Releases · gunn4r/earwig

Added a manual namer (--namer manual) that prompts you to type each speaker's name interactively, with no LLM or inference required. Names remain available via --namer claude / --namer local. (#10)
Added earwig update, which detects how earwig was installed (uv tool, pipx, or pip) and runs the matching upgrade; a source checkout is told to git pull. Added earwig --version (and an earwig version subcommand) to print the installed version. (#12)
Added a --device {cpu,cuda} flag to pick the compute device for transcription and diarization (default cpu; cuda uses float16 on an NVIDIA GPU). There is no mps option — the faster-whisper backend has no Apple-GPU support, so Apple Silicon transcribes on CPU. (#14)
Add earwig setup, an interactive wizard that walks you through the Hugging Face token, stores it in ~/.config/earwig/env without echoing it, records your default namer, and verifies ffmpeg, your token, both gated model licenses, and your namer with actionable errors. earwig now reads that config file and ./.env at startup, so the token file the README has always described actually works.
Add a one-line installer: curl -fsSL https://raw.githubusercontent.com/gunn4r/earwig/main/install.sh | sh installs earwig from git with uv or pipx, warns if ffmpeg is missing, and launches earwig setup.

Escape untrusted text at the Markdown render boundary. Video title, channel, source URL, speaker names, and transcript body — all attacker-influenced — are now neutralized before they're written to the output file: HTML/Markdown metacharacters are escaped, control characters (including terminal/ANSI escapes) are stripped, and the [source](...) URL is validated as http(s) and percent-encoded so a crafted URL can't break out of the link. LLM-inferred speaker names are likewise stripped of control characters before they reach the interactive confirm prompt. Closes the output-injection finding from the §4.6 security audit. (#18)

Removed the heuristic speaker namer and the auto namer policy. Regex name-inference could not be made accurate enough to be safe (a wrong name is worse than none), so naming is now opt-in and the default is off (speakers keep their SPEAKER_xx labels). A persisted EARWIG_NAMER=heuristic/auto now warns and falls back to off. (#10)

Add a CODE_OF_CONDUCT.md (Contributor Covenant 2.1, linked from CONTRIBUTING.md) and tidy the README: drop the stale "earlier versions did X" naming note, trim internal-sounding compatibility phrasing, and document the trust model for the downloaded pyannote models.
Add a project CLAUDE.md guide and note beta/pre-1.0 status plus a changelog link in the README.
Document the config and setup modules in CONTRIBUTING.md's architecture guide, and point the contributor dev-setup at earwig setup.
Unwrap hard-wrapped prose in README.md, CONTRIBUTING.md, and changelog.d/README.md so each paragraph is a single line — no wording changed, only line breaks.

Pin the tested dependency ranges: whisperx>=3.8,<3.9 and pyannote-audio>=4,<5, since whisperX's diarization API and default model drift across minor versions. yt-dlp is intentionally left uncapped so it keeps pace with YouTube changes. (#16)
Harden the CI workflow: declare a top-level least-privilege permissions: contents: read (all jobs only read the repo), and pass github.base_ref to the changelog check through an env: variable instead of interpolating it directly into the run: script. Defense-in-depth from the §4.6 security audit; no functional change. (#17)
Add scripts/release.sh to cut a release in one step (build changelog, bump version, reinstall, commit, tag), with a --dry-run preview.
Keep the local CLAUDE.md agent guide private (gitignored); public contributor conventions live in CONTRIBUTING.md.

Navigation Menu