This article is from the rec.audio.* FAQ, by with numerous contributions by Bob Neidorff others.
Audio data is stored on CD as 16-bit words. It is the job of
the digital to analogue converter (DAC) to convert these numbers
to a varying voltage. Many DAC chips do this by storing electric
charge in capacitors (like water in buckets) and selectively
emptying these buckets to the analogue ouput, thereby adding
their contents. Others sum the outputs of current or voltage
sources, but the operating principles are otherwise similar.
A multi-bit converter has sixteen buckets corresponding to the
sixteen bits of the input word, and sized 1, 2, 4, 8 ... 32768
charge units. Each word (ie sample) decoded from the disc is
passed directly to the DAC, and those buckets corresponding to
1's in the input word are emptied to the output.
To perform well the bucket sizes have to be accurate to within
+/- half a charge unit; for the larger buckets this represents
a tolerance tighter than 0.01%, which is difficult. Furthermore
the image spectrum from 24kHz to 64kHz must be filtered out,
requiring a complicated, expensive filter.
Alternatively, by using some digital signal processing, the
stream of 16-bit words at 44.1kHz can be transformed to a
stream of shorter words at a higher rate. The two data streams
represent the same signal in the audio band, but the new data
stream has a lot of extra noise in it resulting from the
word length reduction. This extra noise is made to appear
mostly above 20kHz through the use of noise-shaping, and the
oversampling ensures that the first image spectrum occurs at a
much higher frequency than in the multi-bit case.
This new data stream is now converted to an analogue voltage
by a DAC of short word length; subsequently, most of the noise
above 20kHz can be filtered out by a simple analogue filter
without affecting the audio signal.
Typical configurations use 1-bit words at 11.3MHz (256 times
over-sampled), and 4-bit words at 2.8MHz (64 times oversampled).
The former requires one bucket of arbitrary size (very simple);
it is the basis of the Philips Bitstream range of converters.
The latter requires four buckets of sizes 1, 2, 4 and 8 charge
units, but the tolerance on these is relaxed to about 5%.
MASH and other PWM systems are similar to Bitstream, but they
vary the pulse width at the ouput of the digital signal
processor. This can be likened to using a single bucket but with
the provision to part fill it. For example, MASH allows the bucket
to be filled to eleven different depths (this is where they get
3.5 bits from, as 2^(3.5) is approximately eleven).
Lastly it is important to note that these are all simply
different ways of performing the same function. It is easy to
make a lousy CD player based around any of these technologies;
it is rather more difficult to make an excellent one, regardless
of the DAC technology employed. Each of the conversion methods
has its advantages and disadvantages, and as ever it is the job
of the engineer to balance a multitude of parameters to design a
product that represents value for money to the consumer.
All sampling techniques (so also D/A techniques) require an
analog reconstruction filter following the converter. This
filter inherently adds phase shift, frequency response ripple
and high frequency roll-off, depending on the characteristic of
the reconstruction filter (which depends on the position of its
poles and zeros).
An oversampling data converter generates a higher output
sampling rate than a simpler converter, so you can use a more
simple reconstruction filter, which is cheaper and more stable
in time and temperature and produces less noise. Also, modern
oversampling systems include digital filters which compensate
the response of the analog filter in the passband, so you can
achieve systems with an overall performance of 20 Hz...18 kHz
+/-0.05 dB. Also deemphasis is mostly done in the digital
So the "sound" of a CD player is more than just the number of
bits. It's the quality of the converter, the filter requirements
imposed by that converter, the quality of the filter, and of
course, the quality of the following analog components. Power
supply quality and clock jitter also influence the sound.