Stream Speech

Reviewed streaming contracts are funded before connecting to the provider. Sessions are limited to five minutes, 8,000 submitted characters and 32 MiB of output. Final native completion settles the frozen tariff; interrupted or ambiguous outcomes retain the reservation. Use the exact supported selector and provider-native control frames below.

Authenticate the HTTP upgrade with Authorization: Bearer <api-key>. Native browser WebSocket cannot set that header; use a trusted server or a WebSocket client that supports custom upgrade headers. Never put an API key in the URL. The first client frame must be a session.start object with an exact Alibaba Qwen3 realtime, Amazon Polly, ElevenLabs, Cartesia, Deepgram, Hume, Inworld, MiniMax, Murf, Soniox, or xAI model. session.start accepts provider.data_region, as a request does.

Alibaba accepts native input_text_buffer.append, input_text_buffer.commit, input_text_buffer.clear, and session.finish events. Its native session, response, base64 audio, completion, usage, and error objects are forwarded unchanged. Qwen3 Flash uses built-in voices; Instruct adds common instructions and optimize_instructions; VC and VD require a custom voice ID created for the exact Alibaba enrollment model.

Amazon Polly uses its generative-only HTTP/2 bidirectional stream. Send { "type": "input_text", "text": "...", "flush": true }, optional { "type": "flush" }, then { "type": "session.finish" }; audio is emitted as base64 audio events and the final provider character count is emitted as stream.closed.

ElevenLabs standard and multi-context payloads are proxied according to their native protocols. Cartesia accepts its current generation and cancellation payloads, supports multiplexed context IDs, and uses the AudioLab-only { "type": "session.finish" } control to drain active contexts and settle the session.

Deepgram accepts native Speak, Flush, Clear, and Close controls; raw upstream audio is emitted as { "type": "Audio", "audio": "<base64>" }, while Metadata, Flushed, Cleared, and Warning objects retain their native fields. Its start frame requires a full Aura voice id (for example aura-2-thalia-en) and a pcm, ulaw, or alaw response format. session.finish maps to Deepgram's graceful Close.

Hume accepts native InputMessage text, voice, description, speed, trailing-silence, flush, and close controls; session.finish maps to Hume's native close:true.

Inworld supports up to five multiplexed contexts using native create, send_text, flush_context, and close_context objects; AudioLab forwards native context, audio, alignment, usage, and status results, and session.finish drains all open contexts. Inworld TTS-2 start-frame instructions are translated to bracketed steering on the first text frame in each context; TTS 1.5 Max and Mini reject steering.

xAI accepts native text.delta, text.done, and text.clear controls across multiple utterances on one connection; use session.finish after the last audio.done or audio.clear.

MiniMax accepts text or native task_continue frames, maps session.finish to task_finish, and forwards task_continued, task_finished, and task errors.

Murf accepts native text, voice_config, clear, context_id, end, min_buffer_size, and max_buffer_delay_in_ms fields (including camelCase voiceId), multiplexes independent contexts, and forwards native base64 audio and final events.

Soniox tts-rt-v2 accepts {"text":"...","text_end":true} after session.start. session.finish ends input. Native audio_end followed by terminated completes the stream; output is capped at two minutes and charged by generated audio duration at the published tariff.

Provider audio, timing, alignment, flush, completion, and error objects are forwarded without changing their provider fields.