Live Demo

MiniASR

A 44.7M-parameter speech recognition model, trained from scratch on LibriSpeech. Built end-to-end with PyTorch.

44.7M

Parameters

17.9%

WER (dev-clean)

460h

Training Audio

12L

Transformer Layers

Record from microphone

00:00

Tap to record

Speak clearly · auto-stops at 15s

📁

Click to browse or drag a file here

WAV · FLAC · MP3 · OGG · WebM · mono · ≤ 12 s recommended

Record audio or upload a file to see transcription