Why whispering is hard
A whisper carries almost the same amount of energy as ordinary room noise. To a speech-to-text pipeline they look nearly identical, so any loudness threshold low enough to catch your whisper is also low enough to transcribe the fan, the street, and the person at the next desk. A tool that simply listens harder ends up typing the room.
Evoglyph solves this by asking a second question about the quiet audio that loudness alone cannot classify: is this actually you? That question only has to run on the ambiguous, quiet clips, which is what keeps it cheap and keeps your normal-volume dictation completely untouched.
The decision, in three steps
Every clip runs through a cascade that goes from free to more expensive, and stops as soon as it has an answer.
- Loudness. Evoglyph already measures how loud your speech was while it trims silence. Anything at or above a normal-speech floor (around −48 dBFS) is plainly speech rather than room noise, so it skips the rest of the check entirely. This is the core guarantee: normal-volume dictation is never gated, so a voice profile can never get in the way of everyday speaking.
- Identity. For quiet clips of at least about a second and a half, and only once you have enrolled, Evoglyph turns the audio into a compact voiceprint and compares it to your stored profile. A close match (similarity above roughly 0.55) clears and is transcribed. A poor match (below roughly 0.35) is treated as not-you. If you have no profile, this step stands down and does nothing. If you have a profile but the voiceprint cannot be computed, it fails safe and treats the clip as not-you.
- Content check. For the narrow band between those thresholds, where identity is genuinely uncertain, a short list of known ambient-noise phrases and a transcription-confidence floor settle the last few cases. Once identity is confirmed, this step is skipped.
When a clip is rejected, a short tone plays and the pipeline returns to idle. You never get partial or garbled text out of a rejected clip.
Enrolling your voice
Enrollment takes four short samples: three at a normal volume and one whispered. Recording both registers is what lets the profile recognise you when you drop to a whisper. Evoglyph checks each sample for loudness and register, averages them into a single profile, down-weights any sample that looks like an outlier, and always keeps the whispered one so the profile spans your full range.
You can enroll during onboarding (it is skippable) or later from Settings. A "Test my profile" control lets you record a fresh sample and see whether it Matched, was Uncertain, or did Not Match. It never shows a raw similarity number, just the plain result.
How the identity check works
The voiceprint is a compact numeric summary of your voice (256 numbers), computed on your Mac. It is compared to your enrolled profile by similarity, and it is never transmitted anywhere. When you are enrolled, Evoglyph captures raw audio so the sound at the moment you dictate matches the sound you enrolled with, which keeps the comparison honest.
If several quiet clips are rejected in quick succession (five within a minute), Evoglyph takes that as a sign your profile may be stale and suggests re-enrolling. You can re-enroll at any time from Settings.
Without enrollment, nothing breaks
If you never enroll, you lose nothing about normal dictation. You speak at a normal volume and this feature never engages. Only quiet, whisper-level dictation needs a profile. If you whisper without one, Evoglyph shows a one-time nudge inviting you to enroll, then gets out of the way.
Because the loudness step runs first, a poorly recorded profile can, at worst, reject a quiet clip. It can never block your normal-volume speech.
Privacy
Everything here runs on your Mac. The voiceprint is computed locally and never leaves the device. Your profile is stored with restrictive file permissions and an integrity check, in the app's support folder rather than the login Keychain, and it is versioned and written atomically so it cannot be left half-written. Like the rest of your data, it is protected by FileVault when you have it enabled. For the full picture of what stays local and every network call the app makes, see Where your audio goes.
The loudness signal that starts the whole cascade comes from Silero VAD, the same openly licensed voice-activity model Evoglyph uses to trim silence before transcription. The whisper gate builds on that open foundation and adds the identity and content checks on top.
Known edge cases
Whisper-level dictation is honest about its limits:
- Very short clips. The identity check needs roughly a second and a half of speech. Shorter quiet clips fall through to the content check instead.
- Someone else talking. Identity gating runs on quiet audio. Evoglyph does not currently filter out a second person speaking at a normal volume.
- Room and microphone. The margin between a whisper and true silence is genuinely small and depends on your environment, so results vary from room to room and mic to mic.
In on-device testing, none of 37 normal-volume clips were mistakenly gated, and every whispered clip from the enrolled speaker was accepted while stand-in impostor voices were rejected. Those impostor voices were synthetic rather than human, so treat the impostor result as an early signal rather than a finished guarantee.