A speech model can produce plausible words when the input contains little or no speech. Whisper's official model card explicitly warns that output can include text not spoken in the audio. In SuprFlow 1.0.0, short recordings, streaming, and long chunked recordings take different decode paths. The app filters some placeholder tokens and attempts to reject vocabulary-prompt echoes, but those checks cannot prove every inserted word was spoken.
Treat an invented sentence as a transcription error, not as a real note. It is especially important to stop before an extra phrase enters a client email, CRM record, or audit working paper.
Isolate the trigger with fictional text
- Record the signed app version, speech model, selected language, microphone, formatter, and destination app. Use a blank local note and a made-up sentence, not a confidential file.
- Dictate “Project Maple has not approved seven units” once, then stop promptly. Check whether any word was added after the sentence. Compare the inserted text with the raw history version if available; a hosted formatter can add a second kind of rewrite after speech recognition.
- Repeat with five seconds of silence before the sentence, then with silence after it. Keep the microphone and model fixed. Note whether the extra phrase appears only in one condition.
- Repeat in a quieter room or with the built-in microphone. Check System Settings → Sound → Input for the device macOS actually uses. Background TV, another speaker, or a Bluetooth input change can explain some apparent “phantom” words.
- If you use Dictionary entries, try a harmless repeat with a distinctive fictional word removed from the test vocabulary. The app passes vocabulary words as a recognition prompt and filters outputs dominated by those words; a filter can miss other invented text. Do not remove a real work dictionary merely to troubleshoot one test.
- If the issue remains, compare another downloaded model with the same script and configuration. Record actual errors rather than declaring one model universally better.
SuprFlow 1.0.0 disables some no-speech filtering for recordings under three seconds because padding can cause real short speech to be dropped. Streaming and long-chunk paths also make their own threshold choices. Those are implementation decisions, not user-facing sliders in the current settings. A user should test the microphone and workflow before looking for an undocumented threshold control.
Decide whether a note is safe to use
Cross out words with no source in what was spoken. Check the material terms: not, names, amounts, dates, percentages, and who owns an action. If an AI formatter also ran, compare raw, cleaned, and formatted history where the signed app exposes them. For a client or audit record, have the responsible person approve the corrected text in the official destination system. The app's fluent wording is not evidence that a statement was spoken or true.
If the same fictional test repeatedly produces unspoken text, send support the test script, model and language, microphone, duration, and whether the error occurs on silence before or after speech. SuprFlow 1.0.0 can print excerpts to local diagnostics, so inspect any log before sharing it. Do not attach a real client recording or raw transcript to an ordinary bug report.
Questions people ask
Does a bigger speech model eliminate invented words?
No model size guarantees that. Compare models on the same Mac, microphone, language, and script; record material errors and latency together.
Can Dictionary words cause a phrase to appear when I was silent?
The app passes vocabulary terms to WhisperKit and includes a filter for outputs that mainly echo those terms. That indicates the risk was anticipated, but it does not establish the cause of your specific output. Test with fictional vocabulary and document the result.
Is this the same as an AI formatter changing my sentence?
No. Recognition can invent text before formatting. A formatter can separately rewrite recognized text. Compare the raw history version with the final inserted version to locate the stage.
Related guides
- Compare SuprFlow speech models
- Check Bluetooth microphone quality
- Teach SuprFlow names and acronyms
- What raw and formatted history contains
- Report a bug without client data