TTS providers

Every TTS plugin here is a wire protocol over the shared @kuralle-syrinx/tts-core lifecycle — multi-context streaming and usage billing work identically regardless of which one you pick.

Cartesia

import { CartesiaTTSPlugin } from '@kuralle-syrinx/cartesia';

const tts = new CartesiaTTSPlugin(socketFactory);
await tts.initialize(bus, {
  api_key: process.env.CARTESIA_API_KEY!,
  voice_id: process.env.CARTESIA_VOICE_ID,
  model_id: 'sonic-3',
  language: 'en',
});

Key	Default	Notes
api_key	—	Required.
voice_id	a premade Cartesia voice	Override with your own voice.
model_id	sonic-3
language	en
output_container / output_encoding	raw / pcm_s16le
cartesia_version	2024-06-10	API version header.

ElevenLabs

Multi-context WebSocket streaming — several concurrent TTS contexts on one connection, which is what makes fast barge-in cheap.

import { ElevenLabsTTSPlugin } from '@kuralle-syrinx/elevenlabs';

const tts = new ElevenLabsTTSPlugin(socketFactory);
await tts.initialize(bus, {
  api_key: process.env.ELEVENLABS_API_KEY!,
  voice_id: process.env.ELEVENLABS_VOICE_ID,
  model_id: 'eleven_flash_v2_5',
  sample_rate: 16000,
});

Key	Default	Notes
api_key	—	Required.
voice_id	a free-tier-accessible premade voice	Library voices (e.g. Rachel) require a paid plan.
model_id	eleven_flash_v2_5
sample_rate	16000

Usage is billed on audio received, not on the provider’s isFinal flag — a rejected generation never bills.

Deepgram Aura

import { DeepgramTTSPlugin } from '@kuralle-syrinx/deepgram';

const tts = new DeepgramTTSPlugin(socketFactory);
await tts.initialize(bus, {
  api_key: process.env.DEEPGRAM_API_KEY!,
  model: 'aura-2-thalia-en',
});

Key	Default	Notes
api_key	—	Required.
model	aura-2-thalia-en	Picks both voice and language.

Gemini

Non-streaming (chunked): the full text is sent and the complete audio comes back in one response.

import { GeminiTTSPlugin } from '@kuralle-syrinx/gemini';

const tts = new GeminiTTSPlugin();
await tts.initialize(bus, {
  api_key: process.env.GEMINI_API_KEY!,
  model: 'gemini-3.1-flash-tts-preview',
  voice_name: 'Kore',
  sample_rate: 24000,
});

Key	Default	Notes
api_key	—	Required.
model	gemini-3.1-flash-tts-preview	Also: gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts.
voice_name	Kore	One of Kore, Puck, Charon, Fenrir, Leda, Aoede, Zephyr, Orus.
instruction	empty	A style/persona lead-in prepended to every utterance. Empty by default — set it explicitly if you want one.
language_code	—
sample_rate	24000

Grok (xAI)

import { GrokTTSPlugin } from '@kuralle-syrinx/grok/tts';

const tts = new GrokTTSPlugin(socketFactory);
await tts.initialize(bus, {
  api_key: process.env.XAI_API_KEY!,
  voice_id: 'eve',
  sample_rate: 16000,
});

Key	Default	Notes
api_key	—	Required.
voice_id	eve
sample_rate	16000

Any OpenAI-compatible endpoint

OpenAICompatibleTTSPlugin speaks the OpenAI POST /v1/audio/speech streaming contract, so it works against OpenAI’s own TTS or any self-hosted / third-party endpoint that implements the same API.

import { OpenAICompatibleTTSPlugin } from '@kuralle-syrinx/openai-tts';

const tts = new OpenAICompatibleTTSPlugin();
await tts.initialize(bus, {
  api_key: process.env.OPENAI_API_KEY!,
  model: 'gpt-4o-mini-tts',
  voice: 'alloy',
});

Key	Default	Notes
base_url	https://api.openai.com/v1	Point this at a self-hosted endpoint to use a different backend.
api_key	—
model	gpt-4o-mini-tts
voice	—
response_format	pcm
source_sample_rate_hz	24000	The provider’s output sample rate.
sample_rate	16000	Engine rate to resample to.
tempo	1.0	Pitch-preserving time-stretch (0.5–1.5) via a built-in WSOLA stretcher — use 0.9 to slow speech 10% without changing pitch.
extra_body	—	Merged into the request body last; use it for any provider-specific field the plugin doesn’t enumerate.

Choosing between them

• Want the widest voice library and fastest barge-in? ElevenLabs’ multi-context streaming.
• Want a large multilingual catalog with per-language model selection? Cartesia.
• Already on Deepgram for STT? Aura keeps you on one vendor and one bill.
• Have a self-hosted or fine-tuned TTS model? OpenAICompatibleTTSPlugin against your own base_url — no custom plugin required as long as it speaks the OpenAI speech API shape.

See Usage & pricing for cited per-character rates.