Compress hallucinated XTTS silence

platform/docker/tts-gateway/test_tts_gateway.py

@@ -184,6 +184,19 @@ class AudioJoinTests(unittest.TestCase):
         sentence = gateway._join_pcm([("Fertig.", spoken), ("Weiter", spoken)])
         self.assertGreater(len(sentence), len(inline))
 
+    def test_long_internal_silence_is_shortened(self):
+        spoken = self.pcm([900] * 1000)
+        long_silence = self.pcm([0] * int(24000 * 1.5))
+        result = gateway._compress_internal_silence(spoken + long_silence + spoken)
+        expected_keep = int(24000 * gateway.INTERNAL_SILENCE_KEEP_MS / 1000) * 2
+        self.assertEqual(len(result), len(spoken) * 2 + expected_keep)
+
+    def test_natural_short_pause_is_preserved(self):
+        spoken = self.pcm([900] * 1000)
+        short_silence = self.pcm([0] * int(24000 * 0.3))
+        original = spoken + short_silence + spoken
+        self.assertEqual(gateway._compress_internal_silence(original), original)
+
 
 if __name__ == "__main__":
     unittest.main()

platform/docker/tts-gateway/tts_gateway.py

@@ -46,6 +46,9 @@ TRIM_THRESHOLD = int(os.getenv("XTTS_TRIM_THRESHOLD", "90"))
 TRIM_PADDING_MS = int(os.getenv("XTTS_TRIM_PADDING_MS", "18"))
 CROSSFADE_MS = int(os.getenv("XTTS_CROSSFADE_MS", "8"))
 SENTENCE_PAUSE_MS = int(os.getenv("XTTS_SENTENCE_PAUSE_MS", "65"))
+INTERNAL_SILENCE_THRESHOLD = int(os.getenv("XTTS_INTERNAL_SILENCE_THRESHOLD", "512"))
+INTERNAL_SILENCE_TRIGGER_MS = int(os.getenv("XTTS_INTERNAL_SILENCE_TRIGGER_MS", "650"))
+INTERNAL_SILENCE_KEEP_MS = int(os.getenv("XTTS_INTERNAL_SILENCE_KEEP_MS", "220"))
 
 SYNTHESIS_LOCK = threading.Lock()
 STATE_LOCK = threading.Lock()
@@ -432,6 +435,35 @@ def _fade_edge(pcm: bytes, *, fade_in: bool = False,
     return samples.tobytes()
 
 
+def _compress_internal_silence(pcm: bytes) -> bytes:
+    """Shorten XTTS silence hallucinations while preserving normal pauses."""
+    samples = array("h")
+    samples.frombytes(pcm)
+    if not samples:
+        return pcm
+    trigger = int(24000 * max(0, INTERNAL_SILENCE_TRIGGER_MS) / 1000)
+    keep = int(24000 * max(0, INTERNAL_SILENCE_KEEP_MS) / 1000)
+    if trigger <= 0 or keep >= trigger:
+        return pcm
+    output = array("h")
+    index = 0
+    while index < len(samples):
+        if abs(samples[index]) > INTERNAL_SILENCE_THRESHOLD:
+            output.append(samples[index])
+            index += 1
+            continue
+        end = index + 1
+        while end < len(samples) and abs(samples[end]) <= INTERNAL_SILENCE_THRESHOLD:
+            end += 1
+        run = end - index
+        if run >= trigger:
+            output.extend(samples[index:index + keep])
+        else:
+            output.extend(samples[index:end])
+        index = end
+    return output.tobytes()
+
+
 def _join_pcm(parts: list[tuple[str, bytes]]) -> bytes:
     """Join clips without clicks; pause only at real sentence boundaries."""
     if not parts:
@@ -459,7 +491,7 @@ def _join_pcm(parts: list[tuple[str, bytes]]) -> bytes:
                 faded = _fade_edge(bytes(output), fade_out=True)
                 output[:] = faded
             output.extend(_fade_edge(pcm, fade_in=True))
-    return bytes(output)
+    return _compress_internal_silence(bytes(output))
 
 
 def _wav(pcm: bytes) -> bytes: