BatchalignHK 0.7.22.post34__tar.gz → 0.7.23.post1__tar.gz

{batchalignhk-0.7.22.post34 → batchalignhk-0.7.23.post1}/BatchalignHK.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.1
 Name: BatchalignHK
-Version: 0.7.22.post34
+Version: 0.7.23.post1
 Summary: Python Speech Language Sample Analysis
 Author: Brian MacWhinney, Houjun Liu
 Author-email: macw@cmu.edu, houjun@cmu.edu
@@ -54,6 +54,7 @@ Requires-Dist: pyannote.audio
 Requires-Dist: onnxruntime
 Requires-Dist: certifi>=2025.10.5
 Requires-Dist: regex
+Requires-Dist: pycantonese
 Provides-Extra: dev
 Requires-Dist: pytest; extra == "dev"
 Provides-Extra: train
@@ -61,15 +62,6 @@ Requires-Dist: accelerate~=0.27; extra == "train"
 Provides-Extra: docs
 Requires-Dist: mkdocs-material; extra == "docs"
 Requires-Dist: mkdocs-click; extra == "docs"
-Dynamic: author
-Dynamic: author-email
-Dynamic: classifier
-Dynamic: description
-Dynamic: description-content-type
-Dynamic: license-file
-Dynamic: provides-extra
-Dynamic: requires-dist
-Dynamic: summary
 
 # TalkBank | Batchalign2
 

{batchalignhk-0.7.22.post34 → batchalignhk-0.7.23.post1}/BatchalignHK.egg-info/SOURCES.txt

@@ -128,6 +128,7 @@ batchalign/pipelines/diarization/pyannote.py
 batchalign/pipelines/fa/__init__.py
 batchalign/pipelines/fa/iic_fa.py
 batchalign/pipelines/fa/wave2vec_fa.py
+batchalign/pipelines/fa/wave2vec_fa_canto.py
 batchalign/pipelines/fa/whisper_fa.py
 batchalign/pipelines/morphosyntax/__init__.py
 batchalign/pipelines/morphosyntax/coref.py

{batchalignhk-0.7.22.post34 → batchalignhk-0.7.23.post1}/BatchalignHK.egg-info/requires.txt

@@ -44,6 +44,7 @@ pyannote.audio
 onnxruntime
 certifi>=2025.10.5
 regex
+pycantonese
 
 [dev]
 pytest

{batchalignhk-0.7.22.post34 → batchalignhk-0.7.23.post1}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.1
 Name: BatchalignHK
-Version: 0.7.22.post34
+Version: 0.7.23.post1
 Summary: Python Speech Language Sample Analysis
 Author: Brian MacWhinney, Houjun Liu
 Author-email: macw@cmu.edu, houjun@cmu.edu
@@ -54,6 +54,7 @@ Requires-Dist: pyannote.audio
 Requires-Dist: onnxruntime
 Requires-Dist: certifi>=2025.10.5
 Requires-Dist: regex
+Requires-Dist: pycantonese
 Provides-Extra: dev
 Requires-Dist: pytest; extra == "dev"
 Provides-Extra: train
@@ -61,15 +62,6 @@ Requires-Dist: accelerate~=0.27; extra == "train"
 Provides-Extra: docs
 Requires-Dist: mkdocs-material; extra == "docs"
 Requires-Dist: mkdocs-click; extra == "docs"
-Dynamic: author
-Dynamic: author-email
-Dynamic: classifier
-Dynamic: description
-Dynamic: description-content-type
-Dynamic: license-file
-Dynamic: provides-extra
-Dynamic: requires-dist
-Dynamic: summary
 
 # TalkBank | Batchalign2
 

{batchalignhk-0.7.22.post34 → batchalignhk-0.7.23.post1}/batchalign/cli/cli.py

@@ -111,6 +111,7 @@ batchalign.add_command(train, "models")
 @click.option("--wav2vec/--whisper_fa",
               default=True, help="Use Whisper instead of Wav2Vec for English (defaults for Whisper for non-English)")
 @click.option("--iic", is_flag=True, default=False, help="Use IIC forced alignment (for Chinese).")
+@click.option("--wav2vec_yue", is_flag=True, default=False, help="Use Wav2Vec with chantonese fixes forced alignment (for Chinese).")
 @click.option("--tencent/--rev",
               default=False, help="Use Tencent instead of Rev.AI (default).")
 @click.option("--funaudio/--rev",
@@ -119,7 +120,7 @@ batchalign.add_command(train, "models")
 @click.option("--wor/--nowor",
               default=True, help="Should we write word level alignment line? Default to yes.")
 @click.pass_context
-def align(ctx, in_dir, out_dir, whisper, wav2vec, iic, tencent, funaudio, **kwargs):
+def align(ctx, in_dir, out_dir, whisper, wav2vec, iic, wav2vec_yue, tencent, funaudio, **kwargs):
     """Align transcripts against corresponding media files."""
     def loader(file):
         return (
@@ -128,11 +129,13 @@ def align(ctx, in_dir, out_dir, whisper, wav2vec, iic, tencent, funaudio, **kwar
         )
 
     def writer(doc, output):
-        CHATFile(doc=doc).write(output, write_wor=kwargs.get("wor", True))
+        CHATFile(doc=doc).write(output, write_wor=kwargs.get("wor", True), dont_merge_letters=True)
 
     # Determine FA engine
     if iic:
         fa_engine = "iic_fa"
+    elif wav2vec_yue:
+        fa_engine = "wav2vec_fa_canto"
     elif not wav2vec:
         fa_engine = "whisper_fa"
     else:

{batchalignhk-0.7.22.post34 → batchalignhk-0.7.23.post1}/batchalign/formats/chat/file.py

@@ -93,7 +93,7 @@ class CHATFile(BaseFormat):
             self.__doc = doc
 
 
-    def write(self, path, write_wor=True):
+    def write(self, path, write_wor=True, dont_merge_letters=False):
         """Write the CHATFile to file.
 
         Parameters
@@ -102,13 +102,13 @@ class CHATFile(BaseFormat):
             Path of where the CHAT file should get str.
         """
         
-        str_doc = self.__generate(self.__doc, self.__special_mor, write_wor=write_wor)
+        str_doc = self.__generate(self.__doc, self.__special_mor, write_wor=write_wor, dont_merge_letters=dont_merge_letters)
 
         with open(path, 'w', encoding="utf-8") as df:
             df.write(str_doc)
 
     @staticmethod
-    def __generate(doc:Document, special=False, write_wor=True):
+    def __generate(doc:Document, special=False, write_wor=True, dont_merge_letters=False):
         utterances = doc.content
 
         def __get_birthdays(line):
@@ -130,7 +130,7 @@ class CHATFile(BaseFormat):
                 main.append(generate_chat_utterance(i,
                                                     special and doc.langs[0] == "eng",
                                                     write_wor=write_wor,
-                                                    merge_letters="yue" in doc.langs,
+                                                    merge_letters=("yue" in doc.langs and (not dont_merge_letters)),
                                                     merge_abbrev="eng" in doc.langs))
         main.append("@End\n")
 

{batchalignhk-0.7.22.post34 → batchalignhk-0.7.23.post1}/batchalign/pipelines/__init__.py

@@ -7,7 +7,7 @@ from .morphosyntax import StanzaEngine, CorefEngine
 from .cleanup import NgramRetraceEngine, DisfluencyReplacementEngine
 from .speaker import NemoSpeakerEngine
 
-from .fa import WhisperFAEngine, Wave2VecFAEngine, IICFAEngine
+from .fa import WhisperFAEngine, Wave2VecFAEngine, IICFAEngine, Wave2VecFAEngineCantonese
 from .utr import WhisperUTREngine, RevUTREngine, TencentUTREngine, FunAudioUTREngine
 
 from .analysis import EvaluationEngine

{batchalignhk-0.7.22.post34 → batchalignhk-0.7.23.post1}/batchalign/pipelines/dispatch.py

@@ -9,7 +9,7 @@ from batchalign import (WhisperEngine, WhisperFAEngine, StanzaEngine, RevEngine,
                         StanzaUtteranceEngine, CorefEngine, Wave2VecFAEngine, TencentEngine,
                         OAIWhisperEngine, TencentUTREngine, AliyunEngine, FunAudioEngine,
                         FunAudioUTREngine, SeamlessTranslationModel, GoogleTranslateEngine,
-                        OAIWhisperEngine, PyannoteEngine, IICFAEngine)
+                        OAIWhisperEngine, PyannoteEngine, IICFAEngine, Wave2VecFAEngineCantonese)
 
 from batchalign import BatchalignPipeline
 from batchalign.models import resolve
@@ -135,6 +135,8 @@ def dispatch_pipeline(pkg_str, lang, num_speakers=None, **arg_overrides):
             engines.append(CorefEngine())
         elif engine == "wav2vec_fa":
             engines.append(Wave2VecFAEngine())
+        elif engine == "wav2vec_fa_canto":
+            engines.append(Wave2VecFAEngineCantonese())
         elif engine == "iic_fa":
             engines.append(IICFAEngine())
         elif engine == "seamless_translate":

{batchalignhk-0.7.22.post34 → batchalignhk-0.7.23.post1}/batchalign/pipelines/fa/__init__.py

@@ -1,3 +1,4 @@
 from .whisper_fa import WhisperFAEngine
 from .wave2vec_fa import Wave2VecFAEngine
+from .wave2vec_fa_canto import Wave2VecFAEngineCantonese
 from .iic_fa import IICFAEngine

batchalignhk-0.7.23.post1/batchalign/pipelines/fa/wave2vec_fa_canto.py

@@ -0,0 +1,185 @@
+from batchalign.models import Wave2VecFAModel
+from batchalign.document import *
+from batchalign.pipelines.base import *
+from batchalign.utils import *
+from batchalign.utils.dp import *
+from batchalign.constants import *
+
+import pycantonese as pc
+
+import logging
+L = logging.getLogger("batchalign")
+
+import re
+
+import pycountry
+import warnings
+
+
+
+class Wave2VecFAEngineCantonese(BatchalignEngine):
+    tasks = [ Task.FORCED_ALIGNMENT ]
+
+    @staticmethod
+    def cantonese_to_mms_chars(text: str) -> str:
+        pairs = pc.characters_to_jyutping(text)
+
+        try:
+            jyut = " ".join(j for _, j in pairs)
+        except TypeError:
+            return text
+        jyut = re.sub(r"[1-6]", "", jyut)
+        jyut = re.sub(r"\s+", "'", jyut).strip("'")
+        return jyut
+
+    def _hook_status(self, status_hook):
+        self.status_hook = status_hook
+
+    def __init__(self):
+        self.status_hook = None
+        self.__wav2vec = Wave2VecFAModel()
+
+    def process(self, doc:Document, **kwargs):
+        # 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}'"
+        assert "yue" in doc.langs, "Please use normal wav2vec to align non-cantonese speech."
+
+        # load the audio file
+        L.debug(f"Wave2Vec FA is loading url {doc.media.url}...")
+        f = self.__wav2vec.load(doc.media.url)
+        L.debug(f"Wav2Vec FA finished loading media.")
+
+        # collect utterances 30 secondish segments to be aligned for whisper
+        # we have to do this because whisper does poorly with very short segments
+        groups = []
+        group = []
+        seg_start = 0
+
+        L.debug(f"Wav2Vec FA finished loading media.")
+
+        for i in doc.content:
+            if not isinstance(i, Utterance):
+                continue
+            if i.alignment is None:
+                warnings.warn("We found at least one utterance without utterance-level alignment; this is usually not an issue, but if the entire transcript is unaligned, it means that utterance level timing recovery (which is fuzzy using ASR) failed due to the audio clarity. On this transcript, before running forced-alignment, please supply utterance-level links.")
+                continue
+
+            # pop the previous group onto the stack
+            if (i.alignment[-1] - seg_start) > 15*1000:
+                groups.append(group)
+                group = []
+                seg_start = i.alignment[0]
+
+            # append the contents to the running group
+            for word in i.content:
+                group.append((word, i.alignment))
+
+        groups.append(group)
+
+        L.debug(f"Begin Wav2Vec Inference...")
+
+        for indx, grp in enumerate(groups):
+            L.info(f"Wave2Vec FA processing segment {indx+1}/{len(groups)}...")
+            if self.status_hook != None:
+                self.status_hook(indx+1, len(groups))
+
+            # perform alignment
+            # we take a 2 second buffer in each direction
+            try:
+                transcript = [word[0].text for word in grp]
+                # replace ANY punctuation
+                for p in MOR_PUNCT + ENDING_PUNCT:
+                    transcript = [i.replace("_", " ") for i in transcript if i.strip() != p]
+
+                transcript_mms = {r:self.cantonese_to_mms_chars(r) for r in transcript}
+                transcript_mms_rev = {v:k for k,v in transcript_mms.items()}
+                # if "noone's" in detokenized:
+                    # breakpoint()
+                res = self.__wav2vec(
+                    audio=f.chunk(grp[0][1][0], grp[-1][1][1]),
+                    text=[transcript_mms[r] for r in transcript]
+                )
+            except:
+                # utterance contains nothing
+                continue
+
+            # create reference backplates, which are the word ids to set the timing for
+            ref_targets = []
+            for indx, (word, _) in enumerate(grp):
+                for char in word.text:
+                    ref_targets.append(ReferenceTarget(char, payload=indx))
+            # create target backplates for the timings
+            payload_targets = []
+            timings = []
+            try:
+                for indx, (word, time) in enumerate(res):
+                    timings.append(time)
+                    for char in transcript_mms_rev[word]:
+                        payload_targets.append(PayloadTarget(char, payload=indx))
+            except:
+                continue
+            # alignment!
+            alignments = align(payload_targets, ref_targets, tqdm=False)
+
+            # set the ids back to the text ids
+            # we do this BACKWARDS because we went to have the first timestamp
+            # we get about a word first
+            alignments.reverse()
+            for indx,elem in enumerate(alignments):
+                if isinstance(elem, Match):
+                    grp[elem.reference_payload][0].time = (int(round((timings[elem.payload][0] +
+                                                                      grp[0][1][0]))),
+                                                           int(round((timings[elem.payload][1] +
+                                                                      grp[0][1][0]))))
+
+        L.debug(f"Correcting text...")
+
+        # we now set the end alignment of each word to the start of the next
+        for doc_ut, ut in enumerate(doc.content):
+            if not isinstance(ut, Utterance):
+                continue
+
+            # correct each word by bumping it forward
+            # and if its not a word we remove the timing
+            for indx, w in enumerate(ut.content):
+                if w.type in [TokenType.PUNCT, TokenType.FEAT, TokenType.ANNOT]:
+                    w.time = None
+                elif indx == len(ut.content)-1 and w.text in ENDING_PUNCT:
+                    w.time = None
+                elif indx != len(ut.content)-1:
+                    # search forward for the next compatible time
+                    tmp = indx+1
+                    while tmp < len(ut.content)-1 and ut.content[tmp].time == None:
+                        tmp += 1
+                    if w.time == None:
+                        continue
+                    if ut.content[tmp].time == None:
+                        # seek forward one utterance to find their start time
+                        next_ut = doc_ut + 1 
+                        while next_ut < len(doc.content)-1 and (not isinstance(doc.content, Utterance) or doc.content[next_ut].alignment == None):
+                            next_ut += 1
+                        if next_ut < len(doc.content) and isinstance(doc.content, Utterance) and doc.content[next_ut].alignment:
+                            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
+
+                    # just in case, bound the time by the utterance derived timings
+                    if ut.alignment and ut.alignment[0] != None:
+                        w.time = (max(w.time[0], ut.alignment[0]), min(w.time[1], ut.alignment[1]))
+                    # if we ended up with timings that don't make sense, drop it
+                    if w.time and w.time[0] >= w.time[1]:
+                        w.time = None
+
+            # clear any built-in timing (i.e. we should use utterance-derived timing)
+            ut.time = None
+            # correct the text 
+            if ut.alignment and ut.text != None:
+                if '\x15' not in ut.text:
+                    ut.text = (ut.text+f" \x15{ut.alignment[0]}_{ut.alignment[1]}\x15").strip()
+                else:
+                    ut.text = re.sub(r"\x15\d+_\d+\x15",
+                                     f"\x15{ut.alignment[0]}_{ut.alignment[1]}\x15", ut.text).strip()
+            elif ut.text != None:
+                ut.text = re.sub(r"\x15\d+_\d+\x15", f"", ut.text).strip()
+
+        return doc

{batchalignhk-0.7.22.post34 → batchalignhk-0.7.23.post1}/batchalign/pipelines/utr/tencent_utr.py

@@ -80,7 +80,8 @@ class TencentUTREngine(BatchalignEngine):
         self.__client = AsrClient(cred, "ap-hongkong")
 
 
-    def replace_cantonese_words(self, word):
+    @staticmethod
+    def replace_cantonese_words(word):
         """Function to replace Cantonese words with custom replacements."""
         word_replacements = {
             "系": "係", 
@@ -194,7 +195,7 @@ class TencentUTREngine(BatchalignEngine):
             roman_cache_end = i.StartMs
             for j in i.Words:
                 word = j.Word
-                if self.__lang == "yue":
+                if lang == "yue":
                     word = cc.convert(word)
 
                     word = self.replace_cantonese_words(word)

batchalignhk-0.7.23.post1/batchalign/version

@@ -0,0 +1,3 @@
+0.7.23-post.1
+Janurary 4th, 2026
+Cantonese FA

{batchalignhk-0.7.22.post34 → batchalignhk-0.7.23.post1}/setup.py

@@ -75,7 +75,8 @@ setup(
         "pyannote.audio",
         "onnxruntime",
         "certifi>=2025.10.5",
-        "regex"
+        "regex",
+        "pycantonese"
     ],
     extras_require={
         'dev': [

batchalignhk-0.7.22.post34/batchalign/version

@@ -1,3 +0,0 @@
-0.7.22-post.34
-December 30th, 2025
-CA