sayloom.Lifetime ↗Start with a clear, small result.
This guide walks through one task: “Separate speech models from writing models.” The useful outcome is a complete, reviewable piece of text, rather than simply seeing an animation or a successful provider response. Use a synthetic example first, then repeat the task in the app and field where you actually work.
Speech recognition models turn audio into text. Writing models transform text after recognition. Keeping these jobs separate makes troubleshooting much easier. A model with fewer parameters can use fewer resources, but that does not establish that it preserves names, actions and dates well enough for your task.
One thought. Check the words.
My microphone creates the source transcript.
My microphone creates the source transcript.
Use the example as a test case, not a promise of identical automatic output. Speak in your own natural voice and note the exact words you said. If the result differs, separate a speech-recognition mistake from a change introduced by formatting or a destination-field problem.
A repeatable way through it.
- 01
Test plain Dictation first. If its original transcript is wrong, changing the optional writing model will not repair the microphone.
- 02
Use English Small for the initial English speech test and multilingual Base for language detection when needed.
- 03
For optional local writing, start Ollama on the Mac and select one of the models actually installed there.
- 04
Run the same synthetic sample through 1B and 3B. Compare facts, question intent, omissions and the time taken.
- 05
Choose based on observed output and available memory, not a model name alone. A lightweight model can still change an important action.
- 06
Keep speech and writing settings explicit. Mac models and API keys do not automatically synchronize to phones.
Know which layer you are changing.
| Layer | What to check | What it proves |
|---|---|---|
| Device / edition | Desktop · Full Mac when using full Mac controls | You are testing the intended implementation |
| Microphone | Selected input, permission and Listening state | Capture is ready, not that the words are correct |
| Speech / text | Original transcript and optional formatting output | You can identify where meaning changed |
| Destination | Original field, cursor and complete result | The reviewed text reached the right place |
A successful microphone permission check does not establish recognition accuracy. A good original transcript does not establish that an AI rewrite preserved meaning. A finished result in Sayloom does not prove that another app accepted the paste. These distinctions make a failure more specific and easier to resolve.
A short review is worth the pause.
- ✓ The recording reaches Listening before the first word.
- ✓ The first and last intended words are present.
- ✓ Names, actions, dates and negation still mean what you intended.
- ✓ The destination or copy result contains the complete reviewed text.
Downloading an Ollama text model does not enable microphone capture.
If the task still fails, keep the expected sentence, actual result, edition and build number together. Change one setting per attempt. Use the free microphone diagnostic for signal checks; do not repeatedly spend trial uses trying to diagnose an unrelated formatting problem.
Keep the workflow honest.
Does this require an optional AI provider?
Plain Dictation does not require Ollama or a text-provider key. Email can use local layout. AI enhancement depends on the provider configured for that mode, and provider fees or privacy terms are separate. Check the original transcript before changing a writing provider.
Will this work the same way on every platform?
No. The full Mac Desktop, current Mac Store edition and mobile implementations have different capabilities. Mobile speech depends on device and language support. Review the relevant edition and use a supported editable field rather than assuming a Mac shortcut, model file or permission applies everywhere.
Product task guide · prepared from the Sayloom implementation and linked official references. Examples use synthetic text. Capabilities and prices can change; check the app and store before making a decision.