Skip to content

Draft 0.3.1

Latest
Latest

Choose a tag to compare

View all tags
@TirupMehta TirupMehta released this 29 Oct 09:39
v0.3.1
13f9735

Draft 0.3.1 is the first release with the CPU path tuned as a first-class fallback, plus opt-in lists.

What changed

  • CPU fallback is faster and no longer blocks startup. On a machine without a usable GPU, Draft now asks the inference session for three quarters of the logical CPUs instead of all of them. Oversubscribing the decoder measured 26 to 40 percent slower on the reference machine. The GPU-only warm-up pass is skipped on CPU, so the app is ready to dictate in about 3 seconds instead of about 12. The model, precision, decode settings, and text are unchanged.
  • Live preview is honest on CPU. A preview is a full extra encode, so the switch is not offered on the CPU backend and a saved preference is turned off with a visible notice instead of doing nothing.
  • Opt-in lists. A run of three or more spoken ordinals, or a make a list / make a numbered list block, becomes one item per line. Lists add only a line marker; no word is added, dropped, or reordered.
  • draftctl gained --threads N, and DRAFT_THREADS pins the count. The measured CPU table and the thread policy are in docs/PERFORMANCE.md.

Installer

  • File: Draft-Setup-0.3.1.exe
  • Size: 1,425,649,621 bytes (about 1.33 GiB), including the F16 Whisper large-v3-turbo model
  • SHA-256: D9162170756FC3AFA19793BD0AD9F0828BB3314A4F097F4F8E8DB97050BA5DF0
  • Windows 10 or 11, x64 with AVX2. Per-user install under %LOCALAPPDATA%\Programs\Draft, no administrator prompt.

The installer is still unsigned, so SmartScreen may warn on first launch. Verify the hash before running it.

Known limits

  • Windows x64 is the supported release. The Apple Silicon macOS preview is ad-hoc signed and has not been run on a physical Mac.
  • Real-speech WER parity with the previous generation is still being measured. No accuracy claim is made ahead of it.
  • The History pane is not built. The bounded, text-only store and draftctl history exist.
  • There is no in-app updater and no update ping.
  • On CPU, the same F16 model takes a few seconds per clip, not milliseconds. A 4 GB machine is tight because the 1.6 GB model is held in RAM; Q8_0 remains the smaller catalog alternative and is never selected silently.
Assets 4
Loading