Skip to content

YazSes and Hugging Face

Hugging Face shows up in three different roles around YazSes. They have different privacy properties, so this page keeps them apart.

1. The installed app: Hugging Face is where the model comes from

YazSes on your computer recognises speech locally, on your own CPU. It needs no account, no subscription and no cloud inference API.

The one time it uses the network is to fetch the speech model. The default models are open faster-whisper conversions published by SYSTRAN on the Hugging Face Hub. After that single download, dictation works with the network off. Which models, and how to install one by hand on a machine that cannot reach the Hub, is on Models; every network connection YazSes can make is listed on the privacy statement.

YazSes does not train these models, and using the Hub as a download host does not make Hugging Face the maintainer of YazSes. Each model's own licence and terms apply to it.

2. The hosted demo: audio does leave your computer

There is one YazSes page on Hugging Face that runs speech recognition for you:

YazSes hosted transcription demo

This is not the private, offline app

This is a hosted demonstration. Audio submitted here is processed on Hugging Face infrastructure. For private, offline voice typing, install YazSes on your own computer.

What it is for, and what it is not:

Hosted demo (Hugging Face Space) YazSes installed on your computer
Where speech is processed Hugging Face servers Your own CPU
Does your audio leave your computer? Yes No (the model is downloaded once, then nothing leaves)
Hold-to-talk anywhere on the desktop No Yes
Types into the focused application No Yes
Purpose Judge transcription quality before installing Everyday dictation

The demo uses the same engine family and the same default model size as the desktop app (faster-whisper base.en, CPU, int8), but it is not the same pipeline, so a transcript there is an indication of accuracy, not a guarantee of what the app would type. It trims silence with the decoder's voice-activity filter, which the desktop app deliberately does not use, and it applies none of the app's text post-processing.

It accepts recordings of up to 60 seconds and 25 MB, and refuses longer or unreadable files with a message. Its code adds no logging or analytics of its own and writes no audio or transcripts anywhere. Gradio, the framework it runs on, does keep each recording or upload in a temporary cache on the server: the Space checks that cache every 10 minutes and deletes files older than 10 minutes, so a recording can stay there for up to about 20 minutes. It makes no promise about what Hugging Face itself logs or keeps; that is governed by Hugging Face's own policies.

3. Research and source

  • GitHub is the canonical home of the source code, the issue tracker and every contribution: https://github.com/MSKazemi/yazses.
  • The first paper, YazSes: An Offline, Privacy-First, Cross-Platform Hold-to-Talk Voice-Dictation System (arXiv:2607.28878), describes the system. Cite it, or the software's concept DOI, as CITATION.cff says.
  • Benchmark results behind the published numbers are archived as JSON in paper/results/, with a manifest naming the machine each one was taken on. Two numbers are comparable only if their machine matches. The method and headline results are on Benchmarks.

The rest of the Hub presence

The project's official organization is huggingface.co/yazses; it holds a short introduction. The pages below live under the maintainer's own account, huggingface.co/mskazemi:

Page What it is
Showcase Space A static one-page overview. It does not run YazSes and processes no audio
Benchmark results dataset The archived results as plain CSV tables with provenance (machine, versions, corpus), under a dataset card. It contains no audio, no transcripts and no corpus text
Research collection A curated list: the paper, the demo, the showcase, the dataset and the upstream models
Paper page The Hub's page for the arXiv preprint, linked to the dataset and the Spaces that cite it

The files behind these pages live in the repository's huggingface/ directory, which is their source of truth. The pages online today were first put up, and later moved or renamed, outside the publishing workflow; from now on, changes are published through that workflow rather than edited on the Hub. The files are checked offline by scripts/hf_preflight.py before anything can go out, and the benchmark tables are regenerated from the archive by scripts/export_hf_benchmarks.py (--check fails on drift). The publishing workflow is manual-only, uses Hugging Face Trusted Publishers with short-lived credentials rather than a stored token, and refuses any destination the maintainer has not approved in a reviewed commit.