Try it without installing it¶
The honest objection to any offline speech-to-text tool is "but is it actually accurate?" — and no amount of prose answers that. So here are four ways to find out that cost you nothing and leave nothing behind.
What you can and cannot try this way
These trials run the file → text side of YazSes (yazses transcribe). They use the same local Whisper engine as live dictation, so the accuracy you see is the accuracy you get.
They cannot do hold-to-talk dictation. That needs a real microphone, the kernel input device for the hotkey, and the ability to type into whatever window you have focused — three things a container does not have. For dictation, install it properly.
1. Docker — one command, nothing installed¶
git clone https://github.com/MSKazemi/yazses.git && cd yazses
docker build -f packaging/docker/Dockerfile -t yazses .
docker run --rm -v yazses-models:/models \
-v "$PWD/data/librispeech-sample:/data:ro" -v /tmp:/out \
yazses jfk.wav -o /out/jfk-heard.txt
That transcribes a clip that ships with the repo. Now compare the two:
cat /tmp/jfk-heard.txt # what YazSes heard
cat data/librispeech-sample/jfk.txt # the reference transcript
They should match word for word. They will not match byte for byte, and that is correct: the reference is the LibriSpeech transcript, which carries no punctuation, while YazSes punctuates what it hears — so you get Americans, ask not against Americans ask not. Compare the words, not the commas.
Why the sample is mounted read-only
YazSes writes its transcript as a sidecar next to the input, so jfk.wav would produce jfk.txt — which is the name of the reference file. Mounting the clip directory writable would overwrite the known-correct text with the model's own output, and the comparison above would then always look perfect no matter what the model actually said. The :ro and the separate -o path make that impossible rather than merely unlikely.
What actually happened when this page was written (4-core CPU, base.en):
| Run | Wall time | Why |
|---|---|---|
| First | 43 s | includes downloading the 141 MB base.en model |
| Every run after | 2.3 s | model is cached in the yazses-models volume |
The clip is 11 seconds of audio, so the cached run is roughly 5× faster than real time on a CPU, with no GPU.
Transcribe your own file by pointing the volume at your own folder:
docker run --rm -v yazses-models:/models -v "$PWD:/data" yazses meeting.m4a
docker run --rm -v yazses-models:/models -v "$PWD:/data" yazses talk.mp4 -f srt
docker run --rm -v yazses-models:/models -v "$PWD:/data" yazses talk.wav --model small.en
wav, mp3, m4a, ogg, flac, opus and mp4 all work. Output formats are txt, md, srt, vtt and json.
The container has no network access to anything but the model download
After the first run the model is on disk and you can prove the point: add --network none and it still transcribes. That is the whole pitch, testable in one flag.
The image is 833 MB. It deliberately drops PySide6 (the desktop overlay, ~648 MB), which transcription never uses.
2. In your browser — one click, nothing to install¶
Upload a file or record straight from the page. It runs the same engine and the same default model (base.en, int8) that YazSes uses on your machine, so the accuracy you see is the accuracy you get.
This one page is not private — and that is the point of saying so
The hosted demo runs on Hugging Face's servers, so audio you send to it does leave your computer. That is true of that page only, and it exists purely to let you judge accuracy before installing.
Everything else on this site is the opposite: YazSes runs entirely on your own CPU. If privacy is your reason for being here, use the Docker method above — it proves the claim with --network none — rather than trusting either page.
3. In your browser — GitHub Codespaces¶
No Docker, no local install, nothing on your machine at all:
- Open the repository
- Code → Codespaces → Create codespace on main
- Wait for it to finish setting up — it prints these instructions itself — then run:
uv run yazses transcribe data/librispeech-sample/jfk.wav -o /tmp/jfk-heard.txt
cat /tmp/jfk-heard.txt # what YazSes heard
cat data/librispeech-sample/jfk.txt # what it should say
The -o matters: without it the transcript is written beside the input as jfk.txt, which would overwrite the reference file you are comparing against.
The container definition is in .devcontainer/. GitHub's free tier covers this comfortably.
The same container is the project's contributor environment, so if you decide you want to change something, the test suite is already installed and runs fully offline in about 30 seconds (uv run python -m pytest tests/ -q) — no microphone, model or GPU needed.
4. Just read the numbers¶
If you would rather not run anything:
- Benchmarks — word error rate and decode speed per model, measured over 200 utterances from 40 speakers.
- The 40-second demo — the real loop, end to end.
- Comparison — where YazSes wins and where it honestly does not.
Ready to install?¶
Dictation is the part worth having, and it needs a real install:
Want to know the cost before you commit? What installing actually costs lists the disk, time and downloads, and how to uninstall is a single page — written before you install, on purpose.