Piper is the neural text-to-speech engine that quietly ended up inside a large share of the self-hosted voice stack — Home Assistant, the NVDA screen reader, LocalAI, Open Voice OS and a long tail of hobby projects all speak through it — and piper1-gpl is its current home under the Open Home Foundation. The pitch has not changed since the Rhasspy days: synthesis runs entirely on the local machine, fast enough for a Raspberry Pi, with no API key and no audio leaving the device. The repository sits at 5,193 stars and 500 forks, is written mostly in C++, and ships as pip install piper-tts.

The architecture is deliberately unfashionable. Voices are trained with VITS and exported to ONNX Runtime, so a voice is just two files — an .onnx model and an .onnx.json config — that you drop next to the binary. Text is turned into phonemes by an embedded espeak-ng, which is where both the breadth and the license come from: espeak-ng's own GPL terms are why this repository is GPL-3.0 rather than the permissive license the older Piper carried. That single dependency is what lets a small model cover the ground it does — 44 language/locale entries in the voice catalogue, from en_US and de_DE through Georgian, Kazakh, Malayalam, Swahili and Nepali.

Release v1.7.0, published 15 August 2026, is the most interesting change in a while because it breaks the espeak monopoly on phonemization. Japanese now has a dedicated phonemizer built on OpenJTalk (pyopenjtalk-plus) behind a new ja extra, added precisely because espeak-ng has no kanji coverage — it reads out Unicode character names — and no pitch accent at all. Full-context labels are parsed for pitch accent and mapped onto IPA, which keeps Japanese voices compatible with the existing IPA-based warmstart rather than forking the training path. The same release added piper_create_options and piper_create_with_options() to libpiper, with the old piper_create() retained as a wrapper for ABI compatibility.

Surface area is broad for a project this size: a command-line interface, an HTTP web server, a Python API, a C/C++ API via libpiper, a documented path for training new voices, and community bindings such as Piper JNI for Java. That combination is why it shows up as the TTS backend in other people's products rather than as a product itself.

Two caveats belong on the record. First, the Open Home Foundation has an open call in the README stating it is looking for maintainers for Piper — a governance signal worth weighing before building a product on it. Second, licensing is not uniform: the project describes itself as intended for personal use and text-to-speech research, and each voice ships its own MODEL_CARD, some of which carry restrictive terms. The engine being GPL-3.0 tells you nothing about whether a specific voice is cleared for commercial use; that has to be checked per voice.