Skip to content

hayamimi 早耳 v0.1.0

Choose a tag to compare

View all tags
@oboroge0 oboroge0 released this 25 Aug 14:57
· 127 commits to main since this release
v0.1.0
1eb3132

hayamimi 早耳 v0.1.0 — first public release

Real-time, multilingual speech-to-text on CPU only. No GPU, no cloud, under 2GB RAM.

🎬 Demo video attached below (hayamimi_demo.mp4) — real engine output on real CC-BY audio, replayed frame-accurately.

Why it exists

Running Whisper-class models in real time normally requires a GPU or a cloud API. hayamimi routes each utterance to a small specialist model per language instead, which is how it stays real-time on an ordinary 6-core CPU.

What you get

  • Japanese accuracy that beats much larger models: 5.8% character error rate on real broadcast audio — less than half of whisper-large-v3-turbo's 13.8% on the same clips — while running at 10-50x realtime.
  • Fast feedback: a finalized line lands about 100ms after you stop talking, with draft subtitles streaming while you speak.
  • 5 specialist language routes (Japanese / Chinese / Korean+Cantonese / English+24 European languages) plus a fallback model covering roughly 1600 languages.
  • Practical extras: two-pass refinement for a clean transcript, speaker labels, ja→en/zh/ko translation, hotwords and a user dictionary, Japanese punctuation restoration, Korean word spacing.
  • Streaming-ready output: --serve starts a browser dashboard and a transparent OBS overlay.

Every accuracy number above comes from evaluation on real speech (broadcast, interviews, lectures) — not synthetic audio. The measurement scripts ship in the repository, so the results are reproducible.

Install

python -m venv .venv && .venv/Scripts/pip install -r requirements.txt
.venv/Scripts/python scripts/download_models.py   # ~3.1GB (or --minimal, ~1.1GB ja/en)
.venv/Scripts/python scripts/realtime_transcribe.py --serve

License notes

The code is MIT. Model weights are downloaded from their original publishers and carry their own licenses (THIRD_PARTY_NOTICES.md has the full table). The optional --translate en model (FuguMT) is CC BY-SA 4.0. Demo-video audio comes from FLEURS and LibriSpeech (both CC-BY 4.0); attribution is in demo/CREDITS.md.

日本語での紹介

hayamimi（早耳）は、GPUもクラウドも使わずにCPUだけでリアルタイム多言語文字起こしをするツールです。言語ごとに小さな専用モデルへ振り分ける構成で、実放送の日本語音声で文字誤り率5.8%（同じ音声でwhisper-large-v3-turboは13.8%）を、6コアCPUの実時間の10〜50倍の速度で達成しています。話し終えてから約100msで確定字幕が出て、--serve を付ければOBS用の透過オーバーレイとダッシュボードがそのまま使えます。精度の数値はすべて実音声での測定で、測定スクリプトもリポジトリに同梱しています。

🤖 Generated with Claude Code

Assets 3
Loading