batchalign 0.7.16__tar.gz → 0.7.17__tar.gz

{batchalign-0.7.16/batchalign.egg-info → batchalign-0.7.17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: batchalign
-Version: 0.7.16
+Version: 0.7.17
 Summary: Python Speech Language Sample Analysis
 Author: Brian MacWhinney, Houjun Liu
 Author-email: macw@cmu.edu, houjun@cmu.edu

{batchalign-0.7.16 → batchalign-0.7.17}/batchalign/models/resolve.py

@@ -12,7 +12,7 @@ resolver = {
     },
     "whisper": {
         'eng': ("talkbank/CHATWhisper-en-large-v1", "openai/whisper-large-v2"),
-        # 'yue': ("alvanlii/whisper-small-cantonese", "alvanlii/whisper-small-cantonese"),
+        'yue': ("alvanlii/whisper-small-cantonese", "alvanlii/whisper-small-cantonese"),
     }
 }
 

{batchalign-0.7.16 → batchalign-0.7.17}/batchalign/models/utterance/cantonese_infer.py

@@ -150,15 +150,45 @@ class BertCantoneseUtteranceModel(object):
             final_passage.append(self.tokenizer.convert_tokens_to_string(res_toks))
 
         # Step 4: Join processed chunks together into the final passage
-        final_text = ' '.join(final_passage)
+        final_passage = ' '.join(final_passage)
 
         print("Text processing completed. Generating final output...")
-
+        
         # Optionally, tokenize the final text into sentences based on punctuation
-        try:
-            split_passage = sent_tokenize(final_text)
-        except LookupError:
-            nltk.download('punkt')
-            split_passage = sent_tokenize(final_text)
+        def custom_sent_tokenize(text):
+            # Regular expression to match sentence-ending punctuation marks (. ! ?)
+            sentence_endings = re.compile(r'([.!?])')  
+            split_passage = []
+        
+            # Split the passage based on punctuation marks and keep them
+            parts = re.split(sentence_endings, text) 
+            
+            # Debug: Output the parts after splitting
+            print(f"Parts after splitting: {parts}")
+        
+            # Combine parts and punctuation together
+            for i in range(0, len(parts) - 1, 2):
+                sentence = parts[i] + parts[i + 1]  # Join sentence with punctuation
+                print(f"Sentence formed: {sentence}")  # Debug: Output the current sentence
+                
+                if sentence.strip():  # Only add non-empty sentences (check for non-whitespace content)
+                    split_passage.append(sentence)
+        
+            # If the last part doesn't have punctuation, we handle it here
+            if len(parts) % 2 != 0:  # If there's no punctuation at the end
+                last_part = parts[-1].strip()
+                print(f"Last part without punctuation: {last_part}")  # Debug: Output the last part
+                
+                if last_part:  # Only add non-empty sentences
+                    split_passage.append(last_part)
+        
+            # Final output
+            print(f"Final split passage: {split_passage}")
+            return split_passage
+            
+        split_passage = custom_sent_tokenize(final_passage)
+
+        # Debugging: Output the sentences after splitting
+        print(f"Final sentences: {split_passage}")
 
         return split_passage

{batchalign-0.7.16 → batchalign-0.7.17}/batchalign/models/whisper/infer_asr.py

@@ -68,6 +68,22 @@ class WhisperASRModel(object):
         self.__config = GenerationConfig.from_pretrained(base)
         self.__config.no_repeat_ngram_size = 4
 
+        
+        if language == "Cantonese":
+            self.__config.no_repeat_ngram_size = 4
+            self.__config.no_timestamps_token_id = 50363
+            self.__config.alignment_heads = [
+                [5, 3],
+                [5, 9],
+                [8, 0],
+                [8, 4],
+                [8, 8],
+                [9, 0],
+                [9, 7],
+                [9, 9],
+                [10, 5]
+            ]
+
         self.pipe = pipeline(
             "automatic-speech-recognition",
             model=model,
@@ -159,7 +175,7 @@ class WhisperASRModel(object):
         if self.lang == "Cantonese":
             config = {
                 "repetition_penalty": 1.001,
-                # "generation_config": self.__config,
+                "generation_config": self.__config,
                 # "task": "transcribe",
                 # "language": self.lang
             }

{batchalign-0.7.16 → batchalign-0.7.17}/batchalign/pipelines/asr/utils.py

@@ -71,7 +71,7 @@ def retokenize_with_engine(intermediate_output, engine):
     final_outputs = []
 
     for speaker, utterance in intermediate_output:
-        # becasue we are using an utterance engine, we need
+        # because we are using an utterance engine, we need
         # to get rid of all the preexisting punctuation
         for i in utterance:
             for j in MOR_PUNCT+ENDING_PUNCT:
@@ -84,8 +84,12 @@ def retokenize_with_engine(intermediate_output, engine):
         joined = joined.replace("。", ".")
         split = engine(joined)
 
+        # Initialize current index to track position in original utterance
+        current_index = 0
+
         # align the utterance against original splits and generate final outputs
         for i in split:
+            # Check if the split has ending punctuation
             if i[-1] in ENDING_PUNCT:
                 new_ut, delim = (i[:-1].split(" "), i[-1])
             else:
@@ -94,12 +98,18 @@ def retokenize_with_engine(intermediate_output, engine):
             tmp = []
 
             for s in new_ut:
-                try:
-                    tmp.append((s, utterance.pop(0)[1]))
-                except IndexError:
-                    continue
+                if current_index < len(utterance):
+                    # Use current element and move index forward
+                    tmp.append((s, utterance[current_index][1]))
+                    current_index += 1
+                else:
+                    # Append with default timestamp if utterance is exhausted
+                    tmp.append((s, [None, None]))
+                    
+            if current_index >= len(utterance):
+                tmp.append((delim, [None, None]))  # Append the punctuation
 
-            final_outputs.append((speaker, tmp+[[delim, [None, None]]]))
+            final_outputs.append((speaker, tmp))
 
     return final_outputs
 
@@ -220,7 +230,6 @@ def process_generation(output, lang="eng", utterance_engine=None):
                     seen_word = True
                     words.append(Form(text=word, time=(int(start), int(end))))
             else:
-                if seen_word:
                     words.append(Form(text=word, time=None))
 
         final_utterances.append(Utterance(

{batchalign-0.7.16 → batchalign-0.7.17}/batchalign/pipelines/morphosyntax/ud.py

@@ -143,7 +143,7 @@ def handler(word, lang=None):
         pos,target = verbform(pos,target,word.text)
         target = target.replace(',', 'cm')
 
-
+    target = re.sub(r'@\w$', '', target)
     return f"{'' if not unknown else '0'}{pos}|{target}"
 
 # POS specific handler

batchalign-0.7.17/batchalign/version

@@ -0,0 +1,3 @@
+0.7.17
+March 4th, 2025
+Cantonese patches + removing special mor line annotations

{batchalign-0.7.16 → batchalign-0.7.17/batchalign.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: batchalign
-Version: 0.7.16
+Version: 0.7.17
 Summary: Python Speech Language Sample Analysis
 Author: Brian MacWhinney, Houjun Liu
 Author-email: macw@cmu.edu, houjun@cmu.edu

batchalign-0.7.16/batchalign/version

@@ -1,3 +0,0 @@
-0.7.16
-Feburary 27th, 2025
-More Hebrew features