fix(transcription): drop whisper-hallucinated segments on silent windows

src-tauri/src/transcription/mod.rs

@@ -87,6 +87,17 @@ fn resolve_language(requested: Option<&str>, multilingual: bool) -> Option<Strin
     }
 }
 
+/// Segments whisper.cpp itself flags as more likely silence than speech are
+/// dropped rather than emitted (see `run_full`) — whisper.cpp's CLI ships
+/// this same 0.6 default for `--no-speech-thold`.
+#[cfg(feature = "cpu-transcription")]
+const NO_SPEECH_THRESHOLD: f32 = 0.6;
+
+#[cfg(feature = "cpu-transcription")]
+fn is_likely_speech(no_speech_probability: f32) -> bool {
+    no_speech_probability <= NO_SPEECH_THRESHOLD
+}
+
 /// whisper.cpp-backed transcriber (CPU baseline; GPU via Cargo features).
 #[cfg(feature = "cpu-transcription")]
 pub struct WhisperTranscriber {
@@ -185,6 +196,15 @@ impl WhisperTranscriber {
             if text.is_empty() {
                 continue;
             }
+            // whisper.cpp's own `no_speech_thold` gate is a no-op (unimplemented
+            // upstream as of whisper-rs 0.16 / whisper.cpp v1.3.0+), so silent/
+            // near-silent windows still decode — and greedy short-window decode
+            // reliably hallucinates a short filler word ("you", "Thank you.")
+            // instead of emitting nothing. Drop those ourselves: matches
+            // whisper.cpp's own CLI default threshold for "this was silence".
+            if !is_likely_speech(seg.no_speech_probability()) {
+                continue;
+            }
             // Whisper timestamps are centiseconds (10ms units).
             let start_ms = offset_ms + seg.start_timestamp().max(0) as u64 * 10;
             let end_ms = if single_segment {
@@ -383,6 +403,17 @@ mod tests {
         assert_eq!(audio_ctx_for_window(60 * 16_000), 1500); // 60s window
     }
 
+    #[test]
+    fn is_likely_speech_drops_high_no_speech_probability_windows() {
+        // Silence/near-silence: whisper.cpp's own greedy hallucination case
+        // ("you", "Thank you.") on an otherwise-quiet window.
+        assert!(!is_likely_speech(0.9));
+        assert!(!is_likely_speech(NO_SPEECH_THRESHOLD + 0.01));
+        // Confident speech kept, including right at the threshold.
+        assert!(is_likely_speech(0.0));
+        assert!(is_likely_speech(NO_SPEECH_THRESHOLD));
+    }
+
     /// GPU spike (opt-in): times whisper.cpp on a chosen backend against a real
     /// model + wav. With `--features vulkan` and `WA_BACKEND=intel` (or nvidia/amd)
     /// whisper.cpp offloads to the GPU; `WA_BACKEND=cpu` is the baseline. Run: