An SSML Prosody Setting Depends on the Synthesizer

A voice assistant reads a confirmation while someone is deciding whether to send a message. The product team wants the destination and action spoken slowly, without making the whole interaction drag. SSML offers a local request:

<speak version="1.1" xmlns="http://www.w3.org/2001/10/synthesis" xml:lang="en-US">
  Send the message to <prosody rate="85%" volume="soft">the support team</prosody>?
</speak>

The SSML 1.1 prosody specification gives rate, pitch, and volume distinct jobs. rate changes speaking speed relative to a voice’s default. pitch addresses the baseline pitch of the contained speech. volume can request a labelled level such as soft or a signed decibel change relative to the current level. These are controls on the enclosed text, so a prompt can focus an important phrase without imposing the same treatment on every sentence.

The markup expresses intent, while the synthesis processor produces the sound. A rate of 85% is relative to a default that depends on the voice, language and dialect. A soft label can cause a different kind of adjustment from a numerical volume change. The specification permits a processor to limit an unsupported value or substitute another value, and in some circumstances to ignore prosodic markup it considers redundant, erroneous or harmful to speech quality. An accepted SSML document therefore does not establish an exact duration, frequency or perceived loudness.

Interactions between controls matter too. Within one prosody element, duration takes precedence over rate, and contour takes precedence over pitch and range. If an application emits both settings, its author should know which one governs the requested rendering. Avoid stacking controls merely because the interface exposes them.

For a voice product, keep the critical wording clear before adjusting its sound. Audition the complete prompt with each supported synthesizer, voice and language. Listen for intelligibility, clipping, awkward emphasis and whether the user can distinguish the action from the destination. Keep the visible confirmation text available when audio cannot carry the distinction reliably. Treat a change of engine or voice as a reason to repeat that listening check. Judge the prompt by whether people understand the decision across supported voices and engines.