step3

src/vad/mod.rs

@@ -44,6 +44,20 @@ pub struct VadSettings {
     pub speech_pad_ms: u32,
     /// Speech required to fire the barge-in onset (`VAD_BARGE_IN_MIN_SPEECH_MS`).
     pub barge_in_min_speech_ms: u32,
+    /// Drop an utterance shorter than this before transcription
+    /// (`VAD_MIN_UTTERANCE_MS`, default 300). `0` disables the check. Filters
+    /// the short VAD false-triggers (breaths, clicks, echo).
+    pub min_utterance_ms: u32,
+    /// Drop an utterance whose average speech probability is below this before
+    /// transcription (`VAD_MIN_AVG_PROB`, default 0.25). `0.0` disables the
+    /// check. Filters the soft/noisy segments that the STT would otherwise
+    /// turn into hallucinated text. The average is taken over the **voiced
+    /// windows only** (probability ≥ `neg_threshold`), so the trailing
+    /// hangover silence (`min_silence_ms`) does not dilute short utterances
+    /// like "uno" / "due". A genuine utterance still fluctuates around the
+    /// speech threshold on its onset/release, so keep this well below
+    /// `threshold` (default 0.5).
+    pub min_avg_prob: f32,
     /// Feed the STT engine with VAD-segmented utterances instead of letting it
     /// run its own RMS segmentation (`VAD_SEGMENT_STT`).
     pub segment_stt: bool,
@@ -62,9 +76,11 @@ impl Default for VadSettings {
             threshold: 0.5,
             neg_threshold: 0.35,
             min_speech_ms: 250,
-            min_silence_ms: 300,
+            min_silence_ms: 700,
             speech_pad_ms: 30,
             barge_in_min_speech_ms: 150,
+            min_utterance_ms: 300,
+            min_avg_prob: 0.25,
             segment_stt: true,
             barge_in: true,
             debug: false,
@@ -91,6 +107,8 @@ impl VadSettings {
         s.min_silence_ms = env_u32("VAD_MIN_SILENCE_MS", s.min_silence_ms);
         s.speech_pad_ms = env_u32("VAD_SPEECH_PAD_MS", s.speech_pad_ms);
         s.barge_in_min_speech_ms = env_u32("VAD_BARGE_IN_MIN_SPEECH_MS", s.barge_in_min_speech_ms);
+        s.min_utterance_ms = env_u32("VAD_MIN_UTTERANCE_MS", s.min_utterance_ms);
+        s.min_avg_prob = env_f32("VAD_MIN_AVG_PROB", s.min_avg_prob);
 
         // Hysteresis must stay below the speech threshold, and the fast
         // barge-in onset can never be slower than the utterance onset.

src/vad/silero.rs

@@ -219,8 +219,10 @@ pub struct VadDetector {
     speech_run_ms: u32,
     /// Consecutive silence milliseconds while in speech.
     silence_run_ms: u32,
-    /// Running probability average over the current utterance.
+    /// Running sum of speech probabilities over the voiced windows of the
+    /// current utterance.
     prob_sum: f32,
+    /// Number of voiced windows contributing to `prob_sum`.
     speech_windows: u64,
     /// Windows analysed since the detector was created (for logging).
     window_count: u64,
@@ -365,8 +367,14 @@ impl VadDetector {
         }
 
         self.utterance.extend_from_slice(window);
-        self.prob_sum += prob;
-        self.speech_windows += 1;
+        // Average only over the voiced windows. The trailing hangover (the
+        // `min_silence_ms` of silence that closes the segment) must not dilute
+        // the score: for a short word like "uno" it is most of the segment and
+        // would drag the mean below any sane gate.
+        if prob >= self.settings.neg_threshold {
+            self.prob_sum += prob;
+            self.speech_windows += 1;
+        }
 
         if prob < self.settings.neg_threshold {
             self.silence_run_ms = self.silence_run_ms.saturating_add(WINDOW_MS);