Self-hosted transcription with word-level timestamps
Scriberr is a transcription server written in Go that you run on your own hardware and use from a browser. Give it an audio or video file, a YouTube link or a recording made in the web app, and it returns text with word-level timestamps, split by speaker when diarization is on. The models run on the machine hosting Scriberr. Its author wrote it to stop paying for the cloud transcription plan of his Plaud Note recorder: $100 a year for 20 hours a month, or $240 for unlimited use.
Meetily transcribes a meeting live on the computer sitting in on it, and Murmure turns dictation into text on the desktop. Scriberr starts from recordings that already exist.
The comparison of local speech-to-text tools sets out what each one asks of the machine.
Click-to-seek transcripts from three local speech engines
- Three local engines: WhisperX (90+ languages), plus NVIDIA's Canary 1B v2 and Parakeet TDT 0.6B v3. Scriberr registers Parakeet as English-only, so other languages go to WhisperX or Canary, both of which can also translate into English.
- Two diarization models: PyAnnote, up to 20 speakers, or NVIDIA's Sortformer, tuned for four voices.
- Word-by-word playback: the current word lights up as the audio plays, and Cmd/Ctrl-click on any word jumps playback to it. Transcripts export to JSON, SRT or TXT.
- LLM summaries and Q&A: Ollama locally, or any OpenAI-compatible API.
- REST API and webhooks: API keys, plus a
callback_urlthat receives the transcript when a job finishes, enough to wire Scriberr into n8n. A command-line client uploads every new file that lands in a watched folder.
Last release in December 2025, and one account only
GPU acceleration means NVIDIA. The docs provide a CPU Docker image, a CUDA image for GTX 10 through RTX 40 cards, and a separate image for the RTX 50 series. There is no image for AMD cards, and on macOS, Parakeet, Canary and PyAnnote install the CPU build of PyTorch. The docs give no memory figures; the code recommends 2 GB for WhisperX, 4 GB for Parakeet and 8 GB or more for Canary, where a GPU is "strongly" recommended.
The first start downloads several gigabytes. Each model family runs in its own Python environment, which Scriberr installs on first launch and keeps apart from app data. In the sample startup log from the docs, the Parakeet model weighs 2.3 GB and Canary 5.9 GB.
One account, and HTTPS expected. Registration only works while no user exists, and the maintainer confirmed in March 2026 that multiple users are not supported (issue #433). In production, cookies are flagged Secure: served over plain HTTP, Scriberr fails to load audio unless SECURE_COOKIES=false is set.
Hugging Face for PyAnnote, OpenAI as an option. PyAnnote needs a Hugging Face account, accepted model terms and an access token. According to the docs, the token is for downloading the models and no audio is sent. Two settings do send content off the machine: transcription through the OpenAI API, and summaries handled by a remote provider.
No release since December 2025, and a rewrite on the table. Version 1.2.0, from 17 December 2025, is the latest. Commits through March 2026 added, among other things, Mistral's Voxtral Mini, which no release includes. The maintainer, who authored 636 of the 828 commits, paused development in March 2026 and announced his return on 20 September. In discussion #485 he leans toward rebuilding the engine on onnx-asr or audio.cpp rather than working through the 21 open pull requests, and warns that the rewrite would "probably" bring breaking changes. Upgrading to v1.2 already forced one: deleting the old Python environment and splitting the volumes.