MediaChef Audio to text
Transcribe audio to text — offline, on your own computer
MediaChef runs OpenAI's Whisper locally through whisper.cpp. Download a model once, then drop a recording or a video on the board and pick what you need out: plain text, SRT or VTT subtitles with timings, or JSON with the time of every segment. The audio is never uploaded, and there is no per-minute price.
Free and open source — read the code on GitHub
What you get out
TXT without timings, SRT and VTT with them, JSON with the start and end of every segment.
How to transcribe a recording
-
Download MediaChef
One file for macOS, Windows or Linux. whisper.cpp v1.7.6 is already inside the download; only the model is fetched separately.
-
Pick a Whisper model
Open Models and download one — small (488 MB) is the balanced default. It downloads once; after that transcription works with the network off.
-
Drop the recording or the video
Audio and video both work: MediaChef reads the sound out of a video itself, so a meeting recording and an MP4 take the same path.
-
Choose the output you need
“Transcribe audio to text” for TXT, “Transcribe audio to SRT subtitles” or “Transcribe audio to WebVTT” for timed captions, “Transcribe audio to JSON with timings” for tooling.
-
Press start and read the result
The text lands next to the file as talk.transcript.txt. The queue shows progress; if there is no speech in the file, MediaChef says so instead of writing an empty one.
Which Whisper model to pick
Four models ship in the catalogue, downloaded inside the app. Bigger means better text and a longer run on the same machine, so pick per task rather than once and forever.
| Model | Download | When to pick it |
|---|---|---|
tiny | 78 MB | A first pass on clear speech, or when you only need to find where a topic starts. |
base | 148 MB | Notes for yourself: readable text you will skim and edit anyway. |
small | 488 MB | The default in every transcription recipe, and the one most people keep. Interviews, meetings, lectures. |
large-v3-turbo | 1.62 GB | Text going out to someone else: subtitles you publish, quotes you cite. Tuned for Apple Silicon. |
Language is detected automatically, and you can pin it instead. Two extra recipes translate any language into English — as text or as timed subtitles — in the same single pass.
Why transcribe on your computer
Confidential audio stays confidential. Interviews, therapy notes, legal calls, anything under an NDA: the file is read by a process on your own machine and by nothing else.
No per-minute pricing. Cloud transcription bills by the minute. Locally, the tenth hour of audio costs what the first one did — nothing.
No length or size limit. A four-hour recording is one job in the queue, not a paid tier or a split-the-file workaround.
Works with no network at all. Once the model is on disk, transcription is offline: on a plane, in a lab, on an air-gapped machine.
Subtitles and text from the same run. SRT and VTT carry timings for a player, TXT is clean prose, JSON has the segment times for your own scripts.
FAQ
How accurate is Whisper transcription?
Accuracy scales with the model: tiny is a rough draft, small is the balanced default, large-v3-turbo is near-human on clear speech. Clean single-speaker audio does best; strong accents, people talking over each other and music under the voice all cost accuracy — as they do for any speech recogniser.
Which languages does it handle?
Whisper covers around a hundred languages and detects the language on its own; you can also pin it if the guess is wrong. Two recipes translate speech in any of them into English in one pass, as text or as SRT subtitles.
What formats does the transcript come in?
TXT for plain text, SRT and VTT with timestamps for players and video editors, and JSON with the start and end time of each segment for scripts and tooling.
Do I need internet?
Once, to download a Whisper model on the Models screen — 78 MB to 1.62 GB depending on which one. After that transcription runs entirely offline.
What if the recording has no speech?
MediaChef reports “No speech detected” instead of writing an empty file and marking the job green. Silence, music without vocals and a wrongly picked file all end the same honest way.
Turn that recording into text
MediaChef 0.4.0 — Whisper running locally, free and open source.
MediaChef is young: builds are not yet signed by Apple or Microsoft, so the first launch asks for confirmation — a plain-text how-to ships inside every download.