the device (if noise is not a limiting factor), while preventing unwanted attenuation of low-order harmonics, especially the second and third, so that the measured values are correct without needing meter correction factors.

The notch depth of the Wien Bridge filter is limited only by the ability of the tuning circuitry to achieve and hold a null; this is limited by the resolution of the variable resistances used in tuning and by the sensitivity and response speed of the amplitude and phase detector systems used for auto tuning. Remember, we're talking microvolts and milli- or micro-hertz here.

Disadvantages    A limit on performance of the Wien Bridge is, in part, the 10dB insertion loss, which means the rejection amp as a whole has to make up the gain and then some in order to have enough feedback to achieve a narrow and deep notch. This requires an overall main-loop feedback of about 16-20dB for acceptable performance. This means that the overall rejection amp gain with the main loop open needs to be at least 20dB for the feedback; then the 10dB filter loss reduces the signal-to-noise ratio, regardless of whether its loss is made up or not. This situation obviously creates issues with bandwidth and rejection amplifier system THD, not to mention noise.

The challenge in the rejection amplifier system is balancing the moderate signal levels needed for very low THD of the amp stages, especially the input stage, against the S/N ratio losses from levels that are too low -- that's why the insertion loss of the Wien Bridge matters. Note that Wien Bridges can be made from unequal value components -- if one leg has 2C and 1R, then the other has 2R and 1C, for example, then the insertion loss drops from 10dB to 6dB, a nice improvement that is offset by the shallower peak and the slower rate of phase change through the fundamental frequency. I've never seen a THD analyzer built in this way.

Capacitance tuning    Most Wien Bridge analyzers are capacitance-tuned. Since large value air-dielectric variable capacitors are also physically large, this means using smaller caps, which in turn means using larger resistors, and at low frequencies, the resistors get very large -- tens of megohms -- which has serious impacts on noise performance. Of course, applying noise filtering to low frequency measurements helps to offset the higher noise from high-Z circuitry, but the problem remains and is substantial.

Resistance tuning    The solution to the high-resistance, high-Z problem of capacitance tuning is to do range switching with fixed capacitors, which can easily be as large as needed, and do in-range tuning with either potentiometers or with switched resistors. This advantage is offset for potentiometer tuning systems by the difficulties and expense in making high-precision ganged potentiometers with good tracking. So, designs using switched 1% or better resistors are usually preferred.

The larger issue for me with Wien Bridge systems is that, in my experience, they have quite high residual distortion compared to other bridges, which significantly limits their performance. In practice, they have 10 to 30dB higher residual THD+noise floors.

Bridged-T Filter
One form of the Bridged-T notch (band-reject) filter works in an active circuit that is similar to the way that the Wien Bridge is used, with feedback sharpening the notch -- but in this case, the feedback is positive.

HP 339 notch filter

Like the Wien Bridge, the Bridged-T uses two resistors and two capacitors, but they are used to make a notch filter rather than a peak filter,with the advantage that the slope of the phase change through the fundamental frequency is quite steep, compared to the Wien Bridge, making auto-tuning sharper and more sensitive. As in the Wien bridge, range switching can be either by the resistors or the capacitors, with the other then being used for in-range tuning. Unlike the Wien Bridge, two of the components, either the resistors or the capacitors, have to be unequal in value, often in a 10:1 ratio for practical reasons of switching and impedance.

1-9-2011 -- In looking at this page again, I noticed that I had forgotten to mention the role of U2 in the circuit above. U2 provides positive feedback, whose level is set by the ratio of Rb and 8*Rb, the divider resistors in the input leg of U2. The ratio shown sharpens the notch enough to only attenuate the 2nd harmonic by about 0.3dB, an essentially negligible amount. If the value of Rb is too small, the circuit becomes unstable; too large, and the attenuation of the low-order harmonics becomes too great and THD readings are inaccurate.

This arrangement is useful for capacitance-tuned analyzers but, like the Wien Bridge, does result in large resistor values at low frequencies for the typical relatively small capacitances of ganged air-variable tuning caps.

