SSML

The W3C specification defines elements for structure and for control. Structural elements such as speak, p and s establish phrasing. Control elements do the work: break inserts a pause of a given duration, say-as reinterprets a token as a date, time, telephone number, ordinal or spelled-out characters, sub substitutes an alias for the written form, phoneme specifies pronunciation directly in IPA, and prosody adjusts rate, pitch and volume over a span.

Support is uneven and that is the main practical constraint. Every engine implements a different subset, and neural voices in particular often ignore or only loosely honour prosody adjustments, because forcing pitch and rate onto a model that inferred its own contour tends to sound worse than leaving it alone. Markup written against one vendor rarely transfers cleanly to another.

For voice agents, the durable use of SSML is normalisation rather than performance. Wrapping an order reference in say-as with a character interpretation so it is spelled rather than read as a number, or a date so it is spoken in the local convention, prevents a class of error that no amount of prompt engineering fixes. Many platforms now expose a pronunciation dictionary as a simpler alternative for the common case of one word rendered wrongly.