I used a signal nulling method for most of the distortion measurements of my amplifier designs. This was the subject of my MSc dissertation in 1978, and I have sometimes been asked for more details, so here is the full original version:
The earliest published example of the method I have found appeared in Wireless World in 1953 (E.R.Wigan, "Diagnosis of Distortion" June 1953 pp261-266), and another article (F.Jones, "Dynamic testing of audio amplifiers", Hi-Fi News and Record Review, November 1970 pp 1655, 1657.) mentions that the Acoustical Manufacturing Company Ltd (Quad) had been using this method for the previous 25 years.
For the latest MJR7 amplifier tests I used a very simple circuit, which only works for inverting amplifiers. This is really just a copy of a circuit by Peter Baxandall published in Wireless World, Nov.1977, p65.
Having extracted the distortion signal it is possible to listen to it without any masking from the original signal. These sort of tests were being carried out by Peter Walker at Quad thirty or more years ago, and the conclusion then was that for some of the amplifiers tested the distortion was totally inaudible when reproduced at its original level, and it had to be increased many times before becoming audible. Adding the original signal back the resulting masking effect should further reduce any possibility of audible effect. There are some listeners who are insistant that distortion far below the normally accepted threshold of hearing nevertheless is audible, particularly if it involves high order harmonics or transients. My own view is that if the lowest speaker voltage found to produce an audible output at a 1m listening distance is found to be 300uV at 3kHz, then it makes no difference whether this is the 3rd harmonic of 1kHz or the 7th harmonic of 429Hz, it will be equally inaudible at lower levels. Evidence for sub-threshold audibility is limited, I have only found two published examples, and one is under highly artificial circumstances ('binaural beats in the brain') and the other is at supersonic frequencies, and both appear not to have been found repeatable by other experimenters. An interesting piece of evidence in favour of the conventional masking theory anyone can try can be found at the DiffMaker website, where there are examples of choir music with a marching band added about 65dB lower in level. (You need to download and install DiffMaker to reproduce the sound files.) Reproduced alone the marching band is easily audible, yet heard along with the choir music it is extremely difficult, if not impossible, to detect its presence. Normal daily experience is also fairly convincing, if I am straining to hear someone on a bad telephone line it is never helpful for someone nearby to start talking loudly. Masking is certainly real, but sub-threshold hearing of distortion is the opposite of masking, and seems extremely improbable.
Footnote: The Electronics M.Sc course in the Physics department at the University of Wales included a 9 month course of lectures and written exams, followed by a 3 month project presented as a dissertation. I only had financial support for the first 9 months, so the practical work had to be done quickly, and was finished in less than six weeks. I had hoped to do more extensive investigations of amplifier distortion, together with development of a new idea for a feedforward output stage, but this had to be postponed, and was eventually made into a working prototype, described in my April 1998 Electronics World article.