patch werid hanging utterance bug

batchalign/models/utterance/infer.py

@@ -35,6 +35,7 @@ def __init__(self, model):
         self.model.eval()
 
     def __call__(self, passage):
+        print(passage)
         # input passage words removed of all preexisting punctuation
         passage = passage.lower()
         passage = passage.replace('.','')
@@ -67,7 +68,8 @@ def __call__(self, passage):
         prev_word_idx = None
 
         # for each word, perform the action
-        for indx, elem in enumerate(tokd.word_ids(0)):
+        wids = tokd.word_ids(0)
+        for indx, elem in enumerate(wids):
             # if its none, append nothing or if we have
             # seen it before, do nothing
             if elem is None or elem == prev_word_idx:
@@ -81,23 +83,31 @@ def __call__(self, passage):
             # set the working variable
             w = input_tokenized[elem]
 
-            # perform the edit actions
-            if action == 1:
-                w = w[0].upper() + w[1:]
-            elif action == 2:
-                w = w+'.'
-            elif action == 3:
-                w = w+'?'
-            elif action == 4:
-                w = w+'!'
-            elif action == 5:
-                w = w+','
+            # fix one word hanging issue
+            will_action = False
+            if indx < len(wids)-2 and classified_targets[0][indx+1] > 0:
+                will_action = True
+
+            if not will_action:
+                # perform the edit actions
+                if action == 1:
+                    w = w[0].upper() + w[1:]
+                elif action == 2:
+                    w = w+'.'
+                elif action == 3:
+                    w = w+'?'
+                elif action == 4:
+                    w = w+'!'
+                elif action == 5:
+                    w = w+','
+
 
             # append
             res_toks.append(w)
 
         # compose final passage
         final_passage = self.tokenizer.convert_tokens_to_string(res_toks)
+        print(final_passage)
         try: 
             split_passage = sent_tokenize(final_passage)
         except LookupError:

batchalign/pipelines/analysis/eval.py

@@ -8,7 +8,7 @@
 from batchalign.pipelines.asr.utils import *
 from batchalign.utils.config import config_read
 
-from batchalign.utils.dp import align, ExtraType, Extra
+from batchalign.utils.dp import align, ExtraType, Extra, Match
 
 import logging
 L = logging.getLogger("batchalign")
@@ -38,8 +38,16 @@ def __compute_wer(doc, gold):
         # ie: if we have <extra.payload> <extra.reference> +> substitution
         #     but if we have <extra.reference> <extra.reference> this is 2 insertions
 
+        cleaned_alignment = []
+
         for i in alignment:
+
             if isinstance(i, Extra):
+                if len(cleaned_alignment) > 0 and i.extra_type == ExtraType.REFERENCE and "name" in i.key and i.key[:4] != "name":
+                    cleaned_alignment.pop(-1)
+                    cleaned_alignment.append(Match(i.key, None, None))
+                    continue
+
                 if prev_error != None and prev_error != i.extra_type:
                     # this is a substitution: we have different "extra"s in
                     # reference vs. playload
@@ -64,6 +72,8 @@ def __compute_wer(doc, gold):
             else:
                 prev_error = None
 
+            cleaned_alignment.append(i)
+
         diff = []
         for i in alignment:
             if isinstance(i, Extra):

batchalign/version

@@ -1,3 +1,3 @@
-0.7.5-alpha.6
-September 3nd, 2024
-fix benchmark command, part 2
+0.7.5-alpha.7
+September 7th, 2024
+batch hanging utterance bug

scratchpad.py

@@ -15,7 +15,14 @@
 
 ########
 
-# from batchalign import *
+# from batchalign.models.utterance import infer
+
+# engine = infer.BertUtteranceModel("talkbank/CHATUtterance-zh_CN")
+# engine("我 现在 想 听 你说 一些 你 自己 经 历 过 的 故 事 好不好 然后 呢 我们 会 一起 讨 论 有 六 种 不同 的 情 景 然后 在 每 一个 情 景 中 都 需要 你 去 讲 一个 关 于 你 自己 的 一个 故 事 小 故 事")
+
+# doc = Document.new(media_path="/Users/houjun/Downloads/trial.mp3", lang="zho")
+# pipe = BatchalignPipeline.new("asr", lang="zho", num_speakers=2, engine="rev")
+# res = pipe(doc)
 
 # # with open("schema.json", 'w') as df:
 # #     json.dump(Document.model_json_schema(), df, indent=4)