Turn any video or audio into text on your own machine with OpenAI's Whisper. No cloud, no file-size cap, no paywall, nothing ever uploaded.
A premium, private alternative to tools like Otter.ai: a clean local web app that transcribes your files with Whisper, plays them back in sync down to the word, summarizes them with a local LLM, and keeps every transcript in a searchable library. Everything runs on your machine.
In plain terms: This is a free app you run on your own computer that turns recordings into written text, then helps you read, search, and summarize them. Because nothing is sent to the internet, your private audio and video stay with you.
Word-level karaoke: the transcript lights up word by word, in time with the audio.
A single identity in two themes: Editorial Light by day, Studio Dark by night. The brand indigo carries the design; a live cyan accent marks anything in motion.
Every finished transcript gets an optional one-click summary, chapters, and action items from a local Ollama model, and is saved to a library you can search across.
 A local library of every transcript, searchable across all of them. |
 Drop one or several files and transcribe them at once. |
In plain terms: Most online transcription services limit how big your file can be, charge for the best quality, and require you to hand your recordings to their servers. That is a poor fit when the recording is long or private.
Cloud transcription tools cap file size, hide the good models behind a paywall, and make you upload private recordings to someone else's server. For long talks, client calls, or anything sensitive, that is the wrong trade.
In plain terms: You drag in a recording and get an editable, searchable transcript that plays back in step with the audio, and you can save it in six common file types. The original file is removed once the job finishes.
- Drag and drop a video or audio file of any length (a 30-minute talk is fine).
- Transcribe several at once: drop a batch and each file opens in its own tab,
with a status rail showing every job. A small worker pool runs them a few at a
time and queues the rest (tune it with
MAX_CONCURRENT, default 2). - Runs OpenAI Whisper locally (via
faster-whisper) on your GPU when one is available (a 30-minute file can finish in a couple of minutes), and falls back to CPU automatically. - Live transcript streams in as it runs, with a progress bar, time estimate, and a GPU/CPU badge.
- Synced media player with word-level karaoke: the file plays back in the browser and the current line highlights as it goes, with the exact word lighting up in time with the audio. Click any line to jump there.
- Speaker diarization (optional): tick "Identify who spoke" and each line is labelled and color-coded by speaker, powered by a local, ungated ONNX model (sherpa-onnx). No Hugging Face token, no account, no cloud. See docs/adr/0001-local-speaker-diarization.md.
- Edit the transcript inline, search it with highlighting, and copy it with or without timestamps.
- Export to plain text (
.txt), subtitles (.srt), web captions (.vtt), Markdown (.md), JSON with timestamps (.json), and Word (.docx). - AI insights (optional): a one-click summary, clickable chapters, and action
items for any finished transcript, generated by a local Ollama
model (default
qwen3:8b). Stays entirely on your machine. If Ollama is not running, the button explains how to start it and everything else keeps working. - Library with global search: every finished transcript is saved to a small local database, so you can reopen any past transcript and search across all of them at once. Open one from the library and export, copy, or summarize it again.
- Source files are deleted after each job. Nothing ever leaves your computer.
- Ships with a CPU Docker image so it runs anywhere, plus an optional GPU image.
In plain terms: The diagram below shows the path a file takes, from dropping it in, to pulling out the sound, to writing the words, to the transcript you can edit and export. All of it happens on your computer.
flowchart LR
UP["Drop video / audio<br/>(any length)"] --> FF["ffmpeg<br/>extract audio"]
FF --> W["faster-whisper (Whisper)<br/>GPU, auto CPU fallback"]
W -->|"stream segments (SSE)"| UI["Browser UI<br/>live transcript + synced player"]
UI --> EDIT["edit, search, click to seek"]
EDIT --> EX["Export<br/>txt, srt, vtt, md, json, docx"]
UP -. deleted after job .-> X["(source removed)"]
The browser embeds your file with a local object URL, so playback and sync happen
entirely on your machine. Text exports are built in the browser (so they honor your
edits); the Word export is built locally by the server with python-docx.