Resistance tuning    In resistor-tuned form, the Bridged-T is easy to tune, with auto-tuning via LDRs, which I'm sure was one of the reasons that HP used it in the 339A analyzer (it also is the filter type for the 339's oscillator section) and ShibaSoku chose it for the 725. Note that the bridge capacitor is the smaller element and the pillar capacitor is the larger one.

Capacitance tuning    Like the Wien Bridge, if the in-range tuning is done with equal air-variable capacitors, the range resistor values are very large at low frequencies, and noise is again a problem. This is why commercial analyzers like the HP 339A use unequal capacitors for range switching and equal switched resistors (with a vernier) for in-range tuning. Note that, opposite to the resistance-tuned case, in capacitor tuning, the filter's bridge resistance is the larger element and the pillar resistance is the smaller one -- same basic calculations, but the location is reversed.

Disadvantages    As in the Wien Bridge, the major disadvantage is insertion loss. I simulated the Bridged-T filter circuit of the HP 339A in LTspice. I found that the circuit can easily achieve null depths of over 100dB, if the tuning resistors have a few milliohms resolution. This level of performance is achieved with capacitance range switching and a capacitor ratio of 10:1 for the two bridge caps. The wide-band signal is attenuated by essentially the same amount as the bridge's attenuation of the fundamental, which amounts to about 6 times, or -15.6dB, with the exact value determined by component tolerances. This means a fairly high insertion loss -- more than the Wien Bridge -- which means more noise -- but the sharper phase tuning is a nice trade-off.

A deeper notch (before overall feedback is applied) can be attained with a larger cap ratio, but that also means, unfortunately, more attenuation of the wide-band signal, resulting in a poorer signal-to-noise ratio. HP definitely optimized the overall design of the 339A, using a 10:1 cap ratio in the rejection filter section and a 100:1 ratio in the oscillator section, and it's performance is very good, if not great, by today's standards -- I now have one.

But the real disadvantage to both the Bridged-T and Wien Bridge filter systems, in my experience, is that they both end up with a relatively high 2nd harmonic signal component in the output of the notch filter system. In the case of the HP 339A, this 2nd H. spike is what sets the measurement floor of 0.001%, and it is what limits the resolution of the HP 331-4 series to about 0.01%. I believe that these distortions come from the non-inverting topology of the amplifier, and are the artifacts of common-mode failure in the amp. An all-inverting amplifier system would be highly preferred.

Twin-T Filter
The Twin-T is a notch filter that is ideal for a THD analyzer, since it essentially has zero insertion loss, plus it has the big advantage over the Bridged-T of an intrinsically very deep null, needing active amplification only to sharpen the notch. Here, a FET-input opamp U1 buffers the filter and U2 provides the active positive feedback. There is no need for gain adjustment as seen in the Bridged-T thanks to zero insertion loss. Notch bandwidth, or "Q," is controlled by the ratio of the Rb resistors -- too sharp and it's hard or impossible to tune; too wide and harmonic attenuation is excessive.

1-9-2011 -- 10*Rb is probably too much feedback; a fixed ratio of about 8:1 for the two resistors works fine -- no need to fuss with a pot.

Twin-T notch filter

Advantages    As shown in the example above, the Twin-T is easily adapted to an active filter design, with feedback to sharpen the notch and thus make correction factors unnecessary. Two sets of switched resistors (large pots don't work well for high resolution due to "graininess" and high temperature sensitivity) and a small precision pot in each leg (with scaled values) can take the place of the Rs shown above in order to yield high resolution; and with great care and even greater patience, such an analyzer is capable of extremely high levels of performance, well into the range of a few parts per million or even much better. See my Active Twin-T notch filter page for details.

Disadvantages    But the Twin-T has the huge disadvantage of being made up of three capacitors and three resistors, and they also have unequal values, which makes range switching complicated and makes tuning extremely slow and tedious. Auto-tuning is doubly difficult, and as far as I know, only HP has made a commercial analyzer, the 4333A, that uses a Twin-T filter, and which has a somewhat complicated auto-tuning system. I've never seen or used one of these and don't know how good it is, but I suspect it has high performance.

