WebSocket
Stream expressive dialogue audio over a WebSocket by sending incremental text segments per registered voice.
The connection uses Eleven v3 dialogue models only (model_id must start with eleven_v3). The default model is eleven_v3_conversational.
Session setup
- After connecting, the first JSON message must include
voices(voice IDs to register for the session) and credentials if not already sent via headers or query string. - Optional
voice_settingsandpronunciation_dictionary_locatorsare only accepted on the first message. - For
eleven_v3_conversational, only one voice ID may be registered. Foreleven_v3, you may register up to 10 voices.
Streaming text
- Send
inputs: an array of{ "text", "voice_id", "new_turn"? }. Text for the same turn is buffered until the server has enough context (at least ~40 characters and 8 words), then partial audio chunks are emitted. - Set
new_turntotrue(or switchvoice_id) to finalize the current prosody segment and start a new speaker turn.
Control messages
flush: force generation of any buffered text without closing the socket.close_socket: flush remaining audio, send a final message, and close the connection.keep_alive: reset the 20 second receive timeout (no generation).
Authentication
Use the xi-api-key or Authorization header, single_use_token query parameter, or include xi_api_key, authorization, or single_use_token in the first message body (same pattern as Text to Speech WebSocket). Anonymous sessions are rejected.
For non-streaming dialogue over HTTP, see Create dialogue and Stream dialogue.
핸드셰이크
헤더
쿼리 매개변수
Identifier of the model that will be used, you can query them using GET /v1/models. Must be a v3 model.
Output format of the generated audio. Formatted as codec_sample_rate_bitrate. So an mp3 with 22.05kHz sample rate at 32kbs is represented as mp3_22050_32.
Language code (ISO 639-1) used to enforce a language for the model and text normalization. If the model does not support the provided language code, it will be ignored. This parameter is not supported for multilingual_v2 models.
When true, character timing from the model may be attached to audio chunks as alignment (snake_case field names).
When enable_logging is set to false zero retention mode will be used for the request. This will mean history features are unavailable for this request, including request stitching. Zero retention mode may only be used by enterprise customers.
보내기
JSON text frame from client. The first frame must register voices; later frames stream inputs and control flags.
수신
Partial audio for the current generation, base64-encoded using output_format.
Sent after the final audio for a given turn has been sent. For streaming codecs (MP3, Opus, etc.), the persistent encoder buffers a small amount of audio across turn boundaries, so a few bytes belonging to the marked turn may arrive interleaved with the next turn's first audio chunk. Clients that need exact per-turn boundaries should request a PCM output format, which has no buffering.
Sent after close_socket once all audio has been flushed.