Bring your own streaming server. Vulkan wheels. Continuous mode that ends itself.
Self-hosted streaming
-
Run your own realtime server. Point
customat it; no API key needed.hyprwhspr setup→ Realtime WS now lists NeMo-Speech.cpp, Phonon, and any OpenAI Realtime-compatible server. Anything else is four settings:websocket_protocol,websocket_sample_rate,websocket_session_format,websocket_live_text. (#262, thanks @JacobBanghart; #270, thanks @James-Wolfley) -
NeMo-Speech.cpp. Local, English, streaming Nemotron. See Self-hosted streaming for the serve command and config.
-
Phonon-2. Live OSD preview while you speak; one stream per recording (toggle, push-to-talk, auto).
Live typing (experimental)
-
Words land as you speak.
realtime_live_typing: truewith a model whose words never change once sent (today, NeMo). Enter, the trailing space and the clipboard restore wait for stop; cancel keeps what's typed. Setup offers it when you pick NeMo. (#261, thanks @JacobBanghart) -
Off with a
post_transcription_hook,record capture, continuous or long-form.hyprwhspr config validatesays why, and the service logs it once at start.
Vulkan
-
Prebuilt Vulkan wheels. Setup fetches one on glibc 2.41+ (Arch, Fedora 42+, Debian 13); no local build. Wheels carry a build number and must match the release's checksums. (#275, thanks @sqonde)
-
Honest fallbacks. If a GPU build fails you get CPU, and the service now says so at every start. Setup installs and checks the SPIR-V build deps.
Continuous mode
-
silence_timeoutends a continuous session. Pauses still paste; a longer quiet closes the session. (#273) -
Sturdier silence stops. Pending chunks retry instead of vanishing, silence-only chunks aren't transcribed, and a replaced recording can't be stopped or pasted into by an old monitor. (#274, thanks @Anirban-Majumder)
Fixes
-
Cohere no longer runs out of VRAM on resume. The old model is released before the new one loads. Every local backend now does the same. (#272, thanks @Anirban-Majumder)
-
Missing Python modules are flagged at start, as is a venv that drifted from what setup installed. Fix: re-run setup and reinstall the backend. (#275, thanks @sqonde)
-
"Transcribing…" stays until the text lands. (#275, thanks @sqonde)
-
Custom realtime endpoints work without a key. They used to fail at start.
Upgrade notes
-
The realtime OSD preview now keeps finished segments on screen instead of clearing them.
-
A malformed
websocket_url(notws://orwss://, or with no host) now fails at start with a clear error. -
Vulkan users on glibc 2.41+: re-run setup and reinstall the backend to swap a local build for the prebuilt wheel.
PRs
-
fix(cohere): release model before reinitializing by @Anirban-Majumder in #272
-
Prebuilt Vulkan wheels, portable wheel builds, startup dependency warnings by @sqonde in #275
-
fix(continuous): isolate silence stops and retry pending chunks by @Anirban-Majumder in #274
New Contributors
- @Anirban-Majumder made their first contribution in #272
Full Changelog: v1.46.1...v1.47.0