Deepgram Aura TTS API Pricing & Conversational Voice Guide (2026)

Acoustic Architecture, WebSocket Buffering & Real-Time Voice Synthesis SLAs

Deepgram Aura TTS is specifically engineered for full-duplex conversational voice bots where audio generation latency must not exceed human conversational pause thresholds (~200ms). While traditional text-to-speech pipelines serialize text generation, token conversion, and mel-spectrogram synthesis into sequential blocking operations, Aura pipelines streaming text directly into its neural acoustic decoder. Over persistent bidirectional WebSockets, Aura begins streaming linear 16-bit PCM or Opus audio chunks within 90 milliseconds of text token dispatch.

Production Full-Duplex Audio Pipe in TypeScript

Pipe LLM text stream tokens directly into Deepgram Aura's WebSocket connection to achieve instantaneous acoustic feedback:

import WebSocket from "ws";

const ws = new WebSocket("wss://api.deepgram.com/v1/speak?model=aura-asteria-en&encoding=linear16&sample_rate=24000", {
  headers: { Authorization: `Token ${process.env.DEEPGRAM_API_KEY}` }
});

ws.on("open", () => {
  // Flush streaming LLM delta tokens as they arrive
  for (const token of tokenStream) {
    ws.send(JSON.stringify({ type: "Speak", text: token }));
  }
  ws.send(JSON.stringify({ type: "Flush" }));
});

ws.on("message", (audioBuffer: Buffer) => {
  // Push raw audio buffer directly to client playback sink
  mediaPlayer.write(audioBuffer);
});

Voice Interruption Handling & Buffer Clearing

In real-time conversational agents, handling user barge-in (when a user interrupts the speaking bot) requires rapid state synchronization. When speech-to-text detects user voice activity (VAD), the client application must dispatch a WebSocket `{"type": "Clear"}` message to Deepgram Aura immediately. This cancels pending synthesized frames and purges local audio ring buffers, preventing the agent from talking over the user and preserving a human-like conversational cadence.