Transcribe audio offline and skip the cloud bill

What it does Scriberr is a self-hosted web app that transcribes audio and video locally using offline models such as Whisper, NVIDIA Parakeet, or Canary. It generates word-level timestamps, detects and labels individual speakers, and wraps the results in a polished interface with a built-in recorder, note-taking, and transcript playback that stays in sync with the text. You can also query transcripts via Ollama or OpenAI-compatible LLMs without leaving the app.

The interesting bit The project was born when the author bought a Plaud Note voice recorder and balked at its mandatory cloud transcription and subscription pricing. Rather than pay $100–$240 per year to send private recordings upstream, he built an offline-first alternative that bundles diarization, summarization, and automation hooks into a PWA that behaves like native software.

Key highlights

• Runs fully offline: no audio data is sent to third-party cloud providers during transcription
• Supports multiple local model backends, including Whisper, NVIDIA Parakeet, and Canary
• Includes speaker diarization, word-level timestamps, and a folder watcher for automation pipelines like n8n
• Progressive Web App support gives a native experience on both desktop and mobile, complete with dark mode
• Exposes APIs and integrates with local or remote LLMs for summarization and Q&A inside the app

Caveats

• Active development is paused while the author recovers from a layoff and searches for his next role; he is openly looking for contributors to keep the project moving
• The Go binary relies on a separate managed Python environment for model inference, so it is not a single-static-binary deployment
• Default secure-cookie behavior is strict enough to silently break audio streaming on plain-HTTP deployments

Verdict Self-hosters with a privacy bent and a spare GPU—or patience for CPU transcription—will find a capable alternative to cloud APIs. If you need active development velocity or a fully managed service, this is not it right now.