Skip to main content

Deepgram's Docs

On this pageOverview

Batch vs Streaming: Which Should I Use?

Flux TTS is served on /v2/speak over two transports against the same voices. They're not tiers — pick by how the audio is consumed.

The short answer

Building a voice agent or any live, conversational experience? Use streaming (WebSocket) — it streams audio as text arrives and keeps prosody consistent across turns.
Pre-rendering audio you know up front (IVR prompts, notifications, audiobook lines)? Use batch (REST).

Side by side

	Streaming (WebSocket)	Batch (REST)
Endpoint	`wss://api.deepgram.com/v2/speak`	`POST https://api.deepgram.com/v2/speak`
Input	Text streamed in as it's produced (LLM tokens)	One complete block of text
Output	Audio streams back incrementally	Full audio in one response
Time-to-first-byte	Low — playback starts before the full response exists	Whole clip generated before you get it
Interruption / barge-in	Yes — `Interrupt` with spoken-text feedback	N/A
Turn lifecycle & cross-turn context	Yes	N/A (stateless request/response)
Mid-stream control	`Configure` speed mid-session	Fixed per request (`speed` query parameter)
Encodings	Raw `linear16` / `mulaw` / `alaw`	Containerized/compressed too: `mp3` (default), `opus`, `flac`, `aac`
Operational model	Long-lived connection, lifecycle to manage	Stateless: simple retries, high fan-out

Choose streaming when

The text is produced incrementally (you're streaming from an LLM).
The user may barge in mid-response — Interrupt cancels in-flight synthesis and reports what they heard.
You want the lowest possible time-to-first-audio in a back-and-forth conversation.
You want tone to carry across turns.

Choose batch when

The full text is known before you synthesize.
You're pre-generating reusable assets (prompts, notifications, narration).
You want a stateless request/response with easy retries and high concurrency, and don't need incremental playback or interruption.

Suggest an edit