The Pisary application is a local dictation tool for macOS that utilizes Sber's GigaAM model. It operates entirely on the CPU without requiring a GPU, ensuring that audio data remains on the user's device. The model file size is 204 MB, while the application itself occupies 32 MB.
Performance benchmarks show that a 6-second phrase is processed in 0.08 seconds, while a 30-second recording is transcribed in less than 0.5 seconds. Model loading takes 0.22 seconds. The software is distributed under the MIT license with open-source code available on GitHub.
The tool supports automatic punctuation and capitalization for Russian text. It also includes a secondary local neural network feature, allowing users to process text using either a 6.5 GB Sber model or a 2.5 GB Qwen model to refine dictated content, such as summarizing or editing, via voice commands.
The application features an over-the-air update mechanism and a visual indicator for recording status. Development evolved from a personal utility to a public project within two weeks, with external contributors already submitting code improvements.
