NEBians
search Sign In Register
engine… device — cache — model none

Record or drop audio 16 kHz · VAD · auto-segment

00:00.0
idle — tap the dot to record

Transcript

Nothing yet. Load a model on the right, then record or drop an audio file. Everything runs locally — no audio leaves this device.
words 0audio 0s infer —speed —

1 · Model — English ultrafast

tiny.en fastest · moonshine-tiny 30MB superfast 5-15× vs Whisper · turbo most accurate English (32→4 decoder layers, ~98% large-v3). All cache offline. q8=WASM stable, q4=turbo smallest.
Models fetch once from Hugging Face CDN, then stored in Cache API — offline forever.

2 · Options · English

🇺🇸
English (en) — fixed
No language detection · fastest path
ultrafast
30s
10s30s
Long files split into overlapping chunks (stride 40%). 30s is Whisper native window.
VAD: silence ≥0.55s closes segment, ≥0.25s opens, 0.25s pre-roll, hard split at 28s.

Extras · On-device AI superfast · WASM

Summarizer 80MB Sentiment 67MB Embeddings 22MB Keywords 0MB TTS Kokoro 82M 100% on-device

Post-process your transcript locally — no server, all WASM/WebGPU. Open TTS Studio →

New: turbo 809M & moonshine 30MB in Model dropdown above — try them. TTS now has its own page with tiny/small/medium Kokoro models.
TTS Studio
All tools stream live per-chunk for long files · Sentiment 3-way, keywords TF-IDF, embeddings cosine.

History local only

Transcripts save here automatically (IndexedDB, never leaves device).
⬇ Drop audio to transcribe (wav · mp3 · flac · m4a · ogg · webm)