Word Tracking works. A streaming Zipformer transducer runs through
sherpa-onnx on the CPU, and the model is fetched on first use rather than
shipped, so the installer stays at 6.5 MB and a Classic-only user never
downloads 42 MB they will not use.
sherpa-rs wraps only sherpa-onnx's offline recogniser, which decodes a
finished buffer. A teleprompter cannot wait for the presenter to stop
talking, so speech.rs calls the online C API directly. That means a
hand-written config struct against the C headers with the unused members
zeroed, so there is an ignored test that loads a real model and decodes
through it — a wrong field would otherwise surface in front of a presenter
mid-take rather than in CI.
Endpoints rebase the matcher's transcript window instead of restarting it.
The reading position survives, and the stream stops growing without bound.
Audio capture moves from the webview to Rust. Recognition needs raw PCM, and
opening the same input device twice — WebAudio for metering, the recogniser
for transcription — fails on exclusive-mode hardware. One capture path now
feeds both, which also puts prompt-core's RMS metering to use. The callback
downmixes and hands off through a bounded queue; when the decoder falls
behind, dropping the oldest audio beats growing a backlog the transcript
would never catch up with.
build.rs stages the sherpa DLLs under a fixed path for the bundler. Naming
target/release directly in tauri.conf.json made debug builds try to bundle
release artefacts a fresh clone has never produced.
Building now needs libclang, because sherpa-rs-sys generates its bindings
with bindgen. README documents it.
Tauri 2 app around prompt-core. Classic and Voice-Activated run end to end;
Word Tracking is wired but idle until the on-device recogniser lands, so the
UI marks it unavailable rather than offering a mode that does nothing.
The overlay is placed on the top edge, not the taskbar. The macOS notch
overlay works because it keeps the presenter's eyes near the lens; on Windows
the camera is above the monitor and the taskbar is pinned to the bottom, so a
bottom-docked prompter would have them reading off their own lap. The taskbar
placement carries transport controls only, as a plain always-on-top window
rather than a registered AppBar — an AppBar reserves desktop work area and
leaves it wrong if the process dies.
Three extended style bits and one affinity call do what no cross-platform
window API exposes: WS_EX_NOACTIVATE so showing the prompter never pulls focus
off the app being presented, WS_EX_TOOLWINDOW to stay out of Alt-Tab,
WS_EX_TRANSPARENT for click-through, and WDA_EXCLUDEFROMCAPTURE to stay out of
screen shares. The last one needs Windows 10 2004+, so its result is reported
to the UI instead of being swallowed — the app must not promise privacy it
cannot deliver.
Audio metering rides along with the frame tick rather than using its own
command. Two round trips per animation frame is real cost for something that
runs beside OBS, and a split also let the speech gate be read a frame before
the sample that would have opened it.
HWNDs cross into window_effects as isize so this crate is not pinned to
whichever `windows` version Tauri happens to link.
Ground-up Rust implementation of the algorithms behind Textream's three
guidance modes. Nothing here touches the OS, so it is testable without a
microphone and a future Linux build needs only a new shell.
Offsets are grapheme clusters rather than chars, matching how the macOS app
counts Characters. A char-based index would drift on emoji cues and combining
marks and pull the highlight off the word the reader is looking at.
- text: CJK-aware tokenisation, RTL inference that ignores [cue] contents
- alignment: cue ranges, skip-ahead, and the commit policy that rejects a
large unconfirmed jump from a single dead estimate
- script: parsed script with word table and progress/offset conversions
- matcher: character- and word-level passes reconciled per update
- vad: RMS metering plus a two-threshold gate with hangover
- scroll: constant-pace driver shared by Classic and Voice-Activated
75 unit tests and a doctest cover cue skipping, filler words, misrecognition,
CJK, RTL, and the degenerate empty-script and non-finite-sample paths.