goodroot/hyprwhspr v1.46.1 on GitHub

Lighter on memory, quicker cues. Some small bugs.

Parakeet.cpp keeps sentences spoken after long pauses. Pauses over 1 s are shortened before decoding; Parakeet had been silently dropping whole sentences that followed them. (#268, thanks @EnacheB)
pywhispercpp no longer downloads a model you didn't ask for. With only a .en model installed, it's now loaded directly for English dictation instead of fetching the multilingual model. Other languages still get the multilingual download.
Tray health tells the truth. It never actually recovered anything; it now matches status. For a stuck recording: hyprwhspr record cancel.

Qwen3-ASR uses ~2.6 GB less VRAM. The sidecar now reserves an 8192-token context (llama.cpp's default was 32000) and skips llama-server's up-to-8 GiB RAM prompt cache, which one-shot dictation never reuses. Tune with qwen3_asr_ctx_size; null restores llama.cpp's default. (#269, thanks @doitian)
faster-whisper on CPU loads int8 by default: ~3× less RAM and faster. It also honors threads now. An explicit faster_whisper_compute_type is respected.
Waybar tray: each status tick launches ~9 processes, down from ~60.
Fewer idle wakeups: the realtime receiver no longer polls, and the mic OSD stops checking for theme changes while hidden.

Start and stop cues play through pw-play or paplay when available. They start faster than ffplay, so the "talk now" cue lands sooner, and they follow your volume settings. Custom non-OGG/WAV/FLAC sounds still use ffplay.

Proxy dictation sessions. New realtime_transcription_session_type (default transcription); set it to realtime for proxies such as CLIProxyAPI that reject transcription-only sessions. (#267, thanks @rexlManu)

faster-whisper CPU users on auto move from float32 to int8. To keep float32, set "faster_whisper_compute_type": "float32".
Qwen3-ASR requests are capped at 2 minutes of audio each, well within the new 8192-token context.

PRs

Full Changelog: v1.46.0...v1.46.1