SSML text to speech

It reads the spoken words out of the markup and synthesizes them into natural-sounding audio, so an existing SSML workflow can point at it without first stripping tags back to plain text. Be clear on the boundary up front: this engine renders the words, not the fine-grained acoustic controls SSML can describe.

What SSML is and why it exists

Speech Synthesis Markup Language is a W3C standard built so a writer can describe how a voice engine should say something, not just what to say: where to pause, which syllable to stress, how to pronounce an acronym, whether a string of digits is a date, a phone number or a plain quantity. It has been the industry convention for controlling synthesized speech for years, precisely because plain text is genuinely ambiguous — '2/3' can be a date, a fraction or a score, and only markup resolves that in the abstract. That is the format this endpoint accepts on input; what it does with each tag is described plainly in the next section.

How the request and response work

You submit SSML-tagged text (or plain text) in the request body. The service extracts the spoken words from the markup and synthesizes them; the tags themselves are stripped to their text before synthesis. The task is queued asynchronously and returns a task_id immediately; the rendered audio is delivered afterward via a signed webhook or a signed link valid for 24 hours. Because the words are what gets synthesized, submitting SSML and submitting the equivalent plain text produce the same natural narration — the markup is accepted for compatibility, not because each tag changes the sound.

What this engine does and does not render

This is the honest boundary, stated so nothing on this page oversells it. The current voice engine reads the text content and produces natural, well-paced speech. It does not acoustically apply the fine-grained SSML controls: emphasis, phoneme pronunciation, prosody and exact break durations are accepted in the input and then stripped, not rendered. So a two-second break tag does not hold the line for two seconds, and an emphasis tag does not stress that word harder than the model would on its own. If your use depends on those controls being rendered precisely, that capability is not available on this engine yet — this endpoint exists so an SSML pipeline can produce audio today without a separate plain-text branch.

Fitting it into automated content

SSML is just structured text, so it is easy to generate programmatically — a content pipeline can wrap numbers in tags, insert breaks between sections, or mark up a name with a phonetic hint, all without human intervention. The value of this endpoint for that audience is that it accepts the SSML they already emit as-is: they do not have to build and maintain a second code path that strips their markup down to plain text before sending it. Pricing reflects the light, targeted nature of the task — a small per-request base plus a per-1000-character rate on the submitted text, with failed tasks never billed after automatic retries.

Access and how results are handled

This endpoint requires prepaid balance like every other one in the API; a request without balance returns HTTP 402 rather than a partial or degraded result. Submitted text and generated audio are removed after the retention period and are never used to train models — delivery is strictly through your webhook or a temporary signed link, and nothing is kept beyond that window.