Checking what the model repository publishes…
Uzbek speech, transcribed on your own device.
Auris‑1 is a 315M‑parameter Wav2Vec2ForCTC model that writes Uzbek in the
Latin alphabet. Recording, resampling and recognition all happen in this tab — no audio is
uploaded anywhere.
One‑time download, kept in this browser’s cache.
What gets downloaded, and where it runs
- Model weights
- —
- ONNX runtime
- —
- Runs on
- —
This happens once. The first visit is a real download of a few hundred megabytes and can take several minutes on a slow connection; the bar counts actual bytes, not an animation. Keep this tab open while it runs. Afterwards the files live in this browser’s cache and the page starts offline.
Downloading the model
This happens once — the files are cached in this browser afterwards. Keep the tab open.
Audio input
Record
Upload
Examples
Held‑out recordings from the project’s own evaluation set. Not training data.
Selected clip
—
—
Load the model above before you can transcribe.
Transcript
Transcripts appear here.
Load the model first, then record something or pick an example clip.
Transcript
Decoded greedily, without the repository’s KenLM language model — too large for a browser. Read this as a lower bound on the model’s accuracy; the server‑side numbers on the model card are better.