Input format

The TTS API accepts raw UTF-8 text bytes. The backend normalizes pushed text itself, one sentence at a time, so numbers, dates, currency amounts, acronyms, and abbreviations can be pushed as-is. This page describes what that normalization does, the character set the model ends up speaking, and the SSML markup the model understands.

Pushing text

Pass text as UTF-8-encoded bytes to push():

tts.push("Dr. Smith owed $42 to the WHO.".encode("utf-8"))
# spoken as "doctor Smith owed forty two dollars to the W H O."

What normalization does

The backend runs these steps, in order, on the spoken text:

Step	Example
Spell out acronyms	WHO → W H O, HTTP2 → H T T P two
Normalize dates	12/25/2024 → December twenty fifth, two thousand twenty four
Numbers to words	$42.50 → forty two dollars and fifty cents, 3.14 → three point one four
Read a leading minus as a sign	-5 → negative five, -$3.50 → negative three dollars and fifty cents
Unicode normalization	é → e, — → ,, … → , , ,
Collapse repeated punctuation	!!!! → !
Collapse whitespace	multiple spaces → single space
Remove special characters	drops anything outside the model alphabet
Expand abbreviations	Dr. → doctor, St. → street, etc. → et cetera

A trailing period is added when the utterance has no sentence-ending punctuation, so the synthesis pipeline has a terminal to flush on.

The minus reads as a sign only directly before a digit, at the start of the text or after whitespace; ranges (5-10), compounds (COVID-19), and list bullets (- 5) keep the hyphen.

An all-caps token that the pronunciation dictionary knows as a word is spoken as that word rather than spelled out, so NASA is voiced. Tokens shorter than four characters are always spelled, so US and IT are not mistaken for the everyday words they spell.

The result uses the restricted character set the model’s phoneme dictionary operates on:

Accepted	Examples
Letters (upper and lowercase)	a–z, A–Z
Space
Apostrophe	'
Comma, period	, .
Question mark, exclamation	? !
Hyphen	-

SSML markup

The model supports a subset of SSML (Speech Synthesis Markup Language) for controlling prosody and emotion. Well-formed tags pass through normalization untouched: only the spoken text between tags is normalized. Malformed tags (an unquoted attribute value, a missing closing quote) are dropped with a warning through the SDK log callback, so one bad tag cannot abort the utterance. Tags and attributes outside the table below are passed through and ignored by the model.

Supported tags

Tag	Description
<speak>	Optional top-level wrapper. A closing </speak> forces a pipeline flush.
<prosody rate="..." pitch="..." volume="...">	Controls speech rate, pitch, and volume. Accepts named presets (x-slow, slow, medium, fast, x-fast for rate; x-low through x-high for pitch; silent through x-loud for volume) or percentage values. Tags are nestable.
<abr:emotion type="...">	Applies a composite prosody preset. Nestable with <prosody>.

Supported <abr:emotion> types: apologetic, calm, empathetic, firm, lively.

When SSML tags are present, no trailing period is added between or after the spoken runs. Use explicit sentence-ending punctuation, a </speak> tag, or the flush byte to end a tagged utterance.

Next steps

• Examples: full examples showing how to synthesize text.

• Overview: overview of the TTS streaming model and output format.