Despite the disadvantages, detecting the relative phases of the two branches of the fundamental in the bridge is quite straightforward, since the signals at the centerpoints of the two Ts have signals that are 90° apart at precise tuning. Phase detectors can easily sense this exact difference and adjust an LDR in the resistive leg with a lamp or LED, and can adjust an opamp configured as a variable negative impedance (that is, a capacitor) in the capacitance leg, again with LDR, to achieve precise tuning.

State-variable Filter
The filter type of choice?     For an analog analyzer, the state-variable (phase shift) filter may have no equal, since it has no insertion loss, is easy to tune, can be relatively low-Z for best noise performance and lends itself to precise auto-tuning. HP used it in the 8903 series, as did Bob Cordell in his terrific THD Analyzer, and it has been used in many others as well, including Tektronix and Sound Technology units. It can be capacitance or resistance tuned, with the usual caveats about large resistors for capacitor tuning, so switched-resistor tuning is generally preferred. 1kHz THD floors of 0.0003% (3ppm) or better are achievable, making these hard to beat.

State variable notch filter

Some comments are in order about this circuit. The two variable Rs and the two Cs together tune the filter at integrators U1 and U2. The output from U1 is the band-pass signal, and the output from U2 is a low-pass signal, if one is needed for anything. U3 is a unity-gain inverter to get the overall phase correct. Unexpectedly, changing either of U3's 10k gain-setting resistors will alter the tuning frequency, which provides a clue to auto-tuning the phase differential between the broadband signal's fundamental and the band-pass filtered fundamental. U4 is a unity-gain inverter to drive U5 with the correct polarity wide-band signal.

Resistors R/2 and R*2.5 set the notch sharpening gain, in this case about 14dB, yielding a 2nd harmonic attenuation of a bit more than 0.3dB. U5 is a differential amplifier that algebraically sums the wide-band input signal from U4 and the band-pass signal from U1 and the filter. The two signals to U5 are in phase, meaning that U5 differences them, and so the fundamental is subtracted, leaving only the distortion and noise to be sent to the meter. U5's gain-setting resistors control the amplitude of the differential between the two inputs to U5, again indicating one place where auto-tuning for amplitude can be applied. In essence, U5 reduces the gain in nearly precisely the proportion that U1's gain setting resistors raise it, providing a very deep null. Obviously, changing the ratio between R/2 and R*2.5 or between U5's two resistors, 10k and 2.5k, or between U4's two gain-setting 10k resistors, will change the signal balance at U5, thus changing the null depth.

It only occurred to me after drawing the circuit above that it would be better to put U4 on the output of U1 and invert the band-pass signal, leaving the wide-band signal with the cleanest signal path possible. The band-pass filtering helps to keep the band-pass signal clean, so nonlinearity in U4 will tend to have less effect on the overall result, because we only really care about the fundamental coming from the band-pass filter. Note that the only non-inverting stage is U5, the diff amp, which helps to keep common-mode errors low, but not insignificant -- if distortion residuals will be a problem, here is where they will happen.

Computer software-based analyzers
Digital signal processing is very attractive for distortion analysis, thanks to very good software performance. Most computers have some sound input capability, and very good external or internal "sound cards" are available, so often only a software application (several good ones are free) is needed to do analysis. Software analyzers work by converting the analog input signal to a digital signal using an analog-to-digital converter (ADC). Once in digital form, all of the processing is done in floating-point math at relatively high resolution, so the limits on performance are only the linearity, bandwidth, and signal-to-noise ratio of the input amplifier, and the linearity, bandwidth, and resolution of the ADC. Generally, software systems for distortion analysis work as spectrum analyzers with an amplitude vs. frequency display.

