Parakeet v Whisper h2h #26

Merged
iamdoubz merged 21 commits from test_parakeet into main 2026-07-16 10:23:17 -05:00
Showing only changes of commit a71299803f - Show all commits
+78
View File
@@ -0,0 +1,78 @@
# WhispAssist v0.7.4
**Privacy-first, Windows-native meeting assistant — everything on-device, nothing leaves unless you say so.**
A small, focused performance release for **CPU transcription**. If WhispAssist runs the Whisper
engine on your CPU — either because your machine has no usable GPU/NPU, or because you selected
the CPU backend — this update takes it from far-slower-than-real-time to comfortably real-time.
Machines on the NPU or GPU backends are unaffected in behavior and simply pick up a
better-optimized build. Nothing here changes the privacy posture; there are no new downloads.
One universal installer (**MSI** and **NSIS**) covers every machine: **Vulkan** for all GPUs
(NVIDIA/AMD/Intel), the Intel **NPU** (OpenVINO), a **DirectML** fallback, and **CPU**.
---
## ⚡ CPU transcription fixed (up to ~90× faster on hybrid Intel CPUs)
Whisper-on-CPU previously spawned one worker thread per logical CPU. On modern hybrid Intel
processors (P-cores + E-cores + low-power E-cores) the workers synchronize in lock-step, so the
two slow low-power cores stalled *every* step while the remaining threads burned CPU waiting —
transcription ran ~20× slower than real time while looking like it was using the whole machine.
WhispAssist now caps Whisper's CPU decode at **4 threads** (the whisper.cpp upstream default).
Measured on a Core Ultra 5 135U over a 4-minute test corpus, same accuracy:
| | before | after |
|---|---|---|
| Speed vs real time | 20× **slower** | **4.6× faster** (RTF 0.218) |
| CPU cost per second of audio | 232.6 cpu-sec | 0.83 cpu-sec |
Live (streaming) transcription shares this code path and benefits equally. If you want to
experiment, the `WA_WHISPER_THREADS` environment variable overrides the cap.
## 🔧 Native code now built fully optimized
A build-toolchain issue caused the bundled native libraries (including whisper.cpp's CPU
kernels) to be compiled **without compiler optimization**. That's fixed; this release ships
optimized native code across the board. (For CPU transcription the thread fix above is the
dominant win; this one is belt-and-suspenders for everything else.)
---
## 📦 Install
**Requirements:** Windows 10 or 11 (x64). WhispAssist needs **WebView2** (preinstalled on Windows 11;
the installer fetches it on Windows 10). Importing from a file/URL additionally needs **`ffmpeg`**
(and **`yt-dlp`** for URLs) on your PATH — the Import dialog links to both.
1. Download **`WhispAssist_0.7.4_x64_en-US.msi`** (or the NSIS **`WhispAssist_0.7.4_x64-setup.exe`**).
2. Run it and accept the UAC prompt. If SmartScreen appears, choose **More info → Run anyway**.
3. Launch **WhispAssist** from the Start menu.
On first run WA picks the best transcription backend (**NPU → NVIDIA → AMD → Intel → CPU**). It runs
without admin rights, does **not** add itself to startup unless you opt in, and keeps all data under
`%LOCALAPPDATA%\WhispAssist`.
Deploying to many machines? See [`docs/enterprise-deployment.md`](docs/enterprise-deployment.md).
## 🔐 Checksums (SHA-256)
```
5c4e6d6b23a4b404c6fa200b6491e74f143c3a886042dc3bef3c2861168ef17c WhispAssist_0.7.4_x64_en-US.msi
eb2ececff06373e9cbb43ee6455c433c1b88785e9060b68ef7bed193191d9761 WhispAssist_0.7.4_x64-setup.exe
```
Verify after download:
```powershell
Get-FileHash .\WhispAssist_0.7.4_x64_en-US.msi -Algorithm SHA256
```
---
## Privacy, unchanged
Everything optional is **off by default**. With nothing configured, WhispAssist makes **no content
egress at all**. Recording is opt-in; sync/AI credentials live only in the OS credential store; the
MCP server is loopback-only and adds no egress; the deployment file never carries secrets. The
reachable-host allowlist is derived from your settings and enforced in the core.