Quick Answer
Whisper is free, MIT licensed for both code and model weights, and transcribes unlimited audio on your own machine. You do not need a terminal: Buzz gives you a desktop app with drag-and-drop, speaker identification and subtitle export. Paid services cap you at 45 to 300 minutes a month.
Whisper is free, MIT licensed for both code and model weights, and transcribes unlimited audio on your own machine. You do not need a terminal to use it: Buzz gives you a proper desktop app with drag-and-drop, speaker identification and subtitle export. Paid services cap you at 45 to 300 minutes a month; local transcription has no cap at all.
Key takeaways
- Whisper is MIT licensed for code and weights, so transcripts are unrestricted commercially. Many free AI tools are not.
- Buzz (MIT) is the beginner answer: a real GUI with installers for Mac and Windows, plus speaker identification.
- openai/whisper itself does not do speaker diarization. This is the most common factual error in competing articles.
- faster-whisper at int8 transcribes 13 minutes of audio in 59 seconds using 2.9GB, versus 2m23s and 4.7GB for the original.
- Free tiers elsewhere are tight: Rev 45 min/month, Descript 60 min, Otter 300 min plus only 3 lifetime file imports.
How to actually use Whisper
Most guides send you to a terminal. You do not need one.
The easiest route: Buzz. It is MIT licensed, free, and ships installers. Download it, drag in an audio or video file, pick a model size, and it produces a transcript you can export as TXT, SRT or VTT. It also does live transcription from your microphone and speaker identification.
Two honest friction points: Buzz supports Apple Silicon Macs only (Intel support ended at v1.4.5), and its Windows installer is unsigned, so you will click through a SmartScreen warning.
For meetings specifically: meetily. It ships a Mac DMG and Windows installer, and its real advantage is capturing microphone and system audio simultaneously, so it records both sides of a call rather than just a file you already have. It adds AI summaries through Ollama locally, or Claude, Groq, OpenAI and custom endpoints.
If you are comfortable with a terminal: pip install openai-whisper, then whisper audio.mp3 --model medium. faster-whisper and whisper.cpp are faster alternatives covered below.
Which model size to choose
Whisper comes in sizes that trade speed against accuracy:
| Model | Parameters | VRAM | Relative speed |
|---|---|---|---|
| tiny | 39M | ~1GB | ~10x |
| base | 74M | ~1GB | ~7x |
| small | 244M | ~2GB | ~4x |
| medium | 769M | ~5GB | ~2x |
| large | 1550M | ~10GB | 1x |
| turbo | 809M | ~6GB | ~8x |
Turbo is the sweet spot for most people. It is an optimised version of large-v3 offering "faster transcription speed with a minimal degradation in accuracy."
One caveat worth knowing: turbo is not trained for translation. It transcribes in the source language but will not translate to English. Use large for that.
English-only variants (tiny.en, base.en, small.en) perform better than multilingual ones at the smaller sizes, though the gap narrows at medium and above.
The speed comparison that matters
If transcription feels slow, you are probably using the wrong implementation. Benchmarked on 13 minutes of audio with Whisper large-v2 on an RTX 3070 Ti:
| Implementation | Precision | Time | VRAM |
|---|---|---|---|
| openai/whisper | fp16 | 2m23s | 4,708MB |
| whisper.cpp | fp16 | 1m05s | 4,127MB |
| faster-whisper | fp16 | 1m03s | 4,525MB |
| faster-whisper | int8 | 59s | 2,926MB |
faster-whisper at int8 is more than twice as fast using 40% less memory than the original, with the same model.
For CPU-only machines, whisper.cpp is purpose-built: plain C/C++ with no dependencies, running on macOS, Windows, Linux, iOS, Android, WebAssembly and even Raspberry Pi. On Apple Silicon, its Core ML build runs the encoder on the Neural Engine for "more than x3 faster" performance than CPU alone.
Who said what: the diarization question
This is where most articles get it wrong, so be clear:
openai/whisper does not do speaker diarization. There is no mention of it anywhere in the repository. It produces a single undifferentiated transcript.
If you need to know who said what, your free options are:
- Buzz: has speaker identification, free, with a GUI. Easiest choice.
- WhisperX: best quality, BSD-2-Clause, uses pyannote's community diarization model (CC-BY-4.0, commercial use with attribution). Requires a Hugging Face token and accepting the model agreement. Claims 70x realtime with large-v2 under 8GB VRAM.
- whisper.cpp: has tinydiarize, but it is experimental.
