Why Whisper types "Thank you" on silence, and how to filter it
Hold a dictation key, say nothing, let go, and Whisper may hand you "Thank you." It does the same at the end of a recording that trails off into quiet. It's a well-known Whisper quirk, and in a dictation app it means text you never said lands in a chat window. Here's where it comes from, what whisper.cpp offers against it, and the filter we ended up with in FnFlow.
What the phantoms look like
We went through the first 1,295 recordings in FnFlow's history. In 33 of them Whisper had written something nobody said. Most were a lone "Thank you." Others were the Russian for "To be continued..." or a credit for whoever made the subtitles. The filter we wrote also covers "Thanks for watching", a bare ellipsis and sound descriptions like "*sad music*".
Where they come from
OpenAI trained Whisper on 680,000 hours of audio paired with transcripts collected from the internet. Its model card says that because the training was weakly supervised on noisy data, the model may output text that isn't in the audio. OpenAI's hypothesis is that the model mixes two jobs, transcribing the audio and predicting the next likely words.
The specific phrases point to subtitles. Sign-offs and subtitle credits often sit at the end of a video, over silence or music. That's our reading, not something OpenAI states. When there's nothing to transcribe, the model seems to fall back on text that often came with silence in its training data.
What the decoder already checks
For every 30-second window Whisper estimates the probability that there's no speech. OpenAI's package and whisper.cpp both treat a window as silent only when that probability is above 0.6 and the average log probability of the decoded text is below -1.0. The second condition is why phantoms get through. When the model is confident about "Thank you.", the window doesn't count as silent.
whisper-cli exposes these knobs and a few more:
-nth 0.60 no-speech threshold
-lpt -1.00 log probability threshold
-sns suppress non-speech tokens
--suppress-regex RE suppress tokens that match a regex
-mc 0 don't carry decoded text into the next window
Lowering -nth marks more windows as silent and risks dropping quiet speech. --suppress-regex works on tokens, so it's a blunt tool for whole phrases. OpenAI notes that not conditioning on previous text makes the model less likely to get stuck in a loop. As we read the current whisper.cpp source, -mc 0 also disables your --prompt, so you lose custom vocabulary with it.
Cut the silence before Whisper hears it
whisper.cpp can run a voice activity detector first. The README describes it this way: the audio goes through a Silero VAD model, and only the speech segments are passed to Whisper.
./models/download-vad-model.sh silero-v6.2.0
./build/bin/whisper-cli -m models/ggml-large-v3-turbo-q8_0.bin -f memo.wav \
--vad -vm models/ggml-silero-v6.2.0.bin
Defaults include a 0.5 speech threshold, a 250 ms minimum speech length and 30 ms of padding around each segment. We haven't tested VAD in FnFlow, so we can't tell you how much it reduces phantoms on dictation.
Filter what's left
FnFlow handles this in three places. Presses under 0.6 seconds skip recognition entirely. In our history, 70 recordings had failed, nearly all shorter than 1.3 seconds and caused by brushing fn. When Whisper returns nothing, the app says it heard no speech.
The last step is a filter on the text. It removes a phantom only when it stands as its own sentence, so "Tell him thank you" stays intact. Subtitle credits with distinctive markers are removed wherever they appear. In English text we're stricter still: a "Thank you." next to real sentences stays, unless the whole result is phantoms. Run over all 1,295 recordings, the filter changed exactly the 33 with phantoms and produced no false removals.
If you're building your own, start from your own history rather than a list from the internet. Match whole sentences, ignore case, and re-run the filter over every past transcript to count what it changes. Something as simple as this is a fine start:
re.fullmatch(r"(thank you|thanks for watching)[.!]*", sentence.strip(), re.I)
The full story of the filter is in how I built a local dictation app, and the whisper.cpp setup is in running Whisper locally on a Mac.
Questions
Is this a bug in whisper.cpp?
No. OpenAI's model card says Whisper can output text that wasn't spoken, because of how it was trained. whisper.cpp runs the same model, so it inherits the behavior.
Will a bigger model stop it?
We haven't measured that, so we won't promise it. The phantoms in our history came from the models we used day to day, and the filter handles them regardless of model.
Can't I just delete every "thank you"?
That would eat real speech. People say thank you. FnFlow removes a phantom only when it stands as a sentence on its own, and on 1,295 recordings that removed exactly the 33 phantoms and nothing else.
Does FnFlow use voice activity detection?
Not today. It skips recognition for very short presses, reports when Whisper finds no speech, and filters phantom sentences from the text.
Local Wispr Flow alternativeFnFlow vs VoiceInkVoice input for CursorHow 7 Mac dictation apps get you to your first dictationWhisper.cpp on Apple Silicon: which large-v3-turbo quantization to shipHow I built a local dictation appVoice dictation in Claude Code and the terminalWhy holding fn doesn't work for dictation on Mac (and how to fix it)How to run Whisper locally on an Apple Silicon MacSpokenly alternative for developersFnFlow vs EnviousWisprHow to dictate a good prompt for AI coding toolsVoice dictation in VS Code on a MacTeaching Whisper your jargon: initial prompt and replacementsPrivate dictation on a Mac at work: what actually leaves your machineVoice input for Codex CLI and the ChatGPT appVoice coding on a Mac: what actually worksWriting git commit messages by voiceDictating code review comments on GitHubDictation in IntelliJ IDEA, PyCharm and WebStorm on a MacDictation in Zed on a MacDictation in Terminal, iTerm2 and WarpWhich Whisper model to use on an Apple Silicon MacPush-to-talk dictation on a Mac: hold, toggle or streamSuperwhisper alternative for developersClaude Code voice mode: how to enable /voice and fix itWhisper apps: five ways to run Whisper locallyTalon voice coding vs a dictation appPush to talk on a MacDoes Claude have a voice mode?VoiceInk vs Wispr FlowVoiceInk vs SuperwhisperParakeet vs Whisper on a MacApple SpeechAnalyzer vs WhisperWhisper.cpp vs faster-whisper on a MacMlx-whisper vs whisper.cppGitHub Copilot voice inputGemini CLI voice inputVoice input in Google AntigravityVibe coding with voice input