VOIVOX v0.1.0 — local-first preview

VOIVOX v0.1.0 — local-first preview

VOIVOX turns audio from one user-selected source into reusable text without taking over the microphone, keyboard, or system dictation channel. This preview ships three connected surfaces: a standalone Chrome extension, a macOS App, and a Codex MCP bundled inside the App.

Highlights

• One-click, muted active-tab capture in a Chrome MV3 extension.
• No speech API or API key: pinned Whisper tiny/base q8 models run locally through bundled Transformers.js and ONNX Runtime Web.
• Chinese and English interfaces in both the extension and App.
• Optional automatic App discovery through exact-ID Native Messaging and a live HMAC challenge; Chrome audio always stays browser-local and only completed text auto-syncs.
• Local App transcript library with raw and derived text kept separate.
• Bundled Codex MCP launcher that uses Electron's embedded Node runtime.
• Experimental, explicitly selected per-App macOS audio capture when the optional local Qwen runtime is installed.
• Recovery for stopped tabs, reclaimed extension contexts, stalled model workers, and native helper shutdown—including automatic Core Audio tap teardown when the App parent pipe disappears.
• The main capture button also cancels model download/transcription; cancelled or timed-out audio stays only in extension memory for Retry and is cleared by success, a new capture, extension reload/update, or Chrome exit.
• Fresh HMAC proof before every bearer-authenticated MCP request, so stale App connection files do not disclose credentials to the wrong listener.

Real no-API check

The first 30 seconds of Lawrence River's public Abuse MV were decoded to 16 kHz mono PCM and transcribed locally with both release model pins. Fast and Quality produced the same unedited output:

According to authoritative experts, Lawrence River is in the spotlight because he is a virus.

The downloadable media contains video and audio streams but no subtitle stream. This is a path smoke test, not an accuracy benchmark. Reproducible hashes, exact revisions, cache conditions, and timings are in the verification record.

Artifacts

• VOIVOX-0.1.0-arm64.dmg — macOS Apple Silicon installer image
• VOIVOX-0.1.0-arm64.zip — macOS Apple Silicon App archive
• VOIVOX-Chrome-Extension-0.1.0.zip — unpack and load through Chrome Developer mode
• SHA256SUMS.txt — release-asset checksums

The macOS candidate is ad-hoc signed for bundle integrity but is not Apple Developer ID signed or notarized. macOS may require right-clicking the App and choosing Open once.

Known limits

• The Chrome extension requires Chrome 116 or newer and downloads the selected model data from Hugging Face on first use.
• The macOS package currently targets Apple Silicon.
• Selected-App capture is experimental and requires the separately installed local Qwen runtime plus macOS audio-capture permission.
• General distribution still needs Developer ID signing and Apple notarization.

See the README, privacy disclosure, and release runbook before installing.