Skip to Content
Compatible APIsOpenAI

OpenAI-compatible API

Point any OpenAI SDK at Revolab with only a base_url and API-key swap. Same keys, same quotas, same metering as the native API — the compatible routes are thin adapters over the same engine.

Endpoints

POST https://api.revolab.ai/v1/audio/speech
POST https://api.revolab.ai/v1/audio/transcriptions

Set base_url="https://api.revolab.ai/v1" and pass your rvl_live_ key as the OpenAI API key. Errors use the OpenAI envelope, so SDK exception types (BadRequestError, RateLimitError, …) behave as expected.

Text to speech

from openai import OpenAI
 
client = OpenAI(base_url="https://api.revolab.ai/v1", api_key="rvl_live_...")
 
# Start here: mp3 is streamed as it is synthesized, so playback can begin
# in ~0.45 s no matter how long the text is.
with client.audio.speech.with_streaming_response.create(
    model="nada-1.0-pro",
    voice="<your-voice-id>",
    input="Streaming starts before synthesis finishes.",
    response_format="mp3",
) as response:
    for chunk in response.iter_bytes():
        play(chunk)  # first chunk arrives long before the last
 
# Buffered WAV — our default. Use it when you want the whole file at once
# (a Content-Length up front, and the generation kept in your History).
speech = client.audio.speech.create(
    model="nada-1.0-pro",
    voice="<your-voice-id>",
    input="Hello from Revolab.",
    response_format="wav",
)
speech.write_to_file("speech.wav")

Speech to text

from openai import OpenAI
 
client = OpenAI(base_url="https://api.revolab.ai/v1", api_key="rvl_live_...")
 
audio_file = open("speech.wav", "rb")
transcript = client.audio.transcriptions.create(
    model="aisyah-1.0-pro",
    file=audio_file,
)
print(transcript.text)
 
# stream=True works too (SDK-compatibility emulation: one full-text
# delta + done — latency is unchanged, our engines return the final
# transcript in one shot)
stream = client.audio.transcriptions.create(
    model="aisyah-1.0-pro", file=audio_file, stream=True
)
for event in stream:
    print(event.type, getattr(event, "delta", getattr(event, "text", "")))

Support matrix

AreaSupport
TTS outputEvery response_format OpenAI names: wav, mp3, opus (Ogg-encapsulated), aac (ADTS), flac and pcm (24 kHz mono s16le). The default is wav, not OpenAI’s mp3 — pass response_format explicitly if that matters to you. Delivery returns the audio bytes directly, as the native /v1/tts does.
TTS streamingFollows the container: mp3, opus, aac and pcm stream chunks as the model produces them. wav and flac are buffered, because both finish their header only once the audio is known — that is what gives them a real, seekable duration. stream_format="sse" additionally emits speech.audio.delta / done events and works with any streamed format. Streamed generations are not stored to your History.
STT formatsjson (default), text, verbose_json. srt / vtt and word timestamps are not available (engines return whole-utterance text).
STT streamingstream=true is an emulation: one transcript.text.delta with the full text, then transcript.text.done. No latency benefit.
ModelsRevolab model IDs only (not OpenAI’s) — see Models for the IDs available to you. OpenAI names like tts-1 or whisper-1 return a 400 listing valid models.
Model listingclient.models.list() works — GET /v1/models returns the active catalog in the OpenAI list shape.
Speed0.5–2.0 (OpenAI allows 0.25–4.0; out-of-range values are rejected, never silently clamped).

Voice IDs come from the Voice Library in your dashboard; quotas and error codes match the native API.

Rate limitsElevenLabs