Fix ScaleSilence copying past the pause interval when scale > 1 (#3744)

GeneratedAudio::ScaleSilence() lengthens each detected pause by copying
n = (interval.end - interval.start) * scale samples starting at
interval.start.

When scale > 1 that is n > interval.end - interval.start, so the copy
runs past the end of the pause and into the audio that follows it. Two
consequences:

- The beginning of the next word is duplicated, which is audible as a
  stutter after every pause.
- On the last interval, interval.end == num_samples, so
  samples.begin() + interval.start + n is past samples.end() and the
  insert reads out of bounds.

Copy the pause once and pad with zeros to the requested length instead.
For scale <= 1 the behaviour is unchanged, and scale == 1 still returns
early before reaching this code.

Author: ekenberg

sherpa-onnx/csrc/offline-tts.cc

@@ -79,10 +79,21 @@ GeneratedAudio GeneratedAudio::ScaleSilence(float scale) const {
     ans.samples.insert(ans.samples.end(), samples.begin() + i,
                        samples.begin() + interval.start);
     i = interval.end;
-    int32_t n = static_cast<int32_t>((interval.end - interval.start) * scale);
-
-    ans.samples.insert(ans.samples.end(), samples.begin() + interval.start,
-                       samples.begin() + interval.start + n);
+    int32_t len = interval.end - interval.start;
+    int32_t n = static_cast<int32_t>(len * scale);
+
+    if (n <= len) {
+      ans.samples.insert(ans.samples.end(), samples.begin() + interval.start,
+                         samples.begin() + interval.start + n);
+    } else {
+      // scale > 1: copying n samples from interval.start would run past the
+      // end of the pause into the following speech (audible as a repeated
+      // word onset) and can read out of bounds on the final interval.
+      // Copy the pause once, then extend it with silence.
+      ans.samples.insert(ans.samples.end(), samples.begin() + interval.start,
+                         samples.begin() + interval.end);
+      ans.samples.insert(ans.samples.end(), n - len, 0.0f);
+    }
   }
 
   if (i < num_samples) {