filter [] words

transcribe/helpers/vadprocessor.py

@@ -113,6 +113,7 @@ class VADIteratorOnnx:
                  sampling_rate: int = 16000,
                  min_silence_duration_ms: int = 100,
                  max_speech_duration_s: float = float('inf'),
+                 speech_pad_ms: int = 30
                  ):
         self.model = OnnxWrapper(VAD_MODEL_PATH, True)
         self.threshold = threshold
@@ -123,7 +124,7 @@ class VADIteratorOnnx:
 
         self.min_silence_samples = sampling_rate * min_silence_duration_ms / 1000
         self.max_speech_samples = int(sampling_rate * max_speech_duration_s)
-        # self.speech_pad_samples = sampling_rate * speech_pad_ms / 1000
+        self.speech_pad_samples = sampling_rate * speech_pad_ms / 1000
         self.reset_states()
 
     def reset_states(self):
@@ -158,7 +159,8 @@ class VADIteratorOnnx:
 
         if (speech_prob >= self.threshold) and not self.triggered:
             self.triggered = True
-            speech_start = max(0, self.current_sample - window_size_samples)
+            # speech_start = max(0, self.current_sample - window_size_samples)
+            speech_start = max(0, self.current_sample - self.speech_pad_samples - window_size_samples)
             self.start = speech_start
             return {'start': int(speech_start) if not return_seconds else round(speech_start / self.sampling_rate, 1)}
 
@@ -174,7 +176,8 @@ class VADIteratorOnnx:
             if self.current_sample - self.temp_end < self.min_silence_samples:
                 return None
             else:
-                speech_end = self.temp_end - window_size_samples
+                # speech_end = self.temp_end - window_size_samples
+                speech_end = self.temp_end + self.speech_pad_samples - window_size_samples
                 self.temp_end = 0
                 self.triggered = False
                 return {'end': int(speech_end) if not return_seconds else round(speech_end / self.sampling_rate, 1)}

transcribe/helpers/whisper.py

@@ -52,7 +52,7 @@ class WhisperCPP:
                 initial_prompt=prompt,
                 language=language,  
                 # token_timestamps=True, 
-                # split_on_word=True,
+                split_on_word=True,
                 # max_len=max_len
             )
             return output

transcribe/utils.py

@@ -7,6 +7,51 @@ from scipy.io.wavfile import write
 import config
 import csv
 import av
+import re
+
+# Compile regex patterns once outside the loop for better performance
+p_pattern = re.compile(r"(\s*\[.*?\])")
+p_start_pattern = re.compile(r"(\s*\[.*)")
+p_end_pattern = re.compile(r"(\s*.*\])")
+
+
+def filter_words(res_word):
+    """
+    Filter words according to specific bracket patterns.
+    
+    Args:
+        res_word: Iterable of word objects with a 'text' attribute
+    
+    Returns:
+        List of filtered word objects
+    """
+    asr_results = []
+    skip_word = False
+    
+    for word in res_word:
+        # Skip words that completely match the pattern
+        if p_pattern.match(word.text):
+            continue
+        
+        # Mark the start of a section to skip
+        if p_start_pattern.match(word.text):
+            skip_word = True
+            continue
+        
+        # Mark the end of a section to skip
+        if p_end_pattern.match(word.text) and skip_word:
+            skip_word = False
+            continue
+        
+        # Skip words if we're in a skip section
+        if skip_word:
+            continue
+        
+        # Add the word to results if it passed all filters
+        asr_results.append(word)
+    
+    return asr_results
+
 def log_block(key: str, value, unit=''):
     if config.DEBUG:
         return 

transcribe/whisper_llm_serve.py

@@ -11,7 +11,7 @@ import config
 import collections
 from api_model import TransResult, Message, DebugResult
 
-from .utils import log_block, save_to_wave, TestDataWriter
+from .utils import log_block, save_to_wave, TestDataWriter, filter_words
 from .translatepipes import TranslatePipes
 from .strategy import (
     TranscriptStabilityAnalyzer, TranscriptToken)
@@ -191,6 +191,7 @@ class WhisperTranscriptionService:
             meta_item = self._transcribe_audio(audio_buffer)
             segments = meta_item.segments
             logger.debug(f"Segments: {segments}")
+            segments = filter_words(segments)
             if len(segments):
                 seg_text = self.text_separator.join(seg.text for seg in segments)
                 if self._temp_string: