AzureTTS: work around word ordering issue at 8khz sample rate

src/pipecat/services/azure/tts.py

@@ -277,6 +277,11 @@ class AzureTTSService(WordTTSService, AzureBaseTTSService):
         self._started = False
         self._first_chunk = True
         self._cumulative_audio_offset: float = 0.0  # Cumulative audio duration in seconds
+        self._current_sentence_base_offset: float = 0.0  # Base offset for current sentence
+        self._current_sentence_duration: float = 0.0  # Duration from Azure callback
+        self._current_sentence_max_word_offset: float = (
+            0.0  # Max word boundary offset seen in current sentence (for 8kHz workaround)
+        )
         self._last_word: Optional[str] = None  # Track last word for punctuation merging
         self._last_timestamp: Optional[float] = None  # Track last timestamp
 
@@ -386,8 +391,14 @@ class AzureTTSService(WordTTSService, AzureBaseTTSService):
         word = evt.text
         sentence_relative_seconds = evt.audio_offset / 10_000_000.0
 
-        # Add cumulative offset to get absolute timestamp across sentences
+        # Use base offset captured at start of run_tts to avoid race conditions
-        absolute_seconds = self._cumulative_audio_offset + sentence_relative_seconds
+        # with callbacks from overlapping TTS requests
+        absolute_seconds = self._current_sentence_base_offset + sentence_relative_seconds
+
+        # Track max word offset for accurate cumulative timing
+        # (audio_duration from Azure doesn't always match word boundary offsets at 8kHz)
+        if sentence_relative_seconds > self._current_sentence_max_word_offset:
+            self._current_sentence_max_word_offset = sentence_relative_seconds
 
         if not word:
             return
@@ -492,9 +503,9 @@ class AzureTTSService(WordTTSService, AzureBaseTTSService):
             self._last_word = None
             self._last_timestamp = None
 
-        # Update cumulative audio offset for next sentence
+        # Store duration for cumulative offset calculation
         if evt.result and evt.result.audio_duration:
-            self._cumulative_audio_offset += evt.result.audio_duration.total_seconds()
+            self._current_sentence_duration = evt.result.audio_duration.total_seconds()
 
         self._audio_queue.put_nowait(None)  # Signal completion
 
@@ -530,6 +541,9 @@ class AzureTTSService(WordTTSService, AzureBaseTTSService):
         self._started = False
         self._first_chunk = True
         self._cumulative_audio_offset = 0.0
+        self._current_sentence_base_offset = 0.0
+        self._current_sentence_duration = 0.0
+        self._current_sentence_max_word_offset = 0.0
         self._last_word = None
         self._last_timestamp = None
 
@@ -604,6 +618,12 @@ class AzureTTSService(WordTTSService, AzureBaseTTSService):
                     self._started = True
                     self._first_chunk = True
 
+                # Capture base offset BEFORE starting synthesis to avoid race conditions
+                # Word boundary callbacks will use this value
+                self._current_sentence_base_offset = self._cumulative_audio_offset
+                self._current_sentence_duration = 0.0
+                self._current_sentence_max_word_offset = 0.0
+
                 ssml = self._construct_ssml(text)
                 self._speech_synthesizer.speak_ssml_async(ssml)
                 await self.start_tts_usage_metrics(text)
@@ -627,6 +647,16 @@ class AzureTTSService(WordTTSService, AzureBaseTTSService):
                     )
                     yield frame
 
+                # Update cumulative offset for next sentence
+                # At 8kHz, Azure's audio_duration doesn't match word boundary offsets,
+                # so we use max_word_offset as a workaround. At other sample rates,
+                # audio_duration is accurate.
+                # TODO: Remove after Azure fixes word boundary timing at 8kHz
+                if self.sample_rate == 8000:
+                    self._cumulative_audio_offset += self._current_sentence_max_word_offset
+                else:
+                    self._cumulative_audio_offset += self._current_sentence_duration
+
             except Exception as e:
                 yield ErrorFrame(error=f"Unknown error occurred: {e}")
                 yield TTSStoppedFrame()