fix: reject near-silent audio to prevent Whisper hallucinations

src-tauri/src/managers/audio.rs

@@ -465,8 +465,12 @@ impl AudioRecordingManager {
 
                 // Pad if very short
                 let s_len = samples.len();
-                // debug!("Got {} samples", s_len);
-                if s_len < WHISPER_SAMPLE_RATE && s_len > 0 {
+                const MIN_SPEECH_SAMPLES: usize = 1600; // 100ms at 16kHz
+                if s_len < MIN_SPEECH_SAMPLES {
+                    // Too short to be real speech — SmoothedVad minimum real output is ~8000 samples
+                    // (15 prefill + 2 onset + 15 hangover frames). Anything shorter is leakage.
+                    Some(Vec::new())
+                } else if s_len < WHISPER_SAMPLE_RATE {
                     let mut padded = samples;
                     padded.resize(WHISPER_SAMPLE_RATE * 5 / 4, 0.0);
                     Some(padded)

src-tauri/src/managers/transcription.rs

@@ -458,6 +458,17 @@ impl TranscriptionManager {
             return Ok(String::new());
         }
 
+        const RMS_SILENCE_THRESHOLD: f32 = 0.005;
+        let rms = (audio.iter().map(|&s| s * s).sum::<f32>() / audio.len() as f32).sqrt();
+        if rms < RMS_SILENCE_THRESHOLD {
+            debug!(
+                "Audio RMS {:.6} below silence threshold {:.4}; skipping transcription",
+                rms, RMS_SILENCE_THRESHOLD
+            );
+            self.maybe_unload_immediately("silent audio");
+            return Ok(String::new());
+        }
+
         // Check if model is loaded, if not try to load it
         {
             // If the model is loading, wait for it to complete.