Advantages    at least 16-bit resolution is common in nearly all PC sound cards and on-board systems, In general, these chipsets yield 1kHz THD values in the 0.005-0.01% range. Very good considering the low cost, once the computer is paid for. Some sound cards and on-board chipsets offer higher performance, with 24-bit resolution. These systems often yield 1kHz THD floors in the 0.001-0.002% range, which is quite good, and some offer results in the 2-3 parts-per-million area. Higher performance yet is found in some dedicated digital analyzer systems, such as those from Audio Precision, with measurements better than those from most (or perhaps all) analog analyzers.

Disadvantages    The glaring deficiency of nearly all PC-based solutions is bandwidth. The Nyquist Limit constrains digital systems from measuring signals higher than 1/2 the sampling frequency of the ADC. Many PCs have 16-bit, 44.1kHz (or 48kHz) ADCs, meaning that the highest signal component measurable is around 20-22kHz, although 192kHz, 24-bit capabilities are becoming common. This in turn means a maximum fundamental frequency of from 2kHz to 5kHz, depending on the harmonic content of the signal. The best onboard systems typically have 96kHz or 192kHz sampling, with a bandwidth of around 45kHz to 90kHz, allowing examination of fundamentals up to 8 or even 20kHz for lower order harmonics.

Most users will want more measurement bandwidth. Plug-in sound cards and outboard sound systems can offer 24-bit resolution with 96kHz sampling being common, and with 192kHz sampling increasingly available, meaning that fundamentals of up to 10 or 20kHz can be measured, again depending on harmonic content. However, systems with very good linearity are still quite expensive, often in the same price area as used high-performance analog analyzers that have similar THD floors and better bandwidth.

Another disadvantage is that in general, the display of harmonics in such analyzers is a spectrum display, meaning the harmonics are shown as spikes, and the user has to add their individual levels up as the root of their mean squares to arrive at a THD value, which is tedious. Some software systems do the math of computing the RMS sum of the components, which is very handy. Also, software solutions generally do not provide selectable low-pass noise filtering, a feature which is useful in arriving at trustworthy THD values and for seeing low-level distortion components.

Finally, most PC sound inputs have a limited level range, meaning some external level adjustment is needed to handle a wide range of signal amplitudes. This is a relatively minor issue, but needs to be considered, since maximizing signal-to-noise ratio is essential for best performance.

1-9-2011 -- In fairness, a good way to extend the amplitude sensitivity of a computer based analyzer is to pre-filter the input signal with an active Twin-T filter like the one described above. Reducing the level of the fundamental by 40 or 60dB will relieve the spectrum analyzer's input amp and ADC from needing huge dynamic range and extremely low intrinsic THD, leaving only its signal-to-noise ratio as a measurement concern. Providing that you can accurately measure the amplitude of the fundamental, and that you know the precise 0dB reference level of the spectrum analyzer, you can easily determine the dB level of harmonics referred to the level of the fundamental. If the analyzer is reading the 3rd harmonic at -105dB, the 0dB reference level of the analyzer is 1VRMS, and the oscillator is putting out a 5VRMS signal, then the 3rd harmonic is actually at -119dB relative to the fundamental.

Using my PC's "Intel High Definition Audio" chipset sound input and the ARTA spectrum analyzer software, I was able to see a THD residual of my state-variable oscillator, IG-18 #2, as being around 0.0007 to 0.0012%, depending on a variety of factors relating to level and measurement bandwidth. By using the active Twin-T filter between the oscillator and the sound input, and by nulling the fundamental to -80dB, I was able to see that the only significant harmonic was the 3rd, at 0.0005%, with everything else below -120dB or in the "grass" of the analyzer display. With the active filter, the only limitations to the PC system are the maximum frequency of fundamental and the frequencies of the harmonics that are accurately measurable with the overall measurement bandwidth.

3-10-2011 -- But a concern remains, and that is overall linearity. I recently measured the linearity of the system using a precision 40dB attenuator (1ppm accuracy) driven from the very accurate attenuator in my HP 652A Test Oscillator. My PC system has very good linearity down to -100dBu, then begins to compress rapidly below that, so that a signal at -110dB displays as -105dB. Compression in that direction means that the results are worst-case, which is comforting. Now, -100dB is nothing to sneeze at, but every such system needs to be checked for its linearity deviation, especially the 16-bit ADCs. This loss of linearity has effects on the evaluation of analyzer null depth among other things. See below.

5-5-2011 -- Using the very high spectral resolution of the ARTA software, I was able to separate out the oscillator signal from adjacent noise spikes and determine that the linearity is actually much better than first noted above. I found that significant compression didn't begin until below -110dBV, and amounted to about +2dB at -120dB. This is much better than I expected.

6-19-2012 -- Real resolution has come to my bench via the Active Twin-T notch filter described on my webpage about it, and coupling that with an E-MU 0204 USB 24-bit/192kHz sound module. This combination has given me a measurement bandwidth of about 90kHz and superb linearity approaching -140dBV relative to a 0dBV (1VRMS) input level. Meaningful measurements of 10kHz signals are practical, and useful measurements of even 20kHz signals can be made if they are very low distortion (i.e. have only low-order harmonics) to begin with. Highly recommended.

1-18-2011 --
Notes on some specific THD analyzers

HP 330 series (and various old Heath analyzers with two-digit model numbers)
These still show up on auction sites. They are vacuum tube units of the Wien Bridge type that are physically large and which have limited resolution, with a maximum full scale sensitivity of 0.3% (or 1% for some Heath and other tube units, and even on some older solid-state units from Leader, Meguro, Marconi, and others). Tuning range is typically 20Hz to 20kHz. Regardless of age, they can be useful and can be improved. Whether that's really practical depends on your frame of mind. Once, in leaner times, I replaced the tubes in an HP 330D with cascode amplifiers made up of JFETs and high-voltage NPN transistors, and was able to get residual THD measurements in the 0.01% area. That unit served well for years, but was a bear to tune since these analyzers didn't have auto tuning. Size and weight are the main barriers to using them as a springboard for greening, but they are treasure-troves of parts.

HP 331A thru 334A series (and Heath IM-5258)
These are solid-state, discrete-transistor designs; no opamps. They are Wien Bridge units with very wide range; in the HP units, up to 600kHz fundamentals(!!) and wide measuring bandwidth, above 3MHz except on the most-sensitive range. They offer low noise, and, in the 333 and 334 (and Heath unit), auto tuning. They employ average-responding metering. They have either a 400Hz passive LC high pass filter that is very good, or a passive 30kHz low pass LC filter, also very good, but not both filters. They desperately need expanded low pass filtering for better performance in the audio mid-band. Max full scale sensitivity is 0.1%, fully usable to 0.01% and below, except that their residual THD floor is around 0.01% (I've been told that the Heath has a higher floor but I've not used one). In the HP units I've used, this floor is dominated by 2nd harmonic distortion, which comes from the bridge amplifier/notch feedback system. I've never checked the depth of the fundamental null. I really should but I have good reason to believe that it is better than -90dB and may approach -100dB. Except for their high residual THD floor, and somewhat insensitive metering (300uV/0.1% full scale), the low noise and nulling of these analyzers may make them useful for measurements to 0.003% or even below. I think they may be candidates for greening.

HP 4333A
This is an analyzer only--no oscillator is built-in. It was the last one from HP to use discrete transistor circuitry, although it does have a sprinkling of opamps in locations non-critical for bandwidth, noise, or residual THD. As I've mentioned elsewhere, I've never seen or used one of these analyzers, but I do have a PDF copy of the manual. Effectively, the 4333 bridges between the 331-4 and the 339. It has some auto-tuning features and uses a Twin-T filter with fairly elaborate and clever tuning circuits. It specs out closely to the 339, but does not have built-in low-pass filters. It's voltmeter has a sensitivity of 100uVRMS full-scale, with less than 10uV residual noise over a bandwidth of more than 200kHz, so it is pretty quiet. The resulting maximum full-scale THD is 0.01%, so based on noise alone, the THD floor would be under 0.001% -- the spec is less than 0.0018%. The metering is average-responding, calibrated in the RMS of a sine wave. Probably a reliable performer, but certainly less desirable than a 339.

Sound Technology 1700 series
These solid-state units are often for sale on the auction sites and I judge them a good buy at the prices they currently sell for--better in every way than the HP 331-334 and not much more expensive. They have Wien Bridge oscillators and phase-shift/state-variable notch filters, with balanced input and output, and offer max full scale sensitivity of 0.01%, easily readable to below 0.001%. They were the standard for high-performance measurements for quite a while. Some models included intermodulation measurement capability. The metering is average-responding. I used one for several years, and the residual measurement floor, as I remember it, was in the 0.0012% area. The built-in oscillators have very low THD, using a combination of JFET and LDR AGC, and cover the frequency range from roughly 10Hz to 100kHz (or was it 20Hz to 200kHz?). The analyzers offer both 400Hz high-pass and 80kHz low-pass filtering to improve sensitivity and resolution. Despite the inconvenience of their relatively large size, they are still fine units as-is, and may offer potential for even higher performance, depending on actual nulling capability and noise levels. Sound Tech made extensive use of Harris 2600 series wide-band, low-noise opamps in the design of the 1700.

HP 339A
This instrument, like the ST 1700, has both oscillator and analyzer sections, and has been a favorite of broadcast engineers and techs, audiophiles, and audio service techs for years. It offers true RMS metering with a VU mode if wanted. It has a very good Bridged-T oscillator with JFET AGC. Mine has 1kHz THD well under 0.0002% and it has a Bridged-T analyzer section that is comparable in performance to that in the Sound Technology units. Like ST, HP made extensive use of the Harris 2625 opamp, which has absolutely terrific performance. It's a shame that they are no longer in production. Few current opamps have the 2625's combination of virtues and HP put them to good use. The metering system is true RMS, a great advantage for accuracy.

The 339A I have at the moment has a null depth around -105dB of greater than -110dB, based on the PC system's linearity test, setting an absolute THD floor of less than 0.0003%. Mine has a residual THD of a bit less than 0.001% with 30kHz low pass filter and 400Hz high pass filters engaged. If a 10kHz low pass filter were available, the 1kHz residual might drop another dB or two, but the real measurement floor is set to around 0.001%, as noted above, by the strong second harmonic spike in the notch amplifier.

It isn't clear to me that the null floor or residuals can be any lower in these units, but if they can, then they may be capable of 0.0003-5% floors. While they are not fully automatic, like the 8903 series, they are easy to use, thanks in part to auto set level, and to simultaneous tuning of the oscillator and analyzer. On the negative side, you rarely see HP instruments from this era with knobs that aren't broken or damaged—not HP's finest hour, mechanically. Because of their construction, I find them hard to work on, but it would be fun to get better performance out of one.

My experience is that the low-pass filter amps are quite noisy, and replacing the quad opamp -- I think it's an LM324 -- that is central to them would be a good idea. I have discovered that the 2nd H. of the oscillator can be minimized by "tuning" one of the two 2.00k resistors at the base of the oscillator's JFET AGC, thereby improving residual THD; and a friend has suggested that replacing the quad op-amp in the AGC circuit with a quieter and faster unit can significantly reduce the THD above 10kHz. This may also require replacing or adding some capacitors.

HP 8903A/B/E
The 8903A has oscillator and analyzer sections with unbalanced output and input. The 8903B has state-variable oscillator and analyzer systems, and has balanced out and in, very useful for measuring bridge-type amplifiers or balanced gear. The 8903E is an analyzer only, with balanced in. These units offer true RMS digital display metering and a very usable frequency counter display, and employ computerized operation with fully automatic level setting and tuning—just connect the signal, push the DISTN button, and wait a few seconds. Their most popular application seems to be in radio production testing and servicing. They offer SINAD measurements, and the fully automatic operation makes them great units for final test or the service bench. I have an 8903E on hand at the moment, and the null floor, taking measuring system non-linearity into account, is greater than 120dB at 1kHz; much better than 0.0003%.

