Create a speech

Text to convert to speech. Up to 5,000 characters, or 50,000 with stream: true; longer input returns 400 input_too_long. May contain inline control tags, which don't count toward this limit. Long input is split at natural boundaries and synthesized in full. With timestamps: true the input is not split; we recommend keeping it to roughly 300 characters.

TTS model ID / public alias. Resolved to the served model server-side.

Preset voice name or custom voice ID. Mutually exclusive with ref_audio / ref_text when explicitly provided.

Output audio format. Streaming requires pcm.

If true, stream raw PCM chunks as they are decoded. Requires response_format to be pcm. Speed adjustment is not supported when streaming. Raises the input limit to 50,000 characters, so use it for long input.

Inline reference audio for one-off cloning: an http(s) URL, data URI, or base64-encoded raw audio bytes. Supported formats: AAC, WAV, MP3, FLAC, OPUS. Inline (base64 / data-URI) payloads: max 10 MB.

Recommended transcript of ref_audio.

Expand non-standard words into spoken form before synthesis, including numbers, dates, times, currency, units, URLs, email addresses, phone numbers, abbreviations, and symbols. Enabled by default; set to false to synthesize the text exactly as written. Text with nothing to normalize is passed through unchanged. Inline tags are preserved. Replaces the deprecated enable_tn.

ISO 639-3 language code of the input text, for example eng (English) or cmn (Mandarin Chinese). Used for text normalization. Defaults to null, in which case normalization infers the language from the text; set it when you know the language for the most reliable results. An unsupported code returns 400 invalid_field_value, including when normalization is disabled. Replaces the deprecated tn_language field and ISO 639-1 codes such as en.

Attach word-level timestamps to the response. The response becomes a JSON envelope carrying the word list and the audio base64-encoded in the requested response_format, instead of raw audio bytes. Requires stream to be false; combining the two returns 400 timestamps_streaming_unsupported. Available for English, Chinese, and Spanish; other languages return a null word list. Overrides normalize, which does not run when timestamps are requested.

Base64-encoded audio in the requested response_format. Decode it before writing it to a file.

Format of the encoded audio, echoing the requested response_format.

The text the model was given, which the timestamps index into.