In plain terms: These are the two ways to install and start the app, one for Windows and one that works on any system. Pick whichever matches your setup and follow the numbered steps.
- Install Python 3.11 or newer and
ffmpeg (on your PATH, or set
FFMPEG_PATH). - Double-click
setup.batonce. It builds a virtual environment and installs the pinned dependencies. - Double-click
run.bat. A browser opens at http://localhost:8765. - Drop a file, pick a model and language, click Transcribe.
docker compose up --build # CPU image, then open http://localhost:8765Transcripts are written to your local output/ folder, and Whisper models download
once into a persistent cache volume (never re-downloaded). For NVIDIA GPUs, an
optional image is provided (needs the NVIDIA Container Toolkit):
docker compose --profile gpu up --build gpu| Model | Speed | Accuracy | Use it for |
|---|---|---|---|
| Tiny / Base | fastest | rough | quick drafts |
| Small | fast | good | the sensible default |
| Medium | slower | better | cleaner results |
| Large v3 | slowest | best | important transcripts (great on a GPU) |
Models download once (Small ~0.5 GB, Large v3 ~3 GB) and are cached after that.
In plain terms: This is the list of the main software pieces the app is built from, grouped by the job each one does. It is here so engineers can see the parts at a glance.
| Layer | Stack |
|---|---|
| Server | Python 3.11 + FastAPI, server-sent events for live streaming |
| Transcription | faster-whisper (Whisper) + ffmpeg, GPU/CPU auto-select with a real CUDA self-test |
| UI | Local web page (HTML/CSS/JS), light theme with a dark toggle, synced player |
| Exports | txt / srt / vtt / md / json in the browser, docx via python-docx |
| Packaging | Docker (CPU and optional GPU), persistent model-cache volume |
In plain terms: There is a quick built-in check you can run to confirm the core pieces still work before you rely on the app. It finishes fast and needs no internet.
A fast, offline smoke test verifies the formatters and the real ffmpeg to faster-whisper pipeline:
python test_smoke.py # prints SMOKE TEST PASSED, exits 0 on successSee TESTING.md for the full manual checklist.
In plain terms: This is a short, honest summary of how finished the project is right now and what already works. It helps you know what to expect.
Working, tested, and containerized. Runs fully offline once a model is cached. Light theme by default with an optional dark mode, synced playback, inline editing, search, and six export formats. CPU image verified; optional GPU image provided.
In plain terms: This section describes a bigger setup for handling many recordings, or several people submitting at once, where extra helpers start up only when there is work and shut off when idle. The everyday app above does not need this.
The single-process app above is one shape of this pipeline. For a backlog of long recordings, several people submitting at once, or a GPU you want to keep busy without paying for it while idle, there is a second shape: an async, queue-based service that autoscales Whisper workers from zero on queue depth and back to zero when quiet.
flowchart LR
C["Client"] --> API["transcriber-api<br/>(no inference)"]
API -->|"enqueue"| Q[("Redis queue")]
KEDA["KEDA"] -.->|"scale 0..N"| WK["worker pods<br/>faster-whisper"]
Q -->|"consume"| WK
WK -->|"artifacts"| ST[("Shared storage")]
API --> ST
Same transcribe_core, wrapped in a thin FastAPI, a Redis work queue, and a
KEDA-scaled worker pool. Proven on Docker Desktop Kubernetes: zero to four workers
and back to zero on queue depth, jobs transcribed and returned through the API. Try
it with make k8s-up && make demo. No cluster? The same queue backend also runs on
plain Docker with make docker-queue (API on http://localhost:8001). Full writeup,
chart, and the captured scaling proof are in deploy/README.md.
MIT. See LICENSE.
Built by Jesse Jolly · SFX Tech Innovation · LinkedIn