Ownvoice

What the ownvoice MCP server does

The ownvoice MCP server connects an MCP client to the local ownvoice CLI. Its purpose is to make voice and identity-check operations available through one generic MCP tool, while the CLI supplies the underlying commands and output.

The CLI works with pocket-tts and PEFT LoRA adapters. It can check whether LoRA injection succeeds against pocket-tts's flow_lm module, train an adapter from a directory of WAV clips, and generate speech using a saved adapter. Training writes an adapter.safetensors file and a metadata.json file to the selected output directory.

The ownvoice MCP server is suited to local workflows where recordings, model files, and generated audio should remain on the machine. The supplied material does not describe a hosted service, account system, or API subscription requirement.

How it works

A client sends a request to the generic MCP tool, which wraps the ownvoice command-line workflow. The CLI has three relevant operations:

• check performs a CPU-only dry run of PEFT LoRA injection against pocket-tts. It does not train an adapter.
• train reads .wav clips from a required directory, trains a LoRA adapter, and evaluates a generated sentence against the reference voice.
• infer loads an adapter and synthesizes the requested text to a WAV file.

Each command supports JSON output for machine-readable results. A completed training run saves its adapter and metadata even when the similarity score is below the stated threshold. The check command needs no GPU; training benefits from an NVIDIA GPU, while inference and checking can run on CPU according to the README.

Setup and configuration

The documented CLI installation is:

pip install ownvoice-cli

Python 3.11 or newer is required. An npm package named ownvoice-cli provides a wrapper that bootstraps the Python CLI through uv or pipx; it is not a Node.js reimplementation. The README does not provide a separate MCP launch command, transport configuration, or MCP-specific environment variables.

Training requires a directory of clean WAV voice clips. The primary required option is --voice-clips; other settings include output location, epoch count, LoRA rank, alpha, dropout, learning rate, and evaluation text. Inference requires an adapter path and synthesis text, with optional output and reference-audio paths.

Tools and capabilities

The ownvoice MCP server exposes one generic MCP tool rather than a separately named MCP tool for each CLI subcommand. Through that interface, the documented CLI capabilities include:

• Checking pocket-tts and PEFT LoRA module compatibility.
• Training a LoRA voice adapter from local recordings.
• Reporting a similarity score and whether the result clears the usable threshold.
• Saving adapter weights and training metadata locally.
• Generating speech into a WAV file from a trained adapter.
• Returning structured JSON output for programmatic callers.

Limitations and notes

The ownvoice MCP server's exact request schema, startup command, transport, and error mapping are not included in the supplied material. Client compatibility therefore cannot be confirmed from the README alone.

Training quality depends on the supplied recordings and configuration. The README recommends roughly five to ten minutes of clean audio and requires consent for the voice being trained. A run can finish below the similarity threshold without exiting with an error; the adapter is still saved and can be tested with inference.

GPU acceleration is recommended for faster training, but it is not required for the compatibility check. The documented package and CLI details describe ownvoice 0.1.2, so behavior may differ in other versions.