batchalign 0.7.11a0__tar.gz → 0.7.11b3__tar.gz

{batchalign-0.7.11a0/batchalign.egg-info → batchalign-0.7.11b3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: batchalign
-Version: 0.7.11a0
+Version: 0.7.11b3
 Summary: Python Speech Language Sample Analysis
 Author: Brian MacWhinney, Houjun Liu
 Author-email: macw@cmu.edu, houjun@cmu.edu

{batchalign-0.7.11a0 → batchalign-0.7.11b3}/batchalign/cli/cli.py

@@ -107,12 +107,16 @@ batchalign.add_command(train, "models")
 @common_options
 @click.option("--whisper/--rev",
               default=False, help="For utterance timing recovery, OpenAI Whisper (ASR) instead of Rev.AI (default).")
+@click.option("--pauses", type=bool, default=False, help="Should we try to bullet each word or should we try to add pauses in between words by grouping them? Default: no pauses.", is_flag=True)
 
 @click.pass_context
 def align(ctx, in_dir, out_dir, whisper, **kwargs):
     """Align transcripts against corresponding media files."""
     def loader(file):
-        return CHATFile(path=os.path.abspath(file)).doc
+        return (
+            CHATFile(path=os.path.abspath(file)).doc,
+            {"pauses": kwargs.get("pauses", False)}
+        )
 
     def writer(doc, output):
         CHATFile(doc=doc).write(output)

{batchalign-0.7.11a0 → batchalign-0.7.11b3}/batchalign/models/whisper/infer_fa.py

@@ -79,10 +79,10 @@ class WhisperFAModel(object):
         # and return the audio file
         return ASRAudioFile(f, resampled, self.sample_rate)
 
-    def __call__(self, audio, text):
+    def __call__(self, audio, text, pauses=False):
         L.debug("Whisper Preprocessing...")
         # input features
-        features = self.__processor(audio=audio, text=" ".join(list(text)),
+        features = self.__processor(audio=audio, text=(" ".join(list(text)) if pauses else text),
                                     sampling_rate=self.sample_rate,
                                     return_tensors='pt')
         tokens = features["labels"][0]

{batchalign-0.7.11a0 → batchalign-0.7.11b3}/batchalign/pipelines/fa/whisper_fa.py

@@ -29,6 +29,9 @@ class WhisperFAEngine(BatchalignEngine):
         self.__whisper = WhisperFAModel(model)
 
     def process(self, doc:Document, **kwargs):
+        # check if pauses
+        pauses = kwargs.get("pauses", False)
+
         # check that the document has a media path to align to
         assert doc.media != None and doc.media.url != None, f"We cannot forced-align something that doesn't have a media path! Provided media tier='{doc.media}'"
 
@@ -83,7 +86,7 @@ class WhisperFAEngine(BatchalignEngine):
                 # if "noone's" in detokenized:
                     # breakpoint()
                 res = self.__whisper(audio=f.chunk(grp[0][1][0], grp[-1][1][1]),
-                                     text=detokenized)
+                                     text=detokenized, pauses=pauses)
             except IndexError:
                 # utterance contains nothing
                 continue
@@ -109,19 +112,25 @@ class WhisperFAEngine(BatchalignEngine):
             alignments.reverse()
             for indx,elem in enumerate(alignments):
                 if isinstance(elem, Match):
-                    next_elem = indx - 1 # remember this is backwards, see above
-                    while next_elem >= 0 and alignments[next_elem].payload == elem.payload:
-                        next_elem -= 1
-                    if next_elem < 0:
-                        next_elem = None
+                    if pauses:
+                        next_elem = indx - 1 # remember this is backwards, see above
+                        while next_elem >= 0 and alignments[next_elem].payload == elem.payload:
+                            next_elem -= 1
+                        if next_elem < 0:
+                            next_elem = None
+                        else:
+                            next_elem = alignments[next_elem]
+                        grp[elem.reference_payload][0].time = (int(round((timings[elem.payload]*1000 +
+                                                                        grp[0][1][0]))),
+                                                            int(round((timings[elem.payload]*1000 +
+                                                                        grp[0][1][0])))+500 if next_elem == None else
+                                                            int(round((timings[next_elem.payload]*1000 +
+                                                                        grp[0][1][0]))))
                     else:
-                        next_elem = alignments[next_elem]
-                    grp[elem.reference_payload][0].time = (int(round((timings[elem.payload]*1000 +
-                                                                      grp[0][1][0]))),
-                                                           int(round((timings[elem.payload]*1000 +
-                                                                      grp[0][1][0])))+500 if next_elem == None else
-                                                           int(round((timings[next_elem.payload]*1000 +
-                                                                      grp[0][1][0]))))
+                        grp[elem.reference_payload][0].time = (int(round((timings[elem.payload]*1000 +
+                                                                          grp[0][1][0]))),
+                                                               int(round((timings[elem.payload]*1000 +
+                                                                          grp[0][1][0]))))
 
         L.debug(f"Correcting text...")
 
@@ -153,8 +162,8 @@ class WhisperFAEngine(BatchalignEngine):
                             w.time = (w.time[0], doc.content[next_ut].alignment[0])
                         else:
                             w.time = (w.time[0], w.time[0]+500) # give half a second because we don't know
-                    # else:
-                    #     w.time = (w.time[0], ut.content[tmp].time[0])
+                    elif not pauses:
+                        w.time = (w.time[0], ut.content[tmp].time[0])
 
                     # just in case, bound the time by the utterance derived timings
                     if ut.alignment and ut.alignment[0] != None:

batchalign-0.7.11b3/batchalign/version

@@ -0,0 +1,3 @@
+0.7.11-beta.3
+Feburary 2nd, 2025
+Incorporate additional pauses

{batchalign-0.7.11a0 → batchalign-0.7.11b3/batchalign.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: batchalign
-Version: 0.7.11a0
+Version: 0.7.11b3
 Summary: Python Speech Language Sample Analysis
 Author: Brian MacWhinney, Houjun Liu
 Author-email: macw@cmu.edu, houjun@cmu.edu

batchalign-0.7.11a0/batchalign/version

@@ -1,3 +0,0 @@
-0.7.11-alpha.0
-Janurary 22th, 2025
-Whisper inference with spaces