vocald

Headless text-to-speech / speech-to-text server with a REST API and a native MCP endpoint. No GUI, no GPU, no torch — built to give AI agents and scripts a voice.

• TTS: Kokoro-82M (50 preset voices, 8 languages) on onnxruntime, int8

• STT: Whisper via faster-whisper (CTranslate2)

• API: REST + MCP (Streamable HTTP at /mcp) — same tools either way

• Auth: single bearer key over every route including /mcp; the docker image fails closed

• Footprint: 1.2 GB image, ~350 MiB RSS after a generation (the torch-based ancestor was 7.4 GB / 2.2 GiB)

Install

Docker

export VOICEBOX_API_KEY=$(openssl rand -hex 24)
docker compose -f docker-compose.headless.yml up -d --build
curl -H "Authorization: Bearer $VOICEBOX_API_KEY" http://127.0.0.1:17600/health

The Kokoro model is baked into the image (no first-request download); Whisper downloads into a named volume on first transcription. Without a key the container refuses to serve anything — deny by default.

Homebrew (macOS, native)

bash verify-cli.sh   # builds + installs the formula from your working tree, tests it end to end
# or step by step — see INSTALL.md
vocald-server        # 127.0.0.1:17493, open loopback dev mode
VOICEBOX_API_KEY=$(openssl rand -hex 24) vocald-server --host 0.0.0.0

Data lives in ~/.voicebox; model files (~115 MB) download on first generation. See INSTALL.md for the full environment reference.

Related MCP server: speaches-mcp

API in 30 seconds

AUTH="Authorization: Bearer $VOICEBOX_API_KEY"

# create a preset voice
curl -X POST -H "$AUTH" -H 'Content-Type: application/json' \
  -d '{"name":"Bella","voice_type":"preset","preset_engine":"kokoro","preset_voice_id":"af_bella"}' \
  http://127.0.0.1:17600/profiles

# synthesize, then download the artifact
curl -X POST -H "$AUTH" -H 'Content-Type: application/json' \
  -d '{"profile_id":"<id>","text":"Hello from vocald."}' \
  http://127.0.0.1:17600/generate          # -> {"id": ...}
curl -H "$AUTH" -o hello.wav http://127.0.0.1:17600/audio/<id>

# transcribe
curl -X POST -H "$AUTH" -F file=@hello.wav http://127.0.0.1:17600/transcribe

Preset voice ids: GET /profiles/presets/kokoro.

MCP

Point any MCP client (Claude Code, Cursor, …) at http://host:port/mcp with the Authorization: Bearer <key> header. Tools:

The tool names keep the voicebox. prefix for compatibility with existing clients and skills.

Podcast skill

vocald pairs with a Claude Code skill that turns any file, URL, or topic into a NotebookLM-style two-host audio conversation — the skill writes the dialogue, then voices each turn through vocald's API and plays it locally. See examples/podcast-skill/ for the skill definition; drop it into ~/.claude/skills/podcast/ and run /podcast <source>.

Limits, by design

Kokoro preset voices only — the voice-cloning engines and the LLM "personality" rewrite from the upstream project are not included in this slim build (requests degrade with clear errors, never crashes). If you need cloning, use the upstream desktop app.

Responsible use

Synthetic speech can be misused. See RESPONSIBLE_USE.md — in short: only voices you own or have permission to use, no impersonation, no fraud.

Credits & license

vocald is a headless, torch-free hard fork of jamiepine/voicebox (forked at b542768), which provides the desktop app this server grew out of. MIT license — see LICENSE.