Better WER calculation for compounds and clitics

montreal_forced_aligner/dictionary/mixins.py

@@ -338,6 +338,14 @@ def base_phones(self) -> Dict[str, Set[str]]:
 
         return base_phones
 
+    @property
+    def split_regex(self) -> re.Pattern:
+        """Pattern for splitting arbitrary text"""
+        markers = self.compound_markers
+        if "-" in markers:
+            markers = ["-"] + [x for x in self.compound_markers if x != "-"]
+        return re.compile(rf'[{"".join(markers)} ]')
+
     @property
     def extra_questions_mapping(self) -> Dict[str, List[str]]:
         """Mapping of extra questions for the given phone set type"""

montreal_forced_aligner/transcription/transcriber.py

@@ -9,6 +9,7 @@
 import itertools
 import multiprocessing as mp
 import os
+import re
 import shutil
 import subprocess
 import sys
@@ -1347,7 +1348,8 @@ def evaluate(self):
             utt_name = utterance.name
             if not utterance.text:
                 continue
-            g = utterance.text.split()
+
+            g = self.split_regex.split(utterance.text)
 
             total_count += 1
             total_word_length += len(g)
@@ -1416,6 +1418,12 @@ def evaluate(self):
 
     def _load_transcripts(self):
         """Load transcripts from Kaldi temporary files"""
+        initial_clitics = {
+            x for x in self.clitic_set if re.match(rf"^.*[{''.join(self.clitic_markers)}]$", x)
+        }
+        final_clitics = {
+            x for x in self.clitic_set if re.match(rf"^[{''.join(self.clitic_markers)}].*$", x)
+        }
         for score_args in self.score_arguments():
             for tra_path in score_args.tra_paths.values():
 
@@ -1431,7 +1439,13 @@ def _load_transcripts(self):
                             continue
                         transcription = []
                         for i in ints:
-                            transcription.append(lookup[int(i)])
+                            w = lookup[int(i)]
+                            if len(transcription) and (
+                                w in final_clitics or transcription[-1] in initial_clitics
+                            ):
+                                transcription[-1] += w
+                                continue
+                            transcription.append(w)
                         utterance.transcription_text = " ".join(transcription)
 
     def export_files(self, output_directory: str) -> None: