kesha-voice-kit

Local multilingual voice toolkit — speech-to-text (STT), text-to-speech (TTS), and language detection. Runs entirely offline on Apple Silicon, Linux, and Windows. No API keys, no cloud. NVIDIA Parakeet TDT for STT across 25 European languages, Kokoro-82M + Vosk-TTS for TTS, plus macOS AVSpeechSynthesizer for ~180 system voices with zero install.

Use when a model SHA-256 mismatch surfaces (download_verified errors, manifest_tests failures, fresh download produces a different hash than rust/src/models.rs pins). Walks through the safe bump procedure — verify the upstream weights deliberately, then update the pin. Refuses to suggest commenting out verification.

Cuts a kesha-engine release per CLAUDE.md rules — pre-flight audits, version bump, tag, release-notes BEFORE publish, smoke-test AFTER publish, then npm publish. Refuses to auto-run; user must explicitly invoke. Knows the gh-cli release-notes trap and the draft-URL 404 trap.