Skip to main content
The model reads your text as a performer would, not as a parser. Punctuation, spacing, and capitalization all change the audio. This page covers what actually moves the output.

​
The two controls

Both are optional, both range from 0.0 to 1.0, and both are calibrated for BTTS_V3. Leave them unset to get Bland’s defaults, which suit most conversational copy.
​
number
How much intonation varies. Higher is livelier and more dynamic; lower is flatter and more even. Raise it for marketing reads and characterful agents; lower it for disclosures, confirmations, and anything where a performance would be distracting.
​
number
How much the same input varies between renders. Higher repeats itself closely; lower takes more interpretive latitude. Raise it when you cache audio or A/B test copy and need the only variable to be the text.
They pull against each other. High expressiveness with high stability gives you the same animated read every time; high expressiveness with low stability gives you a different animated read every time.
BTTS_V2 voices synthesize, but these controls are not tuned for them. Read the x-model response header to see which model a voice resolved to.

​
Pauses

Insert a pause with <|N|>, where N is 0.1 to 10.0 seconds:
A marker also splits the line into two segments, and that boundary costs a breath, so the audio grows by more than N. Budget roughly a second on top of what you ask for. Beyond that the relationship is one to one: raising a marker from 1.0 to 2.0 adds a second. Stay inside the range. A marker above 10.0 fails the whole request rather than being clamped. Reach for markers instead of runs of dashes or ellipses when you want time. Punctuation moves emphasis and delivery along with pacing; a marker only moves the clock.

​
Punctuation

Always end a sentence with terminal punctuation. An unterminated final clause can lose its last word, most visibly with numbers.

​
Numbers

Long digit strings are where synthesis most often goes wrong. Three rules cover it:
  • Keep a run to roughly six characters or fewer.
  • Break longer runs with a pause marker.
  • End the sentence with a period.
Spelling a number as words gives you exact control over how it is read. 1500 may come out as “fifteen hundred” or “one thousand five hundred”; one thousand five hundred cannot.

​
Spacing

The model reads text token by token, so spacing carries meaning:
Hyphenated and concatenated constructions reduce accuracy. Prefer from the same array over from-the-same-array. Chinese and Japanese are segmented automatically. Do not add spacing by hand.

​
What is normalized for you

These are handled already. Rewriting them yourself usually makes the result worse:
  • Emails: john@gmail.com reads as “john at gmail dot com”
  • URLs: www.google.com reads as “www dot google dot com”
  • Smart quotes are converted to straight quotes
  • Repeated punctuation is collapsed: Hello!!! reads as Hello!
  • Missing spaces after a period are inserted: End.Start reads as End. Start

​
Streaming text

On WSS /v2/tts/ws, send LLM tokens exactly as the model produces them, including their leading spaces and punctuation. Each speak.text is appended verbatim, and Bland picks the synthesis boundaries itself. Do not re-wrap tokens into sentences before sending. Buffering on the client to hand over “clean” sentences adds latency to every turn and takes context away from the boundary detection. It also costs you the early audio outright. Release is driven by text continuing to arrive, so a turn delivered as one large speak returns nothing until end_of_turn, even when that one message holds several complete sentences. The same words sent as deltas start playing about a second in.
Pause markers and punctuation work the same over the WebSocket, but a marker split across two speak messages still works because the text is concatenated before it is read. Splitting <|0.5|> across turns does not, because a turn is synthesized on its own.

​
Checking your copy

Speech Studio renders a line immediately with the same models the API uses, which makes it the fastest way to compare two phrasings or two control settings before putting either in code.
Docs for agents: llms.txt