Piper
nigiri.piper
Fast neural text-to-speech engine that runs entirely on CPU, producing natural speech from small ONNX voice models at real-time speed on hardware as modest as a single-board computer.
A Docker Compose file for Piper alone, generated from the catalog and checked by the safety inspector. Sign in and use Kitchen Credits to activate it.
Uses the Medusa cart and payment-session path. This test control does not create an order or capture payment.
Piper phonemises input text and runs a VITS model exported to ONNX, which keeps voices to tens of megabytes and synthesis fast enough to feel immediate without a GPU. Voices are published per language and speaker at several quality levels, so a deployment trades model size against fidelity explicitly rather than through a hidden setting. Everything runs locally with no network calls, which is why it became the default speech engine for offline voice assistants. It is invoked as a command reading text and writing audio, or wrapped in a small server for HTTP use.
You know it worked when
- The binary reports its version.
- A downloaded voice model and its configuration load without error.
- A line of text produces an audio file that plays intelligibly.
- Synthesis completes faster than the duration of the audio it produces.
- A second voice model can be selected without reconfiguring the container.
Known sharp edges
- The image is built from a repository archived in October 2025, so it receives no fixes, including for parse errors reported against newer voice files.
- Voice models are downloaded separately and carry their own licences, several of which are more restrictive than the engine's and some of which forbid commercial use.
- It is a command-line binary rather than a service, so a container started without a wrapper exits immediately and exposes no endpoint.
- Each voice model file must be paired with its matching JSON configuration; a missing configuration fails at load rather than falling back.