v2026.7.4

NVIDIA Parakeet-TDT English speech-to-text backend (ONNX via `onnx-asr`, no NeMo/torch).

AI agents and tools can read these release notes as plain Markdown.

This release covers 0 pull requests.

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.

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.

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.

CHANGELOG/records/2026.7.4.md lists every pull request in this release once, with its related issues and thanks.