Text to Speech CapabilitiesWebSockets

Advanced Settings

In real-time generation, input text often arrives in chunks. These advanced settings let you fine-tune text buffering to balance audio quality and Time to First Byte (TTFB).

Default Behavior

By default, our system waits until a sentence is complete—determined by punctuation—before sending it for voice generation.

min_buffer_size, max_buffer_delay_in_ms and predictive_chunking are top-level siblings of text, not members of voice_config. Nesting them inside voice_config fails silently. All three also accept camelCase aliases (minBufferSize, maxBufferDelayInMs, predictiveChunking).

min_buffer_size

When a sentence isn’t complete and no punctuation is detected, the system uses min_buffer_size to decide when to send the text. This parameter sets the minimum number of characters required before sending input for audio generation. A larger buffer provides better context for the model, leading to higher audio quality. Reducing this value can lower TTFB by enabling quicker responses.

  • Range: integer 0 to 1000 characters
  • Default: 100

Setting min_buffer_size: 0 means “do not wait on volume”: whatever is buffered is sent immediately, even a partial fragment. This gives the lowest TTFB at some cost to prosody, because each request is synthesized independently — a partial sentence is rendered as its own utterance, and rejoined fragments can sound abruptly paused.

max_buffer_delay_in_ms

If a sentence is incomplete and text hasn’t reached min_buffer_size as well, max_buffer_delay_in_ms sets the maximum time (in milliseconds) the system will wait before processing the input. Once this delay is reached, the available text is sent—even if it’s below the buffer threshold.

  • Range: integer 0 to 1000 milliseconds
  • Default: 300

max_buffer_delay_in_ms: 0 is the same “do not wait” contract on the time axis. Setting both knobs to 0 is the most aggressive configuration and synthesizes nearly every delta on arrival.

predictive_chunking

Controls how eagerly the segmenter decides a sentence has ended when the ending is ambiguous, such as 2 p.m., the U.S. or Dr..

  • Values: strict (default) or lenient
  • strict waits for one more character when the boundary is genuinely uncertain.
  • lenient takes its best guess and starts sooner.

Unambiguous sentence endings are never delayed in either mode, so strict does not add latency across the board — only on genuinely ambiguous endings.

A non-numeric value for either buffer knob is rejected with a warning (INVALID_BUFFER_SIZE or INVALID_BUFFER_DELAY) and the previous value is kept. Numeric strings such as "100" are still accepted. See Errors & Warnings.

Example

Let’s say the first chunk of text received is: “I just wanted to say that”

  • This sentence is incomplete and has no punctuation, so the system doesn’t immediately send it for audio generation.
  • The current length is 24 characters, which is below the default min_buffer_size of 100 characters.
  • The system waits for more input. If no additional text arrives within the time set by max_buffer_delay_in_ms (e.g., 500 ms), the system sends whatever text it has.

Result: After 500 ms, “I just wanted to say that” is sent for audio generation to avoid further delay. This mechanism ensures fast responses (good TTFB) without compromising too much on naturalness or context.