feat: expose audio_duration_ms in TurnPrediction

crates/wavekat-turn/src/audio/pipecat.rs

@@ -522,11 +522,14 @@ impl AudioTurnDetector for PipecatSmartTurn {
             (TurnState::Unfinished, 1.0 - probability)
         };
 
+        let audio_duration_ms = (self.ring_buffer.len() as u64 * 1000) / SAMPLE_RATE as u64;
+
         Ok(TurnPrediction {
             state,
             confidence,
             latency_ms,
             stage_times,
+            audio_duration_ms,
         })
     }
 

crates/wavekat-turn/src/lib.rs

@@ -64,6 +64,12 @@ pub struct TurnPrediction {
     pub latency_ms: u64,
     /// Per-stage timing breakdown in pipeline order.
     pub stage_times: Vec<StageTiming>,
+    /// Duration of audio in the detector's buffer at prediction time (ms).
+    ///
+    /// For PipecatSmartTurn this reflects how much of the 8 s ring buffer
+    /// was filled. With soft reset the buffer may span multiple speech
+    /// segments, so this can exceed the current segment duration.
+    pub audio_duration_ms: u64,
 }
 
 /// A single turn in the conversation, for context-aware text detectors.

crates/wavekat-turn/tests/controller.rs

@@ -42,11 +42,13 @@ impl AudioTurnDetector for MockDetector {
             TurnState::Unfinished => 0.80,
             TurnState::Wait => 0.70,
         };
+        let audio_duration_ms = (self.buffer_len as u64 * 1000) / 16000;
         Ok(TurnPrediction {
             state,
             confidence,
             latency_ms: 0,
             stage_times: vec![],
+            audio_duration_ms,
         })
     }