Text to speech, file transcription and live realtime sessions behind one HTTP API — with scoped keys, per-second billing and request logs.
xVoice implements the OpenAI audio API, so the official clients work against it unchanged:
from openai import OpenAI
client = OpenAI(api_key="sk_live_...", base_url="https://your-xvoice-host/v1")Throughout these pages the base URL is written $XVOICE_BASE_URL. It is your deployment's host with /v1 on the end.
| Page | What it covers |
|---|---|
| Quickstart | Signup to a working transcript, in about five minutes |
| Authentication | API keys, test and live environments, key hygiene |
| Errors | The error envelope, every code, and what to do about each |
| Models and voices | What is available, and what capabilities means |
| Rate limits and billing | Tiers, headers, credits, spending limits |
| Endpoint | Purpose |
|---|---|
POST /v1/audio/speech | Text to speech, whole file or streamed |
POST /v1/audio/transcriptions | Transcribe an audio file |
WS /v1/realtime | Transcribe a live stream as it is captured |
GET /v1/models | List models, or read one |
GET /v1/voices | List voices |
GET /v1/me | What a key is attached to |
/docs on your deployment serves the generated OpenAPI schema for these endpoints. It cannot describe WS /v1/realtime — OpenAPI has no vocabulary for WebSockets — so that one is documented only here.
Every endpoint has a working script on the example scripts page, built on the stock openai Python client:
export XVOICE_API_KEY=sk_test_...
export XVOICE_BASE_URL=https://your-xvoice-host/v1
python tts.py "Hello from xVoice." # writes output/speech.wav
python stt.py output/speech.wav
python realtime.py output/speech.wavorg_…, proj_…, key_…, req_…. The prefix says what the thing is, which makes a misplaced id obvious in a log.X-Request-Id, and every error body repeats it as request_id. It is the one thing worth keeping when something goes wrong: it ties a report to the exact request.{"object": "list", "data": [...]}. Where a list can grow without bound it is keyset-paginated with an opaque next_cursor.