hi9841/trit is a Rust engine that turns speech into text and reads text aloud using only the CPU. The latest release listed in the provided repository information is v0.3.0. The features presented are running the Parakeet, Canary, Whisper, and Moonshine models without a graphics card or inference libraries. However, since the provided information does not list anything newly added or changed in v0.3.0, the changes in this release cannot be confirmed.
The core of this project is handling speech recognition and synthesis on the CPU without external inference libraries. Input is 16 kHz mono audio, and it also provides the ability to convert it into sentences and organize the recognition results. For speech synthesis, it is introduced as supporting the Kokoro and Meta voice models and VITS voices. It also provides a command-line tool and bindings for Rust and multiple languages. The implementation language is Rust, and the license field in the repository information is marked as NOASSERTION.
The repository introduction presents several public speech models and CPU-based operation, but the provided facts contain no per-model performance figures or comparison criteria. Therefore, this information alone cannot be used to compare processing speed or recognition quality against other approaches. The license is also not clearly identified, so it is difficult to judge the applicable scope before confirming performance and usage conditions.