Ah, that's a good observation. Looking at the datasheet for that TI chip, it is powered by + 3.3 V, but contains an internal charge pump that generates a - 3.3 V supply. So the output can swing across ground. That's what eliminates the need for blocking caps. And it specifies a DC offset error of +/- 1 millivolt. So they've designed around the need for blocking caps with internal circuitry and a guaranteed maximum output offset.

Now what happens if there is an input offset, such as a constant stream of input words other than zero? That's where the system designer has to take care of things. And there are offsets even when blocking caps are used. The caps only prevent compounding those offsets from one stage to the next.

And a certain amount of offset won't kill your speakers. It just has to be kept within reasonable bounds.

I've noticed when using the audio input of a PC for measurement, that there is a constant offset, which is just an artifact of the ADC. The audio recording software has a function for removing that offset. In fact I would consider an audio recording with a large residual offset to be a poor engineering practice.

Often, switchmode amplifiers with BTL output have a large normal mode offset, so they can be powered by a single supply. And single-supply audio circuits tend to be festooned with blocking caps. ;-)