README.md

metadata

title: Notesoop STT Ko
emoji: 🎙️
colorFrom: yellow
colorTo: pink
sdk: gradio
sdk_version: 6.31.0
app_file: app.py
python_version: '3.12'
startup_duration_timeout: 1h
short_description: Ultra-fast Korean speech recognition (on-device STT)
license: apache-2.0

Notesoop-STT-Ko-v1 — Korean Speech Recognition

Gradio demo for Phsoft-ai/Notesoop-STT-Ko-v1, an ultra-fast Korean STT model designed for on-device use.

How it works

1. Frozen audio encoder — google/embeddinggemma-2 (Gemma-4 audio tower, bf16) produces frame embeddings, which are Matryoshka-sliced to 256 dims and L2-normalized. The encoder is never fine-tuned.
2. Conformer CTC reader — a small 10-layer Conformer (21.5 M params) with a cosine-similarity vector-bank head over a 25,000-token Korean morpheme vocabulary (blank id 0, unk id 1).
3. KenLM 3-gram rescoring — CTC top-k logits are rescored by a Korean KenLM language model (alpha=0.12, beta=0.40, beam 8) and merged back across overlapping 16 s windows.

Reported footprint is ~1.5 GB VRAM and ~40× real-time on the author's hardware; the model card reports CER 16.21 % on the Zeroth-Korean test set.

Usage

Upload a clip (or record from the microphone) and press Transcribe. Korean only. Audio is resampled to 16 kHz mono; up to 180 s per run. ffmpeg-decodable containers (wav / mp3 / m4a / flac / …) are supported.

Example clips

The four bundled clips are Korean read-speech utterances from Bingsu/zeroth-korean (CC BY 4.0), the benchmark this model reports on. They are redistributed here unchanged, with attribution, purely as demo inputs for the model.

Notes / limitations

• Korean only — no other language is supported by the vocabulary or the LM.
• Very short clips (< ~3 s) are padded to the 16 s window and may transcribe poorly; 5 s or longer works best.
• chunk_sec / overlap_sec / batch_size are exposed under Advanced options; the defaults (16 s / 1 s / 4) are the reference setting.