Transcription that stays on your laptop and off your cloud bill

What it does Point it at a YouTube link, a TikTok, an X video, or a local file, and it writes a UTF-8 .txt transcript next to your shell. The whole pipeline is local: yt-dlp fetches, ffmpeg decodes to 16 kHz mono, and a streaming Zipformer transducer (Kroko, INT8 ONNX, ~80 MB) does the recognition on CPU. It claims several-times-realtime speed on a laptop, and it’s fully offline after the first model download.

The interesting bit Language detection is a neat bit of thrift: a whisper-tiny language-ID head pruned to ten languages listens to two 5-second samples cut at ~15% and ~55% into the file, and auto-fetches the matching Kroko model. The autotuning is also unusually thoughtful — it counts physical cores, not hyperthreads, because oversubscribing actually slows ONNX down.

Key highlights

• Ten languages (German, English, Spanish, French, Italian, Hebrew, Dutch, Portuguese, Swedish, Turkish), auto-detected by default; pin with --lang fr.
• --timestamps emits [MM:SS] per sentence, correctly scaled even when audio is pre-sped-up 1.4x (the default, pitch-preserved).
• --diarize adds SPEAKER_00-style labels via a second ONNX pass — still no PyTorch, no cloud SDKs.
• Runtime deps are just sherpa-onnx, numpy, and yt-dlp (plus ffmpeg on PATH).
• Works on macOS, Linux, and Windows; output lands in ./transcripts/ by default.

Caveats

• Speaker labels are per-run only — SPEAKER_00 in one file has no relation to SPEAKER_00 in another.
• The default 1.4x speedup trades some accuracy on hard audio; --speed 1.0 opts out.
• yt-dlp is deliberately unpinned, so URL fetching depends on keeping it current.

Verdict Great fit if you want quick, private, dependency-light transcripts of talks, podcasts, or meetings without provisioning anything. Skip it if you need languages beyond the ten supported or speaker identity that persists across files.