Does anything leave my computer?

The model does the opposite: it is downloaded to you, once. whisper.cpp reads a model file you saved yourself, OpenAI's whisper keeps its models in ~/.cache/whisper, and the browser tool fetches its model from Hugging Face and keeps it in the browser's cache. The recording itself is read on your machine by all three and is never sent anywhere.

Which model should I use?

Start with base.en for English: it is 148 MB and got every word of our test clip right. For other languages, use a model without .en in its name and say which language it is. Go up to small or larger only when the transcript has too many mistakes, because each step up is slower and bigger.

Can I get subtitles (SRT or VTT) out of it?

From the command line, yes: whisper-cli -osrt or -ovtt, or whisper --output_format srt. The browser tool saves plain text only. With Timestamps on, each line starts with the time it begins, which is enough to find a quote but not a subtitle file.

Why does whisper-cli say it failed to read my file?

It reads WAV, MP3, FLAC and OGG, not M4A, which is what Voice Memos and most phones record. Convert to a 16 kHz WAV first (afconvert on a Mac, ffmpeg anywhere). OpenAI's whisper and the browser tool decode M4A themselves.

Can it tell who is speaking?

No. Whisper writes down what was said, not who said it, in all three routes here. Telling speakers apart is a separate job (diarisation) that none of them do.

How to transcribe a recording without uploading it

To transcribe a recording without sending it to a service, run OpenAI's Whisper speech model on your own machine. It is free, it is good, and there are three practical ways to run it: whisper.cpp in a terminal, OpenAI's own Python command, or a web page that runs it inside the browser.

The Audio Transcriber with meeting-note.m4a open as a finished transcript: four timed phrases about a draft budget due on Friday, with Copy and TXT buttons, and the Add audio, Settings and Transcribe bar at the bottom — Drop the recording, transcribe it, read it as a document. Nothing is uploaded.

The direct waywhisper.cppmacOS · Linux · Windows

On a Mac with Homebrew, install it and fetch an English model once:

brew install whisper-cpp
curl -LO https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin

Then transcribe a WAV and write subtitles next to it:

whisper-cli -m ggml-base.en.bin -f interview.wav -osrt

It prints each line with its time and saves interview.wav.srt. A phone recording (.m4a) has to become a WAV first; see below.

What the commands actually produce

The test clip was 12.8 seconds of English speech made with macOS say, on an M4 Pro Mac (macOS 26.5). With whisper.cpp 1.9.5 and ggml-base.en.bin:

[00:00:00.000 --> 00:00:01.680]   Thanks for joining the call.
[00:00:01.680 --> 00:00:04.000]   Before we start, a quick reminder,
[00:00:04.000 --> 00:00:06.480]   the draft budget is due on Friday.
[00:00:06.480 --> 00:00:08.660]   Maria will send the figures to the whole team
[00:00:08.660 --> 00:00:10.440]   by Wednesday afternoon.
[00:00:10.440 --> 00:00:12.960]   If anything is missing, tell me before then.

Every word is right. The first run took six seconds, most of it setting up the graphics chip; after that the same clip took a third of a second. The .srt it wrote starts like this:

1
00:00:00,000 --> 00:00:01,680
 Thanks for joining the call.

2
00:00:01,680 --> 00:00:04,000
 Before we start, a quick reminder,

The other switches you are likely to want:

-otxt plain text, -ovtt WebVTT for web video. Combine them freely.
-of meeting names the output meeting.srt instead of interview.wav.srt.
-np prints only the transcript, without the loading chatter.
-l sets the language, and it defaults to English. With a multilingual model (one without .en, such as ggml-base.bin), pass -l nl for Dutch, or -l auto to let it detect.

Homebrew renamed it

brew install whisper-cpp still works, but the formula is now called whisper.cpp, and the command it installs is whisper-cli. Older guides that say ./main or whisper-cpp predate that.

Turn a phone recording into a WAV

whisper-cli reads WAV, MP3, FLAC and OGG. Given an .m4a, which is what Voice Memos and most phones save, it stops with failed to read audio file. Whisper wants 16 kHz mono anyway, so convert to exactly that.

A Mac already has the converter, afconvert:

afconvert -f WAVE -d LEI16@16000 -c 1 memo.m4a memo.wav

Anywhere with ffmpeg:

ffmpeg -i memo.m4a -ar 16000 -ac 1 -c:a pcm_s16le memo.wav

Both gave a WAV that transcribed to the same six lines as above.

OpenAI's whisper command

The original, in Python. It decodes any format ffmpeg can (so .m4a is fine), and it needs ffmpeg installed. It also pulls in PyTorch, which is a large download.

pip install openai-whisper
whisper memo.m4a --model base.en --output_format srt

Always name the model

Without --model, it uses turbo, which is a 1.6 GB download and slow without a graphics card. Without --output_format, it writes five files (txt, vtt, srt, tsv and json).

On the same clip (openai-whisper 20250625, PyTorch 2.14.1 on the processor, in a Linux container), it downloaded the base.en model to ~/.cache/whisper, warned that FP16 is not supported on CPU; using FP32 instead (harmless), and wrote the same six lines with times a few hundredths of a second different. It took 34 seconds including the download. whisper.cpp is the quicker of the two, by a long way on a Mac.

Transcribe in the browser

The Audio Transcriber runs the same Whisper model in the page, with WebGPU (the graphics chip) in Chrome and Edge, and WebAssembly on the processor elsewhere. The recording is converted and transcribed on your computer; the only download is the model.

Open the Audio Transcriber and drop one or more recordings on the page, or choose Add audio. MP3, WAV, M4A, OGG, FLAC, WEBM and MP4 all work.
Choose Settings and pick the Whisper model. Tiny is the default and fine for clear speech; Base and Small are more accurate and slower.
Set Language to the language spoken, or to Auto-detect. It starts on English. Detection can guess wrong on a short or noisy clip.
Timestamps are shown by default. Hide timestamps, in the same panel, takes them out of the page and out of what you copy or save.
Press Transcribe. The first time, the model downloads; after that the model picker says on this device and it starts straight away.
The transcript opens under its file. Press Copy, or TXT to save memo.txt. With several files, All TXT saves them in one text file.

What it gives you is plain text. With timestamps shown, each line starts with its start time, as [00:04], one per phrase rather than per word. That is enough to find a quote in the recording. It is not a subtitle file: for SRT or VTT, use one of the commands above.

What it downloads

The sizes in the model picker are smaller than what is actually fetched. The model is two files, and which versions load depends on the browser:

Model	Listed	Fetched with WebAssembly	Fetched with WebGPU
Tiny	~44 MB	about 64 MB	about 152 MB
Base	~78 MB	about 136 MB	about 291 MB
Small	~250 MB	about 510 MB	about 968 MB

These are the sizes of the files on Hugging Face that the tool asks for. It is a one-off: the model stays in the browser's cache, and a later visit only checks it is there. Clearing the site's data removes it.

Long recordings

The whole recording is decoded into memory before Whisper starts: an hour of audio becomes about 230 MB of samples, on top of the model. A long interview works on a desktop. On a phone, or for anything over an hour or two, split the file first, or use whisper.cpp.

Which to use

Subtitles, or many recordings: whisper.cpp. It is fast, writes SRT and VTT, and a shell loop does a whole folder.
You already use Python, or need M4A without converting: OpenAI's whisper, with --model set.
One recording, no installing, any device: the Audio Transcriber. Copy the text or save it as TXT.

Whichever you pick, read the transcript against the recording before you quote from it. Whisper is good, but names, numbers and crosstalk are where it slips, and it never marks a guess as a guess.