Hear Summaries and Responses¶
Goal¶
Speak a request or observation, let a profile turn it into a short response, and hear that response through the Mac while the profile's other outputs continue normally.
The most useful first version is a spoken summary:
Dictation → local ASR → short LLM summary → local Screechback playback
This is an audio response to one dictation, not a continuous voice conversation. Each capture is processed independently unless the selected external LLM or agent endpoint manages its own session context.
Why This Pattern Is Useful¶
- You can hear the result while looking elsewhere.
- The spoken result can be shorter than the text pasted or saved.
- Local text-to-speech avoids sending the finished response to a separate speech service.
- Screechback runs independently, so playback does not hold up paste, History, Saved Outputs, or Notes.
Preset: Speak a Short Summary¶
1. Download and Test a Speech Model¶
- Open
App Settings→Local Speech Models & Voices. - Select a model for the language you intend to use.
- Review the displayed download source and destination.
- Download the model and its included voices.
- Open
Test Screechback…. - Enter a representative one- or two-sentence response and compare voices.
The first test can take longer while the model loads. For PocketTTS, 6-layer packs favour speed while 24-layer packs favour quality. See Screechback for catalogue details.
Selecting the model checks its installed state, and an in-app download updates that status immediately. To remove it later, use the confirmed Delete Model & Voices… action in App Settings or the profile; this removes only the selected model's owned resources and leaves sibling models and shared resources intact.
2. Create a Summary Processor¶
Create an LLM processor named Spoken Summary with a restrained prompt such as:
Summarize the user's dictated text in one or two short, natural sentences.
Preserve the important conclusion and any requested action.
Use plain spoken language.
Do not use Markdown, headings, bullet points, or preamble.
Return only the summary.
Use a local processor when the dictated content should remain on the Mac. A remote processor receives the transcript before local speech synthesis begins.
3. Add the Processor to a Profile¶
In Profiles → Manage Profiles…:
- Add
Spoken Summaryas an enabled LLM stage. - Choose the stage input that should be summarized.
- Decide whether the summary is also the profile's
Final Output.
Use one of these arrangements:
| Desired Result | Final Output | Screechback Text Source |
|---|---|---|
| Hear and paste the same summary | Spoken Summary |
Final profile output or Spoken Summary |
| Hear a summary but paste the original | Original text |
Spoken Summary |
| Hear a summary but paste a formatted branch | Formatting stage | Spoken Summary |
4. Configure Screechback¶
In the same profile's Screechback section:
- Select
Spoken Summaryas the text source. - Select the downloaded speech model and voice.
- Select the language when using Supertonic-3.
- Use
Test Screechback…if needed. - Enable Screechback.
You can configure all choices while the enable switch is off, which is useful while tuning the processor or voice.
5. Decide What Else the Profile Should Do¶
Screechback is only one independent output. Configure the rest of the profile according to the workflow:
| Setting | Quiet Review Profile | Speak-and-Paste Profile |
|---|---|---|
| Auto paste | Off | On |
| Temporary history | Optional | On |
| Save to Notes | Optional | Optional |
| Saved Outputs | Optional | Optional |
| Trigger routing | Off initially | Off initially |
Generated Screechback audio is temporary and is not currently saved into Notes or Saved Outputs. The transcript artifacts selected elsewhere in the profile remain available.
Simpler Preset: Parrot the Result¶
An LLM is not required. To hear the same text the profile would otherwise deliver:
- Leave LLM post-processing off.
- Select
Final profile outputas the Screechback source. - Select and test a speech model and voice.
- Enable Screechback.
This reads the finalized transcript aloud. It is useful for confirming recognition or exploring voices, although it does not create a new response.
Advanced Pattern: Send to an Agent Endpoint¶
An LLM processor can point at an OpenAI-compatible system that forwards requests to a more capable agent or session-aware service. From OpenScreech's perspective, it is still one processor request per dictation.
If the remote system maintains a session, it—not Screechback—owns conversation history and context. Confirm how that endpoint identifies sessions before relying on continuity. OpenScreech does not currently expose a native conversation/thread model for Screechback.
Validation Checklist¶
- The profile is mapped to the hotkey you use.
- The summary stage returns brief plain text in its tester.
- The selected speech model reports
Ready to use. - The chosen voice sounds appropriate with realistic text.
- The first live dictation produces playback.
- Paste, History, Notes, and Saved Outputs behave independently as configured.
Common Pitfalls¶
| Pitfall | Outcome | Fix |
|---|---|---|
| Summary prompt returns Markdown | Headings or punctuation are spoken awkwardly | Ask for short plain spoken sentences only |
| Wrong profile is mapped to the hotkey | No playback or the wrong voice | Check Manage Hotkeys… |
| LLM stage was removed | Screechback falls back to Final profile output |
Re-select the intended source after editing stages |
| Model has not loaded before | First response appears unusually slow | Wait for the first load, then test again |
| PocketTTS 24-layer is too slow | Playback begins later than desired | Compare the corresponding 6-layer pack |
| Supertonic language is wrong | Pronunciation does not match the response | Set language separately from voice style |
| Long responses are selected | Playback is verbose and blocks the listener's attention | Use a dedicated concise summary stage |