Is less than 24 bits worth it?

My first CD player, a Mission unit, had only 14 bit D/As, (oversampled) but was vastly superior to contemporary Sony units which boasted 16 bits, so I agree with sean (for a change). That being said, and all things being equal (which is not always true) 24 bits has potential for sonic improvement if the rest of the circuitry is up to it.

Let's not forget jitter affect on resolution!
Both circuit noise and data dependent jitter can easily "disqualify" the significance of additional least-order bits in a 20+ bit system. The intrinsic jitter from the oscillators that control the coax or Toslink connection to the DAC can destroy any hope of resolving near 24 bits.

How does jitter have an effect? I don't know how CD players are commonly configured, but, if I were designing it the serial bits would be loaded into a buffer register in whatever irregular timing they might arrive with from the spinning disc. (Actually a set of buffer registers, for a cache of values). Then the 16 bit word would be output parallel to the D/A, and converted on a clock mark. Only the conversion clock mark needs to be timed accurately.

I've heard it said that the state of the art advances by a little more than one bit per decade. Certainly right now more than 20 bits are lost in the noise floor.

As the others have said, a well-executed 16-bit player sounds better than a typical mass-produced 20-bit player, and usually by a wide margin.

It's just like with power amps: not all 100 wpc amps are equal, and there are lot of 50 wpc amps that sound far better than many 100 wpc amps with most speakers.

When the analog signal is finally created, both amplitude and phase inaccuracy create deleterious effects. Using 24 bit resolution, the last bit translates to a signal amplitude of about a 17 millionth of peak amplitude. You can get an "equivalent" effect if the exact timing of that 24 bit sample is off by mere nanoseconds (depending on the exact slope at the sample moment), which is nothing more than jitter. Jitter that can affect the difference in timing from one sample to the next is created by intrinsic oscillator noise in Phase Locked Loops (the thing that tracks and creates the internal clock) or by the very bit stream itself (a phenomenon called data dependent jitter). CDPs can be more successful than transport/DAC combos in addressing these effects. Both have a really tough job getting true 24 bit resolution.