Fix bug: same reference issue

api/src/inference/base.py

@@ -12,12 +12,12 @@ class AudioChunk:
 
     def __init__(
         self,
-        audio: np.ndarray,
+        audio: np.ndarray, # dtype: np.int16
-        word_timestamps: Optional[List] = [],
+        word_timestamps: Optional[List] = None, # Using None instead of `[]` to avoid reference to the same empty list.
         output: Optional[Union[bytes, np.ndarray]] = b"",
     ):
         self.audio = audio
-        self.word_timestamps = word_timestamps
+        self.word_timestamps = word_timestamps if word_timestamps is not None else []
         self.output = output
 
     @staticmethod

api/src/services/tts_service.py

@@ -65,7 +65,7 @@ class TTSService:
                 # Handle silence tags, eg: `[silent](0.5s)`
                 if match := SILENCE_TAG.match(chunk_text):
                     silence_duration = float(match.group(1))
-                    silence_audio = np.zeros(int(silence_duration * 24000), dtype=np.float32)
+                    silence_audio = np.zeros(int(silence_duration * 24000), dtype=np.int16)
                     if not output_format:
                         yield AudioChunk(silence_audio, output=b"")
                         return
@@ -89,7 +89,7 @@ class TTSService:
                         return
                     chunk_data = await AudioService.convert_audio(
                         AudioChunk(
-                            np.array([], dtype=np.float32)
+                            np.array([], dtype=np.int16)
                         ),  # Dummy data for type checking
                         output_format,
                         writer,