How to Transcribe Audio to Text Offline on Windows (Whisper Guide)
Turn recordings, meetings, lectures and interviews into text on your own PC with OpenAI's Whisper — no uploads or per-minute fees. Model sizes, hardware, accuracy tips and easier options.
By Zeeshan Khalid · Updated 26 September 2026
Paid transcription services charge per minute and require you to upload your audio — awkward for confidential meetings, medical dictation, legal interviews or journalism sources. Since OpenAI released Whisper as open source, you can get comparable accuracy on your own PC, offline, for free.
This guide explains your options, from the command line to one-click apps.
What is Whisper?
Whisper is a speech-recognition model trained by OpenAI on 680,000 hours of multilingual audio. It:
- Transcribes 99 languages and can translate them into English
- Adds punctuation and capitalisation automatically
- Handles accents and background noise better than older offline engines
- Outputs plain text or timed subtitles (SRT/VTT)
The model weights are free under the MIT licence, and once downloaded they run with no internet connection.
Choosing a model size
Whisper comes in several sizes. Bigger is more accurate but slower and needs more memory:
| Model | Parameters | Approx. VRAM | Speed | Use it for |
|---|---|---|---|---|
| tiny | 39 M | ~1 GB | Fastest | Quick drafts, old laptops |
| base | 74 M | ~1 GB | Very fast | Clear speech, dictation |
| small | 244 M | ~2 GB | Fast | Best default for most people |
| medium | 769 M | ~5 GB | Moderate | Accents, noisy audio, other languages |
| large | 1.55 B | ~10 GB | Slow | Maximum accuracy |
English-only variants (tiny.en, base.en, small.en, medium.en) are slightly more accurate for English at the same size.
Option 1: Whisper from the command line
If you're comfortable with Python:
- Install Python 3.9+ and FFmpeg (for example
winget install ffmpeg). - Install Whisper:
pip install -U openai-whisper
- Transcribe a file:
whisper "meeting.mp3" --model small --language en --output_format txt
Use --output_format srt for subtitles, or all for every format. The first run downloads the model.
For much faster results on NVIDIA GPUs, you'll need a CUDA-enabled PyTorch install — which is where most people get stuck. faster-whisper, a reimplementation built on CTranslate2, is typically several times faster than the original with lower memory use and is what many apps use under the hood.
Option 2: Built-in Windows options (and their limits)
- Voice typing (Win + H) transcribes live speech into any text box. It doesn't transcribe files, and it processes audio online.
- Microsoft Word's Transcribe (Microsoft 365, web version) transcribes uploaded recordings — in Microsoft's cloud, with monthly limits.
Neither gives you private, offline transcription of existing files.
Getting accurate transcripts
Whatever tool you use, the recording matters most:
- Mic close to the speaker. A laptop mic across a room is the biggest single cause of errors.
- Reduce echo and background noise — soft furnishings help more than software.
- One speaker at a time. Crosstalk confuses every model.
- Set the language rather than relying on auto-detect for short clips.
- Step up a model size before spending time correcting a bad transcript.
- Check names and jargon. Whisper spells unfamiliar names phonetically; a quick find-and-replace fixes them.
What to do with the transcript
- Subtitles: export SRT and load it alongside your video. Use the free subtitle converter to convert to VTT for the web or fix timing offsets.
- Meeting notes: paste the text into your notes app, or into a local LLM to summarise. (See running local AI with LM Studio.)
- Search: plain-text transcripts make hours of recordings searchable with Ctrl+F.
- Need the audio in another format first? The audio converter extracts audio from MP4 video and converts formats in your browser.
The no-setup option
Setting up Python, FFmpeg and CUDA is a hurdle for many people, so I built Whispering Notes AI for Windows: drop in any MP3, MP4, WAV, MKV, FLAC, WEBM or AVI file and get a transcript. It offers five model sizes from Tiny to Large, 90+ languages with auto-detection, an editor to fix the text, export to TXT, SRT or JSON, and an AI summary of key points. It uses your NVIDIA GPU when available and falls back to CPU otherwise. After the one-time model download it runs fully offline, with a one-time price.
If your goal is subtitles for a video rather than a transcript, Subtitle Studio AI is the sibling app, with word-level timing, a cue editor, styling and burn-in — see how to add subtitles to a video automatically.
Frequently asked questions
- Is Whisper free to use offline?
- Yes. OpenAI released Whisper's code and model weights under the MIT licence. Once a model is downloaded, it runs entirely on your computer with no internet connection and no usage fees.
- Which Whisper model should I use?
- Start with 'small' — it's a good balance of speed and accuracy on most PCs. Use 'medium' or 'large' for difficult audio, accents or non-English languages if you have a capable NVIDIA GPU; use 'base' or 'tiny' on older laptops.
- Do I need a GPU to run Whisper?
- No, but it helps a lot. On a CPU, the small model transcribes roughly in real time or slower; an NVIDIA GPU with CUDA can be 5–20 times faster, especially for the larger models.
- How accurate is Whisper?
- On clear English speech, the larger models are comparable to paid services. Accuracy drops with background noise, crosstalk, heavy accents and specialist vocabulary; the larger models cope noticeably better.