- meetily: no. Its repository states diarization is planned for a PRO edition. The free version does not do it, contrary to several roundups.
What the paid services actually cost
Verified on each vendor's own pricing page in September 2026:
| Service | Free tier | Entry paid |
|---|---|---|
| Otter | 300 min/mo, 3 lifetime imports | $16.99/user/mo ($8.49 annual) |
| Fireflies | Unlimited transcription, 400 min storage, 20 AI credits | $18/mo ($10 annual) |
| Rev | 45 AI min/mo, English only | $25.49/seat annual; human $1.99/min |
| Descript | 60 min/mo, watermark, 720p | $16 annual / $24 monthly |
Otter's free tier has a detail people miss: the 3 file imports are lifetime, not monthly. After three uploads you cannot import again without paying.
Fireflies is the most generous free tier here, with unlimited transcription, though storage and AI summaries are capped.
Rev's human transcription at $1.99 per minute is a different product entirely: 99%+ accuracy delivered within 12 hours, which no AI tool matches for difficult audio.
Local versus cloud: the real decision
Cost. Local is free and unlimited forever. Paid services cost $16 to $25 a month and cap your minutes. If you transcribe more than a few hours monthly, local wins outright.
Privacy. This is the strongest argument and it is not close. Whisper, whisper.cpp, faster-whisper, Buzz and meetily all run entirely on your machine. Audio never leaves it.
For client calls, HR conversations, medical or legal discussions, and anything under NDA, uploading to a cloud service is a compliance question. Running locally removes it entirely. meetily positions explicitly around this, using Ollama for fully local summarisation with zero external calls.
Convenience. Cloud services join your meetings automatically, sync across devices, and require no setup. That is genuinely worth paying for if your transcription needs are light and non-confidential.
Accuracy. Whisper large-v3 is competitive with commercial services on clean audio. On difficult audio (heavy accents, crosstalk, poor microphones) human transcription still wins, which is what Rev sells.
Which should you choose?
Choose Buzz if you want free unlimited local transcription with a proper interface and speaker identification. For most people this is the answer.
Choose meetily if your use case is meetings specifically and you want both sides of a call captured with AI summaries.
Choose faster-whisper if you are comfortable with Python and want the fastest local option.
Choose whisper.cpp if you are on CPU-only hardware or something unusual like a Raspberry Pi.
Choose Fireflies free if you want cloud convenience at no cost and your meetings are not confidential.
Choose Rev human transcription if accuracy on difficult audio matters more than cost.
For a direct comparison against the paid services, see the best AI meeting note takers. For meeting-specific tools, read our roundup of the best AI meeting note takers.
The verdict
Free transcription is one of the few areas where the open-source option is not a compromise. Whisper is MIT licensed for both code and weights, runs unlimited on your own hardware, and is accurate enough for professional use.
The only genuine tradeoffs are setup effort and cloud convenience. Buzz removes most of the first, and if your audio is at all confidential, the second is a reason to stay local rather than a reason to pay.
If speed matters, our benchmark of faster-whisper versus whisper.cpp and Buzz shows a 2.4x difference between implementations.
Related reading: best AI meeting note takers and how to clone your voice with AI.
Frequently Asked Questions
Is Whisper free for commercial use?
Yes. Whisper is MIT licensed for both the code and the model weights, so transcripts are unrestricted commercially. That is unusual: many free AI tools release permissive code with non-commercial model weights.
How do you use Whisper without a terminal?
Use Buzz, a free MIT-licensed desktop app with installers for Apple Silicon Macs and Windows. Drag in a file, pick a model size, and export TXT, SRT or VTT. It also does live transcription and speaker identification. For meetings, meetily ships a DMG and Windows installer.
Does Whisper identify different speakers?
openai/whisper does not do speaker diarization at all, contrary to many articles. Buzz has speaker identification, WhisperX offers the best quality using pyannote, and whisper.cpp has experimental tinydiarize. meetily's free version does not, as diarization is marked for its PRO edition.
Which Whisper model should you use?
Turbo for most people. It is an optimised large-v3 offering much faster transcription with minimal accuracy loss, needing about 6GB VRAM. Note turbo is not trained for translation, so use large if you need to translate to English.
Is local transcription better than Otter or Rev?
For cost and privacy, yes. Local is unlimited and free, and audio never leaves your machine, which matters for client calls and confidential meetings. Paid services cap free tiers at 45 to 300 minutes and charge $16 to $25 monthly. Rev's human transcription at $1.99 per minute still wins on difficult audio.