Skip to Content
Compatible APIsCartesia

Cartesia-compatible API

The cartesia Python SDK works against Revolab with a base_url swap — bytes, SSE, the generation websocket, and batch speech-to-text, at Cartesia’s real paths.

Endpoints

POST https://api.revolab.ai/tts/bytes
POST https://api.revolab.ai/tts/sse
WS   wss://api.revolab.ai/tts/websocket
POST https://api.revolab.ai/stt

Bytes & SSE

from cartesia import Cartesia
 
client = Cartesia(api_key="rvl_live_...", base_url="https://api.revolab.ai")
 
fmt_wav = {"container": "wav", "encoding": "pcm_s16le", "sample_rate": 24000}
fmt_raw = {"container": "raw", "encoding": "pcm_s16le", "sample_rate": 24000}
 
# Buffered WAV
audio = b"".join(client.tts.generate(  # v2 SDKs: client.tts.bytes(
    model_id="nada-1.0-flash",
    transcript="Hello from Revolab.",
    voice={"mode": "id", "id": "<your-voice-id>"},
    output_format=fmt_wav,
))
 
# Server-sent events — base64 PCM chunk events + done
for event in client.tts.generate_sse(  # v2 SDKs: client.tts.sse(
    model_id="nada-1.0-flash",
    transcript="Streaming starts before synthesis finishes.",
    voice={"mode": "id", "id": "<your-voice-id>"},
    output_format=fmt_raw,
):
    if event.type == "chunk":
        play(event.audio)

Websocket & speech to text

from cartesia import Cartesia
 
client = Cartesia(
    api_key="rvl_live_...",
    base_url="https://api.revolab.ai",
    websocket_base_url="wss://api.revolab.ai",
)
 
with client.tts.websocket_connect() as ws:
    ws.send({
        "context_id": "turn-1",
        "model_id": "nada-1.0-flash",
        "transcript": "Hello over the websocket.",
        "voice": {"mode": "id", "id": "<your-voice-id>"},
        "output_format": {"container": "raw", "encoding": "pcm_s16le",
                          "sample_rate": 24000},
    })
    for msg in ws:
        if msg.type == "chunk":
            play(msg.audio)
        elif msg.type == "done":
            break
 
# STT (batch)
result = client.stt.transcribe(
    file=open("speech.wav", "rb"), model="aisyah-1.0-flash"
)
print(result.text)

Support matrix

AreaSupport
AuthBoth SDK generations work: v3 (Authorization: Bearer + cartesia-version header), v2 (X-API-Key), and ?api_key= on the websocket — all carrying a normal rvl_live_ key. The version header is accepted and ignored.
Output formatscontainer "wav" (buffered) or "raw" (chunked / streamed) with encoding pcm_s16le at 24000 Hz only. mp3, other encodings, and other sample rates return a 400 naming the supported combination.
Websocket contextsMultiple contexts per connection; each context synthesizes per request. continue:true pieces follow Cartesia’s max_buffer_delay_ms contract — 0 synthesizes each piece as it arrives (what the LiveKit plugin sends), otherwise pieces buffer until the final piece. Whitespace-only final pieces close the context with done.
Timestampsadd_timestamps is accepted but no word/phoneme timestamp events are produced (engines return whole-utterance audio/text).
Streaming STT/stt/websocket is not supported — the socket closes with a pointer to POST /stt (batch). No honest emulation preserves incremental transcription semantics.
ModelsRevolab names only. sonic-* and ink-* return a 400 listing valid models.

Building a voice agent with LiveKit? See the LiveKit agents guide — the plugin drives this websocket for you.

ElevenLabsLiveKit