# gigastt

A local Russian speech-to-text engine written in Rust, built on the GigaAM v3 model and ONNX Runtime, usable as a command-line transcriber, an HTTP and WebSocket server, or an embedded library from several languages.

gigastt is a speech-recognition engine for Russian that runs entirely on-device, powered by the GigaAM v3 model through ONNX Runtime, with a roughly 225-megabyte default INT8 model and no cloud API or key required. It can transcribe a single file to text or subtitle formats, batch-process or watch a folder of recordings, or run as a server exposing HTTP and WebSocket transcription endpoints, including an OpenAI-compatible upload endpoint.

It supports common audio formats including WAV, MP3, FLAC and Opus, and offers several recognition heads: a base model for Russian, a variant that adds punctuation and casing, and multilingual variants covering Russian, English, Kazakh, Kyrgyz and Uzbek, with optional punctuation restoration, text normalization, hotwords and speaker diarization. Beyond the CLI and server, it can be embedded directly from Rust, C, Node and Python, with packaging guides for Swift and Kotlin, and runs on CPU by default with optional CUDA, CoreML and NNAPI backends.

Website: https://github.com/ekhodzitsky/gigastt