leve — labaslabas.dev

About

leve turns long-form audio and video into short-form, captioned clips — all from the command line. Point it at a podcast episode, talk, or YouTube URL; leve transcribes the audio, aligns word-level timestamps, selects engaging moments, and exports captioned clips ready for Shorts, Reels, or TikTok. It is built for podcasters and YouTubers turning long-form content into shorts, and it runs entirely on your machine: CLI-only, no SaaS, no uploads, no account.

Features

leve init — scaffold a workspace from a local file or YouTube URL
leve run — run the pipeline: download → transcribe → align → select → export
leve status — see where you are in the pipeline
leve condense — remove silent pauses from the source video
leve export clip — render a selected clip (or all clips) with captions
leve export range — create a clip from explicit timestamps and export it
leve export source — export the whole source with caption overlays
leve doctor — check system dependencies
leve reset — reset pipeline stage(s) to pending
leve clip — manage clips (add, list, select, unselect)

Requirements

macOS on Apple Silicon
ffmpeg
yt-dlp (only for YouTube URLs)
python3 + uv
Xcode Command Line Tools (for the aligner's C++ dependency)
A downloaded ggml-*.bin whisper model
A downloaded mms_fa_uint8.onnx alignment model

leve does not bundle, download, or host models — the whisper and MMS-FA models are the two most likely source of first-run confusion, so grab them before you start (see "Bring your own models" below).

Download

leve-v0.1.0-darwin-arm64.tar.gz

macOS arm64 · ~10 MB · SHA256 in SHA256SUMS

Install

Option 1 — one-liner (recommended):

curl -sSf https://raw.githubusercontent.com/lucastrunce/leve/v0.1.0/install.sh | sh

Option 2 — from source (requires a checkout for share/scripts or a LEVE_SCRIPTS_DIR override):

go install github.com/lucastrunce/leve/cmd/[email protected]

Option 3 — manual download:

curl -fsSL -O https://pub-a7b67eccf154442b85bd31333515dab4.r2.dev/leve-v0.1.0-darwin-arm64.tar.gz
curl -fsSL -O https://pub-a7b67eccf154442b85bd31333515dab4.r2.dev/SHA256SUMS
shasum -a 256 -c SHA256SUMS
tar -xzf leve-v0.1.0-darwin-arm64.tar.gz
sudo mv leve /usr/local/bin/
sudo mkdir -p /usr/local/share/scripts
sudo cp -R share/scripts/. /usr/local/share/scripts/

Bring your own models

leve does not bundle, download, or host models — you download them once and point leve at them.

Whisper model — download any ggml-*.bin model (recommended: ggml-large-v3.bin) from ggerganov/whisper.cpp, then either set [paths] whisper_model in ~/.leve/etc/config.toml, set LEVE_WHISPER_MODEL, or drop the file at ~/.leve/models/.

MMS-FA alignment model — download mms_fa_uint8.onnx (~303 MB, CC-BY-NC-4.0) from xycld/lyric-align-mms-fa:

curl -L https://huggingface.co/xycld/lyric-align-mms-fa/resolve/main/mms_fa_uint8.onnx -o ~/.leve/models/mms_fa_uint8.onnx

Then either set [paths] mms_fa_model / LEVE_MMS_FA_MODEL, or leave it unset and let align.py use the HuggingFace cache.