New: AI video tools. Cut, clip, caption and dub your videos. See all video tools →
Voice Creator Pro logoVoice Creator Pro
Try For Free

Desktop API

Runs inside the desktop app on your own computer at 127.0.0.1:8100, with no API key. Also covers subtitles, clips and audiobooks.

OpenAPI spec · Base URL http://127.0.0.1:8100

HTTP API for voice cloning, voice design, text-to-speech, transcription, audiobooks, dubbing, clips and AI Cut.

Base URL

The API runs on your computer at http://127.0.0.1:8100. Change the port in Settings or start the app with --api-port. Every route starts with /api/v1/, except the OpenAI-compatible POST /v1/audio/speech. Requests to an unknown path return the app's HTML page, not a 404.

Authentication

No key is needed unless you set one in Settings or start the app with --api-key. Once set, send it as X-API-Key: <key>, Authorization: Bearer <key> or ?api_key=<key>. GET /api/v1/health and GET /api/v1/system/version never need a key.

Quick Start Workflow

  1. Check health: GET /api/v1/health
  2. Pick a model: GET /api/v1/models lists the models and what each supports
  3. Generate speech using one of:
    • POST /api/v1/tts/generate to use a built-in speaker voice
    • POST /api/v1/generation/clone to clone a voice from reference audio or a saved voice
    • POST /api/v1/generation/design to create a voice from a text description
    • POST /api/v1/conversion/convert to convert existing speech to another voice
  4. Get audio: Pass ?sync=true to get audio in the response, or poll GET /api/v1/generation/progress then GET /api/v1/generation/result

The requested model loads automatically, so the first request after a model switch takes longer. POST /api/v1/model/load loads one ahead of time.

Sync vs Async

All generation and transcription endpoints accept ?sync=true. In sync mode, the request blocks until completion and returns the result directly. Generation has a hard 600s cap; transcription uses a 300s inactivity timeout (resets while progress events are flowing) so long audio can finish even if it takes longer than 5 minutes wall-clock. In async mode (default), the request returns immediately and you poll for progress/results.

Streaming

For real-time audio streaming, use the /stream variants (e.g. POST /api/v1/generation/clone/stream). These return a session_id. Connect to GET /api/v1/streaming/{session_id}/audio to receive a chunked WAV stream as audio is generated. Call POST /api/v1/streaming/{session_id}/finalize when done to collect the complete audio as base64.

Audio Format

All audio input/output uses base64-encoded WAV unless otherwise noted. The streaming audio endpoint returns raw WAV (PCM 16-bit mono). The output audio format can be changed via PUT /api/v1/settings/audio-format.

Models

Multiple TTS models are available, each with different capabilities. Query GET /api/v1/models to list all models, or GET /api/v1/models/active for the current model's capabilities.

Models differ in:

  • Supported modes: clone, design, and/or TTS (built-in speakers)
  • Languages: varies by model and variant
  • Quality tiers: each model has its own set of variants
  • Generation parameters: model-specific tuning knobs (temperature, exaggeration, etc.)

Pass model_family on generation requests to target a specific model, or omit it to use the currently configured model.

Validation

All requests are validated against the target model's capabilities. Invalid languages, quality tiers, or generation parameters return a 400 error listing the valid options.

Agents

POST /api/v1/speak speaks text out loud in a saved voice. POST /v1/audio/speech accepts OpenAI text-to-speech requests, so OpenAI clients work by pointing their base URL at this server. An MCP server for AI agents runs at /mcp.

Live avatars

Connect a WebSocket to /api/v1/talking-head/live (pass the key as ?api_key= when one is set) to get a talking face back in real time. Send JSON messages and audio, and receive video frames with their audio.

  1. Start: {"type": "start", "avatar_id": "av_..."}, or "stock_id", or "image_base64". Optional: "model" (lite default, or pro), "size" (512, or 1024 with pro), "idle" (loop default, live or none), "pace" (realtime default, or fast to get frames as quickly as they render), "framing" (face default for a square close-up, or photo for the whole photo with the face blended back in, up to 1280 px on its long side; ready gives width and height) and "voice", "language", "tts_model" and "tts_quality" (a quality tier such as higher_quality or faster, the app setting if left out) as defaults for say. The voice model loads during start so the first say is quick. Send "warm_tts": false if you only send your own audio. The server answers with loading messages while models download and load, then ready.
  2. Speak: {"type": "say", "id": "m1", "text": "...", "voice": "Morgan", "language": "English"} speaks in a saved voice. tts_model and tts_quality can also be set per say. Or send your own speech: {"type": "audio_begin", "id": "a1", "sample_rate": 24000}, then binary messages of 16-bit mono PCM, then {"type": "audio_end"}.
  3. Receive: binary frames, plus started and done messages for each id. Each frame is a 20-byte little-endian header (kind u8 with 1 for speech and 2 for idle, id_length u8, reserved u16, epoch u32, index u32, audio_offset u32, jpeg_length u32), then the id, then a JPEG, then 40 ms of 16-bit mono PCM at the sample_rate from ready. Idle frames carry no audio.
  4. Control: {"type": "interrupt"} stops the current speech and drops queued speech, {"type": "clear"} drops only the queue, and {"type": "played", "samples": N, "epoch": E} reports how many speech samples you have played in the current epoch. With realtime pace the server stays about one second ahead of what you report, or of the clock if you never report.

After an interrupt the server sends interrupted with a new epoch. Drop any frames with an older epoch. One live session runs at a time.

Endpoints

Health

System

Setup

Settings

Model

Generation

Clone

Design

TTS

Transcription

Streaming

Voices

History

Cache

GPU Install

Conversion

Audio

Translation

YouTube

Audiobook

Dubbing

Clips

AI Cut

Subtitles

Images

Agent

Avatars

Uploads

Health check GET

Returns {status: "ok"} if the server is running.