However, from a high performance viewpoint, these units, which should be naturals, are hampered by a couple of things. One is the 400Hz high pass filter, if fitted. It and several other special filter types are optional. They are in the signal chain ahead of the notch filter, adding a pair of opamps to the chain and contributing significant additional THD and noise. This filter really should have been after the notch filter, as in the 339A, which is where the low pass filters of 80kHz and 30kHz are.

A second issue is that there are five opamps in the signal chain ahead of the notch filter, even when no special filters are used. There are seven or eight if a filter is used and this raises the residual THD floor significantly. These opamps, all NE5534s, do balanced to single-ended conversion, auto set level functions and programmable-gain functions for auto level set. The programmable-gain amps are not by-passable without interfering with the computer operation of the analyzer.

A third issue is that using the monitor output for post-processing is difficult, because the output level does not have a fixed reference level. It varies with the gain-setting done by the microprocessor and the various programmable gain amps.

As with the 339, I'm not sure if a deeper null is achievable with the 8903 or how best to reduce the residual THD of the amplifier chain. Better amps than the 5534 do exist, but it isn't clear that using, say, AD797s or LME49710s or OPA134s or LT1122s would actually help. I may have to find out. The 8903E on my bench measures the oscillator in my HP 339A at 0.0009%, with the 30kHz low pass filter in the circuit, and I think that's the best this analyzer can do at present. Having a 10kHz low pass filter available would certainly help. Substantially modding the 8903 would not be easy because of the microprocessor based operation, but it isn't altogether impossible to make constructive changes....

Tektronix AA 501A
I recently (April 2015) had an opportunity to use a Tektronix AA 501A Distortion Analyzer. This fully automatic unit is very easy to use and, except for lacking a frequency display, is fully comparable to the HP 8903A/B/E units in performance and operation.

It looks to me like the notch filter is a Bainter filter followed by a non-inverting summing amp for the notching. And the result is a strong 2nd H., just as with the HP units. The particular one I looked at had a THD residual at 1kHz of 0.0011%, with a 2nd H. level of -105dB, as seen from the Function Output. Its null was deep at -120dB. This analyzer depends utterly on the NE5534 opamp (in singles and duals) and so perhaps a better summing amp, one with less common mode error (an OPA1641?) would improve this 2nd H. issue.

ShibaSoku 725
I recently (October 2014) had one of these on the bench, and it is as good an analyzer as I think exists at this time. Like the HP 8903E, it is an analyzer only; no oscillator is included. Also like the HP 8903E, it offers automatic operation and a digital frequency readout. It uses an active Bridged-T notch filter like the one in the HP 339A, but after that, everything is different.

This instrument uses a 12-bit ADC to digitize and process the post-notch signals, providing a variety of options for analysis, and is capable of resolving THDs of less than 1ppm, 0.0001%! In addition to supplying TRMS metering, it offers several options for average-responding metering, and several LP and HP filter options. It also provides analysis of the levels of individual harmonics from 2nd through 5th, with resolutions into the ppb (yes, parts-per-billion) range.

I could not determine the self-distortion of the 725C that I saw — I have no oscillator with low enough distortion to see where the floor actually is. The noise levels of this unit are certainly low enough (well below -140dBu) to imagine an actual THD+N resolution of perhaps -125 or -130dBu.

What about Boonton, Krone-Hite, and other analyzers?
Great question. The Boonton and Krone-Hite units that I am aware of are basically much like the HP 8903 series, offering microprocessor-based automatic operation, at least 20Hz to 100kHz operating range, RMS metering, digital frequency display, and 1kHz THD resolutions in the 0.001% area. The similarities of specs and performance make me think that all of these units were built to meet some telecom industry or governmental requirements, and this probably includes the ShibaSoku as well, even though it far outperforms all of the others.

back to top

to Home -- article index

© 2010-2014 Dick Moore