## 2026.7.4

### Highlights

- NVIDIA Parakeet-TDT English speech-to-text backend (ONNX via `onnx-asr`, no NeMo/torch).
- A Private-mode language toggle (English / Multilingual) on the home: English uses the on-device Parakeet engine; Multilingual falls back to the hardware-aware Whisper tier.
- Automatic capture gain control and optional noise suppression on the input path, so hot microphones no longer clip into garbage transcriptions and no manual level tuning is required across different mics and environments.
- The canonical Linux install is now per-user (`install.sh --user`, under `~/.local`): one venv process both dictates and serves the UI, and in-app updates apply without `sudo`/`pkexec`.
- The desktop shell is now built through the Tauri CLI (not a raw `cargo build`), which embeds the frontend.
- The webview reconnects its live event stream after the engine restarts (e.g. an in-app update), re-resolving the new port/token and re-syncing state, so the dictation history and quick-copy no longer freeze at their last snapshot.

### Changes

- NVIDIA Parakeet-TDT English speech-to-text backend (ONNX via `onnx-asr`, no NeMo/torch). On CPU it is both faster and more accurate than Whisper `small.en` (~6% WER, ~13x real-time), emits punctuation and casing directly, and — being a transducer — outputs silence instead of hallucinating filler. It is now the default on-device engine for English on CPU.
- A Private-mode language toggle (English / Multilingual) on the home: English uses the on-device Parakeet engine; Multilingual falls back to the hardware-aware Whisper tier. No provider or brand names surface in the UI.
- Automatic capture gain control and optional noise suppression on the input path, so hot microphones no longer clip into garbage transcriptions and no manual level tuning is required across different mics and environments.
- Parakeet decodes the whole utterance in one pass rather than through streaming chunks: it has no prompt input to carry context across chunk seams, so full-utterance decoding is both higher quality and, at ~13x real-time, still low-latency.
- The canonical Linux install is now per-user (`install.sh --user`, under `~/.local`): one venv process both dictates and serves the UI, and in-app updates apply without `sudo`/`pkexec`.

### Fixes

- The desktop shell is now built through the Tauri CLI (not a raw `cargo build`), which embeds the frontend. A plain cargo build omitted the `custom-protocol` feature and the shell tried to load the dev server, showing "Could not connect to localhost: Connection refused" at runtime.
- The webview reconnects its live event stream after the engine restarts (e.g. an in-app update), re-resolving the new port/token and re-syncing state, so the dictation history and quick-copy no longer freeze at their last snapshot.

### Complete contribution record

This audited record covers the complete df1117c834976c82b64ad8dbfd5eea09e2b43d25..9c090267ff360a3c9ae71b834be68bf3f703ade6 history: 0 in-range PRs + 0 retained seed-only PRs = 0 unique PRs. The generation manifest also supplies direct commits as editorial input; the grouped notes above prioritize user impact.

#### Pull requests