Question 1

How do I fix Hold/Push-to-Talk not working in a fresh Windows build when Tap mode works?

Accepted Answer

The build is silently missing `windows-key-listener.exe`. The download script searches only the first 50 GitHub releases with `tagPrefix`, and the required release has moved beyond that page. The script warns but exits with code 0, so the build appears successful.

Fix `scripts/download-windows-key-listener.js` by replacing the paginated prefix lookup with an exact tag lookup:

```javascript
// Before
const release = await fetchLatestRelease(REPO, { tagPrefix: "windows-key-listener-v" });

// After
const release = await fetchLatestRelease(REPO, { tag: "windows-key-listener-v1.0.0" });
```

Then rerun:

```powershell
node scripts/download-windows-key-listener.js
npm run prebuild:win
npm run build:renderer
electron-builder --win
```

Verify that `resources\bin\windows-key-listener.exe` exists before packaging. Rebuild and launch the new app; Hold mode should then capture keydown/keyup events.

Apply the same exact-tag change to the other affected scripts to prevent future failures:

- `download-windows-fast-paste.js`: `windows-fast-paste-v1.0.0`
- `download-text-monitor.js`: `windows-text-monitor-v1.0.0`
- `download-meeting-aec-helper.js`: `meeting-aec-helper-v1.0.0`
- `download-windows-mic-listener.js`: `windows-mic-listener-v1.1.0`
- `download-windows-system-audio-helper.js`: `windows-system-audio-helper-v1.1.0`

Temporary workaround: use Settings → Activation mode → Tap, or provide a locally compiled `windows-key-listener.exe`. This affects OpenWhispr 1.9.2 and earlier source builds once the repository has more than 50 releases.

Question 2

Why does OpenWhispr return 401 "Incorrect API key provided" after changing the BYOK API key or endpoint?

Accepted Answer

This is a known credential-caching issue in OpenWhispr 1.9.2. When you change a Custom/BYOK provider's API key, base URL, or model in Settings, the dictation window can keep using the previously cached key until the app restarts, causing provider B to reject the old key with a 401 error.

Immediate workaround:
1. Save the new Custom/BYOK settings.
2. Fully quit OpenWhispr (do not only close the dictation or settings window).
3. Relaunch OpenWhispr.
4. Start dictation again; the new endpoint and API key will be used.

Permanent fix: upgrade to an OpenWhispr release containing PR #2527, which broadcasts credential changes to other windows, refreshes in-memory credentials, and invalidates cached API keys. A follow-up fix also prevents meeting/note-recording connections from reusing a prepared connection after a key is saved.

Question 3

Why does OpenWhispr Cloud dictation return no transcript when my Dictionary has about 1,000 terms?

Accepted Answer

This happens with BYOK OpenAI transcription using the default `gpt-transcribe` model: the macOS app sends every Dictionary term as a separate multipart `keywords[]` part. When the Dictionary contains roughly 1,000 terms, OpenAI rejects the upload, so normal dictation returns no transcript. The fix in PR #2224 caps `keywords[]` at 900 terms and sends the remaining terms through the `prompt` parameter instead. Update to a release containing this fix. As a temporary workaround, reduce the Dictionary to fewer than 900 terms or use History Retry.

Question 4

Why does the desktop app show Free after I purchased Pro through the iPhone App Store?

Accepted Answer

This is a desktop display bug affecting v1.10.2 and earlier `main` builds. Your Pro entitlement is active when the API returns `isSubscribed: true` and `entitlementSources.provider: true`; the app incorrectly only checks `entitlementSources.personal`, so App Store subscribers are labeled Free. Do not purchase Pro again. Update to a desktop release containing PR #2535. If self-hosting/building, treat `entitlementSources.provider` as a valid subscription in `src/hooks/useUsage.ts`, `src/components/SettingsPage.tsx`, and `src/lib/upsell.ts`—for example, include it when calculating `isPersonallySubscribed`—and route billing management for these users to the App Store instead of the Stripe portal. As a temporary workaround, confirm access via the API/account usage; App Store billing and cancellation remain in iPhone Settings > [your name] > Subscriptions.

Question 5

Why does the first OpenWhispr AI cleanup take about 10 seconds after idle, and how do I keep the reasoning model loaded?

Accepted Answer

The local `llama-server` is unloaded after `IDLE_TIMEOUT_MS = 5 * 60 * 1000` (5 minutes). The next cleanup must reload the full GGUF model from disk—for Gemma 4 E4B Q4_K_M, about 5.1 GB—which causes the ~10-second delay. To fix it, update OpenWhispr if needed, go to **Settings → Language Models**, and enable **Keep model loaded**. The default 5-minute unload behavior remains unchanged unless you opt in; enabling it keeps the local language model resident and removes the cold-start reload at the cost of continued memory usage.

OpenWhispr

Core Features

What It Can't Do

Use Cases

Tags

Getting Started

Similar Projects