vibevoice-api 0.0.1

A model for speech generation with an AR + diffusion architecture.

VibeVoice OpenAI-Compatible TTS API

This is a VibeVoice OpenAI-compatible TTS API.

Community

Join the unofficial Discord community: https://discord.gg/ZDEYTTRxWG - share samples, ask questions, discuss fine-tuning, etc.

Installation

git clone https://github.com/vibevoice-community/VibeVoice-API
cd VibeVoice/

uv pip install -e .

Model Zoo

Model	Context Length	Generation Length	Weight
VibeVoice-1.5B	64K	~90 min	HF link
VibeVoice-Large	32K	~45 min	HF link

Getting Started

Run a local server that is compatible with the OpenAI audio API (client.audio.speech.create). It wraps VibeVoice to synthesize speech from text.

Start the server

python -m vibevoice_api.server --model_path vibevoice/VibeVoice-1.5B --port 8000

API base path (default: /v1)

All routes are mounted on /v1 by default. To override the prefix, set VIBEVOICE_API_BASE_PATH (leading slash required) before launching the server:

export VIBEVOICE_API_BASE_PATH=/api
python -m vibevoice_api.server --model_path vibevoice/VibeVoice-1.5B --port 8000

Clients must include the same prefix when constructing URLs. The static console is served at <base_path>/web/console.html.

Endpoints

POST <base_path>/audio/speech

Synthesize speech from text.

Request fields (OpenAI-compatible):

• model (string): model id or local path (e.g., vibevoice/VibeVoice-1.5B).
• voice (string): name mapped to a reference voice, a filesystem path (prefix with path: or absolute), or an alias from a voice map.
• input (string): the input text.
• response_format (string): wav, pcm (native), or mp3 / opus / aac (require ffmpeg).
• stream_format (string, optional): set to sse for Server-Sent Events (streamed base64 PCM chunks).
• extra_body (object, optional):
  • voice_path: absolute/relative path to a reference audio file.
  • voice_data: base64-encoded WAV bytes (optionally as a data URL).

Python example (OpenAI SDK ≥ 1.40):

from openai import OpenAI

base_path = "/v1"  # or your VIBEVOICE_API_BASE_PATH
client = OpenAI(base_url=f"http://127.0.0.1:8000{base_path}", api_key="<YOUR_API_KEY>")

speech = client.audio.speech.create(
    model="vibevoice/VibeVoice-1.5B",
    voice="Andrew",
    input="Hello from VibeVoice!",
    response_format="wav",
)

with open("out.wav", "wb") as f:
    f.write(speech.read())

Pure HTTP example (cURL):

curl -X POST "http://127.0.0.1:8000/v1/audio/speech"   -H "Content-Type: application/json"   -H "Authorization: Bearer <YOUR_API_KEY>"   -d '{
    "model": "vibevoice/VibeVoice-1.5B",
    "voice": "alloy",
    "input": "Hello!",
    "response_format": "mp3"
  }' --output out.mp3

Streaming (SSE): Set "stream_format": "sse" in the request body to receive a stream of SSE events carrying base64-encoded PCM audio chunks. A JS example client is provided in scripts/js/openai_sse_client.mjs.

Voice Mapping

You can define stable, human-friendly voice names via a YAML file that is auto-loaded on each request.

• Voice YAML mapping: You can use YAML to manage aliases or automatically scan multiple folders (see next section).

Search order (first found):

1. Path from VIBEVOICE_VOICE_MAP (relative to repo root or absolute)
2. ./voice_map.yaml
3. ./config/voice_map.yaml

Example (voice_map.yaml):

alloy: en-Frank_man
ash: en-Carter_man

aliases:
  promo_female: demo/voices/en-Alice_woman.wav

directories:
  - demo/custom_voices

Then call with voice: "alloy", or use extra_body.voice_path / extra_body.voice_data per request.

Formats

• wav, pcm: native outputs (no extra dependencies).
• mp3, opus, aac: require a working ffmpeg binary. Either ensure ffmpeg is on PATH or set VIBEVOICE_FFMPEG to the binary path.

Authentication & Admin (optional)

By default, API-key auth is disabled. To enable:

export VIBEVOICE_REQUIRE_API_KEY=1

With auth enabled, include Authorization: Bearer <YOUR_API_KEY> in client requests.

Admin key management (requires VIBEVOICE_ADMIN_TOKEN; routes respect your <base_path> and default to /v1):

List stored key hashes

curl -sS -H "Authorization: Bearer $VIBEVOICE_ADMIN_TOKEN"   http://127.0.0.1:8000/v1/admin/keys

Create/import a key (omit body to auto-generate with the given prefix)

curl -sS -X POST -H "Authorization: Bearer $VIBEVOICE_ADMIN_TOKEN"   -H "Content-Type: application/json"   -d '{"prefix": "sk-"}'   http://127.0.0.1:8000/v1/admin/keys

Revoke a key by stored hash

curl -sS -X DELETE -H "Authorization: Bearer $VIBEVOICE_ADMIN_TOKEN"   http://127.0.0.1:8000/v1/admin/keys/<key_hash>

Logs are written under logs/ and can be configured via:

• VIBEVOICE_LOG_DIR
• VIBEVOICE_LOG_PROMPTS=1
• VIBEVOICE_PROMPT_MAXLEN=4096

Notes

• Only TTS (/audio/speech) is implemented; there are no STT endpoints.
• Legacy root routes (e.g., /audio/speech, /metrics) remain for backwards compatibility, but new integrations should prefer the explicit <base_path>.

License

The source code and models are licensed under the MIT License. See the LICENSE file for details.

Note: Microsoft has removed the original repo and models. This fork is based off of the MIT-licensed code from Microsoft.