Skip to main content

​
Introduction

Bland TTS lets you clone a voice from short audio samples and use that voice anywhere a Bland agent speaks, on calls, on the Speak endpoint, or over the streaming WebSocket. Two cloning engines are available:
  • V3 (default), single 10-second sample, 17+ languages, highest fidelity.
  • V2, exactly 1 WAV file, 17+ languages, production-stable previous generation.
To create or manage voices in the dashboard: https://app.bland.ai/dashboard/voices?tab=studio To clone programmatically, see Clone Voice.

​
Sample requirements

For any engine, the source sample should have:
  • Clear enunciation
  • No background noise
  • A single speaker
  • At least 100 ms of silence at the start and end (avoid cutting off mid-word)
  • Emotional variety, for more expressive output
Tip: Use any audio editing tool to trim and clean your sample before uploading.

​
Per-engine constraints

V2 and V3 voices are single-sample by design.

​
Voice configuration

Each cloned voice has two adjustable defaults. Set them via the Studio UI or Update Voice Config, or override per call via the Speak body parameters.

​
Consistency (V2 and V3)

consistency controls how many frames the decoder waits before emitting the first audio frame.
  • Lower values: faster first audio, but slightly reduced quality in the opening few hundred milliseconds.
  • Higher values: smoother output, slightly slower time-to-first-audio.
Range: integer 1-32 (V2/V3), float 0.0-1.0 (V1, where higher = more consistent).

​
Boost / Expressiveness (V2 and V3 use boost; V1 uses expressiveness)

Adds prompting to align the output more closely with the original voice sample.
Important Note: Disable boost if your agent needs to code-switch or speak multiple languages not present in your sample. If you see broken or distorted output, try toggling it off.
V1 uses a continuous expressiveness (0.0-1.0) instead of the binary boost switch.

​
Text formatting

​
Spacing

V2 and V3 process text token by token, so spacing matters. Correct:
Incorrect:
Hyphenated or concatenated words can reduce accuracy:
Note: For Chinese and Japanese, the tokenizer handles character segmentation automatically. No manual spacing required.

​
Pause markers

Insert <|N|> where N is a duration in seconds (0.1-10.0) to add a pause:

​
Text normalization

The model automatically normalizes common patterns. You do not need to pre-format these:
  • Emails: john@gmail.com becomes john at gmail dot com
  • URLs: www.google.com becomes www dot google dot com
  • Smart quotes: converted to straight quotes
  • Repeated punctuation: Hello!!! becomes Hello!
  • Missing spaces: End.Start becomes End. Start

​
Numbers

For best results with numbers:
  • Keep sequences to roughly 6 characters or fewer
  • Format numbers as words when possible
  • End sentences containing numbers with a period
  • For long sequences, break them up with pause markers
Important Note: If a sentence does not end with proper punctuation, the model may drop the last number.

​
Punctuation

Punctuation directly affects speech delivery:
  • Ellipses (…): add pauses and weight
  • CAPITALIZATION: increases emphasis
  • Standard punctuation (., ,, !): natural speech rhythm
  • Dashes: can add stutters or breaks

​
Best practices

  • Pick a sample with emotional variety for more expressive output.
  • Use natural speech patterns and proper punctuation.
  • Avoid unusual formatting (excessive dashes or ellipses) for timing control. Use pause markers (<|N|>) instead.
  • Keep word spacing clean.
  • End sentences with appropriate punctuation, especially when numbers are involved.
Docs for agents: llms.txt