Speech recognitionAvailable
NanoASR
An offline speech recognition server built on sherpa-onnx. A single binary, CPU inference, word-level timestamps and an embedded web interface.
What it does
Offline and private
No audio leaves the machine. Models run locally; the network is only used to fetch model weights the first time.
Built for recordings
Calls, interviews and meetings from one minute to half an hour, mono or stereo, including 8 kHz telephony.
Who said what
Word-level timestamps with confidence, speaker diarization, and SRT or VTT subtitles that carry speaker labels.
Clean Russian text
Punctuation and casing, and inverse text normalisation: numbers, dates, times, currency, phone numbers and percentages.
Streaming too
Realtime recognition over a websocket while the audio is still arriving, with partial hypotheses.
One binary
The web interface is built in, four API dialects share one port, and a job queue handles webhooks. CPU by default, NVIDIA GPU with one flag.
Install
curl -fsSL https://github.com/usunrise88/NanoASR/releases/latest/download/install.sh | bashLinux, as a systemd service listening on 127.0.0.1:8080. It verifies the download, fetches the default models and prints the API keys. Windows has a PowerShell installer; see the repository.