Testing Of Asynchronous Sample Rate Converters

The theoretically ideal sample rate converter is a device that converts the data transfer rate without changing the content of the material.  

Listening tests:  

Many devices are judged by their particular sonic quality. A recording engineer may prefer, for example, a small amount of distortion to add some "color" to the sound. The same engineer, testing a number of processing devices separately, may choose his building blocks to achieve a certain characteristic sound. Let us assume that the desired characteristic is based on some small amount of second order distortion which imparts a characteristically "warm" sound. A problem may arise when processing the sound through more than one unit. The "desired distortion" may be compounded beyond the desired level (in our example, the second order distortion may increase each time the sound is processed to an unacceptable end result).  

Sample rate converters may serve to reduce excessive clock jitter. This improvement can take place for any sampling-rate ratio (including 1:1). Jittery incoming data introduces signal dependent noise and distortions (increasing with signal amplitude and frequency). Such jitter reduction yields noticeable sonic improvement, thus complicating the objectivity of listening tests. Some manufacturers' comparison tests inappropriately compare a high jitter input to a low jitter output. Further confusion is due to the fact that digital domain FFT tests do not adequately show the effects of input jitter.  

Listening tests should be based on comparing the audio of a directly applied signal against the converted version of the same material. The greater the difference, the less ideal the converter.  

The sample rate converter should receive a low jitter data source, and drive a reference grade D/A converter, a high quality power amplifier and top grade speakers, all matched to 0.1 dB. "Blind" listening comparisons (A/B/X tests) by recording professionals yield the best unbiased results.  

Measuring performance:  

The most commonly used measurements are based on a standard FFT, THD plus Noise testing (in the digital domain) and phase linearity. Interpreting measurements performed on asynchronous sample rate converters is less straightforward. The asynchronous sample rate converter can not control the input and output rates (these rates are forced by the driving source and required destination devices). The converter is required to reconstruct the data content while accommodating receiver and transmitter clock rates.  

The digital nature of the process introduces quantization effect into the conversion. The ratio does not change smoothly. It tracks the clock rate variations in a quantized fashion (small incremental jumps). Proper design requires that the quantized ratio changes fall below the ear sensitivity levels. Tracking the clock rates sets restrictions on the maximum tolerable ratio step size and the manner in which the ratio tracks the clocks rates. Let us focus on the two extreme cases for ratio tracking:  

a. Steady clocks: The converter is adjusting the ratio up and down by a small amount around the correct average ratio.  

b. Fast "varispeed": The adjustment accuracy is reduced with fast varispeed affecting the accuracy of ratio adjustment.  

Technological limitations require careful consideration for optimizing both ratio step size and the tracking mechanism. Psychoacoustic considerations (listening tests) and practical limitations of varispeed should form the basis for proper performance criteria. Measuring asynchronous sample rate converters may reveal some of these compromises.  

While the input and output of theoretical converters measure identically, real converters continuously track and adjust internal coefficients. Such ratio modulation appears on FFT measurements as a "widening of the main lobe" of a sinusoidal test tone. The amount of widening depends greatly on variables such as ratio step size, ratio tracking, FFT size and type of FFT window used.  

Common digital domain measurements (FFT based measurements) do not show the effects of low levels of incoming jitter. A common indirect approach is based on measuring the THD plus noise reduction at the output of a reference DAC (driven by full scale high frequency tone). While such a measurement does not quantify jitter, it yields (in principle) the desired end result. Real world limitations of reference DAC performance set limits to such measurements (DAC performance is typically lower at high amplitudes and frequencies). Model 3000 utilizes a high Q (steep resonance) LC circuit for de-jittering incoming data. Further jitter reduction may be achieved with a 1:1 sample rate conversion, using a low jitter crystal clock oscillator for the output data clock.  

Sample rate converter performance should be measured over the usable audio range. Poor performance at high frequencies cannot be dismissed as inaudible noise. A typical justification for high frequency performance degradation is based on the fact that real music contains less energy at this frequency range. This could allow, at most, a few dB reduction in THD+N. Unlike many unsampled analog circuits which tend to generate higher frequency distortions, sampling folds back distortion energy to lower frequencies, including the ear's most sensitive mid-range region.  

When processing long words a desirable performance specification should exceed the limitations of a 16 bit word by a significant margin. This will ensure that overall performance is limited almost completely by word length bottleneck. (Truncation of a long word to 16 bits yields theoretical results of approximately 98 dB THD+N). The combination of long word format and good performance specifications is even more desirable when additional processing may take place. Premature truncation amounts to loss of detail.