Whisper invents text on silence in meetings and short dictations; the -50 dBFS gate does not catch it #193

HEX version: 2.3.0 for the recording, installed from the signed DMG. The same code is in 2.3.1; source checked again at main f8774a5.
Platform: macOS
System details: macOS 26.4.1, Apple M1 Pro, built-in microphone. Whisper large-v3-turbo, language Italian, recognition hints set (five product names).

What happened?

When I do not speak, Whisper still returns text. The text is a closing phrase ("Grazie a tutti", "Grazie"), my recognition hints copied word for word, or short fragments.

Meetings (beta). I recorded a 63-minute call and mostly listened. In transcript.ndjson, 152 of 185 microphone entries are text I did not say:

The system track has 10 more (8 "Grazie", 2 hints string) during remote silence. Echo suppression does not remove them, because no remote speech matches them.

Dictation. This is the case from #108, still on 2.3.0. A short press without speech (1–3 s) pastes "Grazie" or "Grazie a tutti". Today's history.json has 8 of these, and process.log for the same hours has no discarded near-silent dictation line, so the gate did not discard them.

Expected: no text when nobody speaks.

Steps to reproduce

1. Select Whisper large-v3-turbo and enter some recognition hints. Any language should do: I use Italian, and in English the closing phrase is usually "Thank you." (as in Inserting 'thank you' at the end of dictation sometimes #108).
2. Meetings: record a call and stay silent for a few minutes. Dictation: press the shortcut for 1–3 s without speaking.
3. Read the transcript or the pasted text. Expect the hints string, a closing phrase in the selected language, or short fragments.

Additional context

Where it comes from in the code

1. Meeting chunks have no silence gate. Every rolling chunk (transcribe_pending) and every chunk of transcribe_file goes to the model.
2. The dictation gate (is_near_silent) checks only the loudest 30 ms window against -50 dBFS. A click, the key press or room noise passes it.
3. The recognition hints are Whisper's initial_prompt. Each meeting chunk is a separate run, so each chunk starts with the prompt. On silence, Whisper often just repeats it.
4. Whisper's own no-speech check is already active with the transcribe-cpp defaults (no_speech_thold 0.6, logprob_thold -1.0). It does not stop these, because it skips a segment only when the logprob is also low.

Why a level threshold is not enough

I checked each microphone entry of the meeting against the audio under it. A gate "keeps" an entry when it would let that audio reach the model.

• The current gate keeps almost all invented text.
• A stricter level rule removes it, but also drops real speech. It would also go against silence_gate_discards_only_near_silent_captures, which keeps 100 ms of whispered speech at -43 dBFS.
• I also tried the PauseSegmenter speech test (12 dB above the adaptive floor). It marks speech in most of my silent stretches, because of room and keyboard noise.
• Silero separates the two groups. The one invented entry it keeps is a "Grazie a tutti." over 16 s of silence. I use threshold 0.4, not the default 0.5, because at 0.5 a whisper at -43 dBFS is not heard. The one real entry it misses is a 40 ms "Certo." at a chunk boundary. With Silero's default min_speech_ms of 250, 3 short real replies on the system track ("Ok", "Ciao") were lost; 160 keeps them.
• It scans the 63 minutes in 1.6 s on the CPU.

Notes on this test:

• one recording, one machine, entries checked one by one, not the real chunk path (the PR has a run on the real chunk path);
• the GGUF model is handy-computer/silero-vad-v6.2-gguf, F32.

Proposed fix

transcribe-cpp 0.4.0 has a Silero VAD (VadSession, with Silero v5.1.2 / v6.2.0 in GGUF or whisper.cpp .bin format). HEX pins =0.2.4. #129 also needs 0.4.0.

1. Run the VAD before the model. Discard a dictation clip with no speech segment. In meetings, skip chunks with no speech, or send only the speech spans.
2. Drop any Whisper segment whose text is the same as the initial prompt. This needs no threshold. Here it removes 29 microphone entries and 2 system entries.

I opened #194 with both parts, on transcribe-cpp 0.4.0. Open questions:

• Should the VAD model ship with the app, or download with the transcription model (as in the PR)?
• Should the VAD run for every model, or only for Whisper?

Related: #108 (same symptom in dictation, closed after the -50 dBFS gate), #129 (transcribe-cpp 0.4.0). The English live draft in meetings (#182, #185) is a separate problem and is not touched.