OpenVoiceOS STT HTTP Server

Turn any OVOS STT plugin into an HTTP microservice for speech-to-text and spoken-language detection.

Pair it with the companion client plugin to offload transcription from an OVOS device, or point existing tooling at the vendor-compatible endpoints below.

Contents

• Install
• Configuration
• Usage
• HTTP API
• AI Agent Integration: MCP & UTCP
• Vendor-compatible endpoints
• Docker
• Documentation
• Examples
• Credits

Install

pip install ovos-stt-http-server

The server only hosts plugins. Install at least one STT plugin alongside it:

pip install ovos-stt-plugin-fasterwhisper

Optional extras:

Configuration

The STT plugin is configured exactly as it would be inside an assistant, under mycroft.conf:

{
  "stt": {
    "module": "ovos-stt-plugin-deepgram",
    "ovos-stt-plugin-deepgram": {"key": "xxxxx"}
  }
}

Usage

$ ovos-stt-server --help
usage: ovos-stt-server [-h] --engine ENGINE [--lang-engine LANG_ENGINE]
                       [--host HOST] [--port PORT] [--multi]

options:
  -h, --help                 show this help message and exit
  --engine ENGINE            STT plugin to be used (required)
  --lang-engine LANG_ENGINE  audio language-detection plugin to be used (optional)
  --host HOST                host to bind (default: 0.0.0.0)
  --port PORT                TCP port (default: 8080)
  --multi                    load one plugin instance per language (more memory)

For example, to serve faster-whisper for transcription with matching audio language detection:

ovos-stt-server \
  --engine ovos-stt-plugin-fasterwhisper \
  --lang-engine ovos-audio-transformer-plugin-fasterwhisper

HTTP API

The native API is unauthenticated. Audio is sent as the raw request body.

POST /stt query parameters:

The body must be raw PCM (16-bit signed, mono). Example with a WAV file decoded to PCM on the fly:

# 16 kHz mono int16 PCM in body
curl -s --data-binary @speech.pcm \
  -H 'Content-Type: application/octet-stream' \
  'http://localhost:8080/stt?lang=en&sample_rate=16000&sample_width=2'

See examples/native_example.py for a runnable script that reads a WAV file and posts its PCM frames. Full reference: docs/index.md.

Transformer pipelines

The server can run OVOS transformer plugins around transcription, on every endpoint: audio transformers process audio before STT (an AudioLanguageDetector in the chain resolves lang=auto) and utterance transformers rewrite the transcript before it is returned. Opt-in via the standard mycroft.conf sections:

{
  "utterance_transformers": {
    "ovos-utterance-corrections-plugin": {}
  }
}

Enabling an utterance transformer server-side means clients receive a different transcript than the raw STT output. Use it for fleet-wide vocabulary corrections. See docs/transformers.md for when to run transformers server-side vs on-device and how to avoid double-processing.

AI Agent Integration

MCP: Model Context Protocol

Install the optional extra to expose the server as an MCP tool provider:

pip install "ovos-stt-http-server[mcp]"

When mcp is installed, the server automatically mounts an MCP endpoint at /mcp using the streamable-HTTP transport (compatible with both the legacy SSE path /mcp/sse and the newer POST /mcp format).

Connecting an MCP client

Claude Desktop / claude-code (claude_desktop_config.json)

{
  "mcpServers": {
    "ovos-stt": {
      "transport": "http",
      "url": "http://localhost:8080/mcp"
    }
  }
}

ovos-tool-adapters persona JSON

{
  "toolboxes": ["ovos-mcp-toolbox"],
  "ovos-mcp-toolbox": {
    "transport": "http",
    "url": "http://localhost:8080/mcp",
    "timeout": 30
  }
}

Available MCP tool

Example call (Python MCP client):

import asyncio, base64
from mcp.client.streamable_http import streamablehttp_client
from mcp import ClientSession

async def main():
    async with streamablehttp_client("http://localhost:8080/mcp") as (r, w, _):
        async with ClientSession(r, w) as session:
            await session.initialize()
            audio_b64 = base64.b64encode(open("speech.pcm", "rb").read()).decode()
            result = await session.call_tool("transcribe", {
                "audio_b64": audio_b64,
                "lang": "en-us",
            })
            print(result.content[0].text)

asyncio.run(main())

UTCP: Universal Tool Calling Protocol

No extra dependencies are required. Every running server exposes a UTCP manual at:

GET /utcp

The response is a UTCP-1.0 JSON document describing the stt, lang_detect, and status tools so any UTCP client can discover and invoke them without separate documentation. The url fields use the server's actual base URL, so the manual is correct even behind a reverse proxy.

Register a UTCP client's provider config at /utcp:

{
  "toolboxes": ["ovos-utcp-toolbox"],
  "ovos-utcp-toolbox": {
    "utcp_config": {
      "tool_providers": [
        {
          "name": "ovos-stt",
          "provider_type": "http",
          "url": "http://localhost:8080/utcp"
        }
      ]
    }
  }
}

Vendor-compatible endpoints

The server mounts compat routers under per-vendor prefixes so existing tools and SDKs that already target a cloud STT API can be pointed at your local OVOS instance with only a base-URL / endpoint override. Every router accepts (and silently ignores) the vendor's auth token. Authentication is the job of your reverse proxy.

A runnable script for each lives in examples/. Full endpoint reference, per-vendor notes, and network-redirect recipes: docs/api-compatibility.md.

Docker

Any plugin can be served with a small Dockerfile:

FROM python:3.11-slim

RUN pip install --no-cache-dir \
    ovos-stt-http-server \
    ovos-stt-plugin-fasterwhisper

EXPOSE 8080
ENTRYPOINT ["ovos-stt-server", "--engine", "ovos-stt-plugin-fasterwhisper"]

Build and run:

docker build -t my-stt-server .
docker run -p 8080:8080 my-stt-server

Each plugin can ship its own Dockerfile in its repository using ovos-stt-http-server as the base.

Documentation

Examples

examples/ holds one runnable script per vendor router (driving each vendor's real client SDK where one exists) plus a native-API script. See examples/README.md.

Credits

Developed by TigreGótico for OpenVoiceOS.

This project was funded through the NGI0 Commons Fund, a fund established by NLnet with financial support from the European Commission's Next Generation Internet programme, under the aegis of DG Communications Networks, Content and Technology under grant agreement No 101135429.