INTRODUCTION
Measurement of distortion is fundamental for design and evaluation of audio circuits. Several techniques have been defined for distortion measurement and have been widely used for improvement of audio circuits. However, the evaluation of top quality circuits via listening test does not tally with the figures given by they poor techniques, and more people in spite ofthey poor distortion figures.
There have been some attempts to define new, sharper measurements better correlated with subjective tests, but with little success. An explanation of this failure may be that these new measurements are based on the classical theoretical model of distortion, regardless of possible misconception concerning distortion in audio circuits. Questioning the theorical basesof audio circuit distortion is fruitfull and leads to breaking new ground in audio circuit measurement.

SUMMARISED THEORETICAL ANALYSIS : TRADITIONAL THEORETICAL ANALYSIS
The classical theoretical model of an audio power amplifier is the base for measuring amplifier distortion. It is made up of a perfect amplifier and two distortion generators : the linear distortion generator corresponds to the amplitude, phase, phase-slope and group delay modifications resulting from the band limitations of a real amplifier ; the non-linear distorsion generator corresponds to the non-linear transfer characteristique of a real amplifier.
The aim of current distortion measurement is to characterise the distortion generators. Band limitation and non-linear transfer function are measured in order to fully characterize the circuit under test and to define its distortion for any audio signal. The characterisation of the distortion generators is made with sinusoidal signal.
This approach is rigorous and valid as long as the model itself is valid. The validity of the distortion model is widely accepted even though this model does not take account a known distortion phenomenon : Transient Intermodulation Distortion. The reason for this is probably that TID (as far as this concept is limited to slew-rate limitation) only affects poorly designed circuits and can easily be avoided. However, slew-rate limitation shows that linear and non linear distortions can be combined in a more complex way than in the classical amplifier model.
Unfortunately, other phenomena combining linear and non-linear distortion occur in many audio amplifiers. Thus their non-linearity is not adequately analysed with sine waves and thus by classical distortion measurements. It is possible to exaggerate these phenomena and to design two simple circuits exhibiting exactly the same classical distortion measurement (band limitations, non-linear distortion figures and spectrum) but showing different distortions with many non sinusoidal and audio signals. These circuits also have a very different sound quality. They prove that the classical measurements of a circuit are usually unable to define its sound quality.
This example highlights a basic limitation of classical measurements in that static measurements are only reliable for stable systems. Classical measurements reply on the implicit hypothesis that the distortion characteristics are immuable. If not, classical measurements fail to fully characterise circuit distortion and to define circuit behaviour with any signal.

NEW THEORETICAL ANALISYS
A thorough theoretical analysis of audio circuits reveals many possible causes making characteristics unstable, and especially, variable according to the signal. There are many sources of memory in audio circuits :

• Memory occurs in components.
• Memory also occurs in circuits and mainly results from combinations of non-linear transfer functions and band limitations.
• Gobal memory is the combination of all these memory effects.

A new circuit model including memory can be proposed for distortion analysis. The linear distortion is produces not only by the band limitation effects, but also by the memorizing of the signal. The non-linear distortion is produced by a non-linear variable transfert function.
The new distortion model is more complex than the previous one and its characteristics are not easy to measure. Memory phenomena are ignored by measurements using static signals like steady-state sine waves (or the signals used for the attemps of new measurements).

RESULTS USING A NEW MEASUREMENT SET
Measurements of memory were made with a new measurement set ;

• on a commercially available high quality transistor amplifier, with a THD of -86dBc at the level of the test signal.
• on a triode tube amplifier (SE 300B) designed by an audiophile, with a THD of -27dBc at the level of the test signal.
• on a new transistor amplifier designed for low memory, with a THD of about -110dBc at the level of test signal.

RESULTS AND CORRELATION WITH LISTENING TESTS
Several listening tests were made with the mesured AC amplifiers in different conditions, with different listeners ; they gave the same results. They seem to show that the measured memory is better correlated with sound quality than the THD.
The tube amplifier, in spite of its poor distortion figure, was judged as giving a much more natural sound than the traditional transistor amplifier, completely in opposition to the traditional distortion measurement values.
The memory-free transistor amplifier (Lavardin Technologies), thanks to its unusual sound quality, was preferred to the tube amplifier even by tube fanatics involved in the listenings tests. This results invalidate an explanation for the preference for tube circuits : the hypothesis of distortion pleasant for the ear.

CONCLUSION :
These limited first results of memory measurements in audio circuits prove that memory really occurs in audio circuits. They show that the proposed model for circuit distortion is closer to reality than the traditional model.
Even if the reason for the audibility of memory distortion is not yet clear, the quality improvement resulting from a memory-free design shows that memory is audible. Low memory distortion is the reason for the good sound of tubes.