Do you need an internet connection?
Only for setup. The distinction matters, because "offline" is used loosely by many products.
Needs internet, once or occasionally
- Downloading the AI model on first launch, 0.5 to 3 GB depending on size
- The login link and subscription check
- App updates
Never needs internet
- Transcribing an audio or video file
- Speaker identification
- Summaries
- Editing and exporting
The simplest test for any tool: disconnect from the network and transcribe a file. If it works, the processing is local. If it fails, "local" was a marketing word.
Local transcription tools compared
Every tool below runs Whisper, OpenAI's open-source speech model, on your own machine. They differ in interface, platforms and whether they identify speakers.
| Tool | Interface | Mac + Windows | Speaker labels | Price |
|---|---|---|---|---|
| Scrieb | Desktop app | ✓ | ✓ | Free 1 h/day; Pro subscription |
| MacWhisper | Desktop app | Mac only | Pro only | Free / Pro |
| Buzz | Desktop app | ✓ | ✓ | Free |
| WhisperX | Command line | ✓ | ✓ | Free |
| Whisper CLI | Command line | ✓ | ✗ | Free |
Which one? Comfortable in a terminal: Whisper CLI, or WhisperX if you need speaker labels. Want a free open-source app: Buzz, which now labels speakers too. On a Mac and working alone: MacWhisper is excellent. Need speaker labels plus summaries, DOCX export and a signed installer, on Windows or across a mixed Mac and Windows team: Scrieb. Full breakdown in Best Local Transcription Software.
Which Whisper model to use
"Local transcription model" usually means one of the Whisper sizes. Bigger is more accurate and slower; the right choice depends on your hardware and your audio.
| Model | Accuracy | Speed | Use when |
|---|---|---|---|
| tiny / base | Low | Very fast | Quick drafts, clean audio, weak hardware |
| small | Good | Fast | Laptops without a GPU, English, clear recordings |
| medium | Very good | Moderate | The default for CPU-only machines and most languages |
| large-v3 | Best | Slow without GPU | Accents, technical vocabulary, noisy audio; any machine with a GPU |
Scrieb lets you switch model size in Settings, so a fast draft on a laptop and a large-v3 pass on a desktop with a GPU use the same app.
Hardware: what your computer needs
Mac
Apple Silicon (M1 and later) is the sweet spot: fast, quiet, fully local. Intel Macs work with smaller models. 8 GB RAM minimum, 16 GB comfortable for large-v3.
Windows with a GPU
An NVIDIA RTX 20-series or newer runs 5× to 15× real time: a one-hour file in 4 to 12 minutes. AMD and Intel GPUs are supported too.
Windows, CPU only
Roughly 0.5× to 1× real time with the medium model. A one-hour file takes one to two hours. Slower, but it works and stays offline.
Disk
Models take 0.5 to 3 GB once. Audio files and transcripts live wherever you put them; nothing is cached on a server.
Platform details: Offline transcription on Mac · Offline transcription on Windows
Local vs. cloud transcription
Cloud transcription
- Audio uploaded to remote servers
- Internet required for every file
- Possible cross-border data transfer
- Data processing agreement often required
- Per-minute billing or monthly caps
- Works on phones and weak laptops
Local transcription
- Audio stays on your device
- Works offline after setup
- No transfer, no third party
- No processing agreement for content
- Flat price, no minute limits
- Needs a reasonably modern computer
Accuracy is not the difference: most cloud services run the same Whisper family on their servers. The difference is who has access to the audio. For confidential content, that decides it. Switching from a cloud service? See the Turboscribe alternative comparison.
Who transcribes on their own computer
Lawyers
Client conversations and dictation that must not reach a vendor. Transcription for lawyers →
Researchers
Interviews with participants who were promised confidentiality. Interview transcription →
Meetings
Recording your own calls without a bot joining and uploading everything. Meeting transcription →
Anyone offline
Trains, courtrooms, hospitals, secure sites, remote locations. Offline transcription on Windows →
Pricing
One flat price for Mac and Windows. No cloud fees, no minute limits, speaker identification included.
Frequently asked questions
- Do I need an internet connection to transcribe?
- No. The software needs the internet once, to download the AI model, and for account functions such as the login link and updates. The transcription itself runs on your computer. In Scrieb you can disconnect completely and transcribe as normal.
- What is the best local transcription model?
- For most people, Whisper large-v3 when you have a GPU, and Whisper medium or small when you do not. Larger models are more accurate, especially with accents, technical vocabulary and background noise, but slower. Scrieb lets you pick the model size, so you can trade speed for accuracy per machine.
- Is there free local transcription software?
- Yes. Buzz is a free open-source desktop app for Mac and Windows, and Whisper's command-line tool and WhisperX are free for people comfortable with a terminal. Buzz now identifies speakers too, as do the free apps Vibe, noScribe and aTrain. Scrieb has a free plan (1 hour a day) and a Pro subscription, and adds summaries, DOCX/PDF export and a custom vocabulary.
- Is local transcription the same as on-premise transcription?
- Close. On-premise usually means a server in your own building; local means each computer transcribes its own recordings. Both keep audio off a vendor's cloud. Local needs no server to run or maintain, which is why small firms and teams use it: Scrieb installs on each Mac or Windows PC and processes everything there.
- Can it identify speakers offline?
- Only if the software includes a speaker identification step and runs it on the device. Whisper alone does not. Scrieb and Buzz run both transcription and speaker identification locally; the Whisper CLI does not label speakers.
- Can I transcribe audio without uploading it to the cloud?
- Yes. With software that runs on your computer, the audio file is read from your disk, processed by a model on your CPU or GPU, and the text is written back to your disk. Nothing is uploaded at any step.
- Will it run on an older computer?
- Yes, with a smaller model. On a laptop without a dedicated GPU, expect roughly real-time speed with the medium model, so a one-hour file takes about an hour. It still runs fully offline.
- How accurate is local transcription compared with cloud services?
- The same, when the same model is used. Most cloud services run Whisper or a close relative on their servers. Accuracy depends on audio quality and model size, not on where the model runs.