make sure the leading and tailing word boundary exists. #1019
Conversation
…on is absorbed into the word boundary symbol.
|
Hi @lesca I am glad to hear that. Thank you for your feedback. |
| if model_lang not in LANGUAGES_WITHOUT_SPACES: | ||
| if not text.startswith(" "): | ||
| text = " " + text | ||
| if not text.endswith(" "): | ||
| text = text + " " |
There was a problem hiding this comment.
Won't this result in double spaces?
There was a problem hiding this comment.
Hi @Barabazs
No, it won't. If there's already a leading or trailing space, no additional space will be added. So no need to worry about double spaces.
| @@ -173,8 +180,8 @@ def align( | |||
| elif char_ in model_dictionary.keys(): | |||
| clean_char.append(char_) | |||
| clean_cdx.append(cdx) | |||
| else: | |||
| # add placeholder | |||
| elif char_ not in '!"\',.:;<=>?[\\]^_`{|}': | |||
There was a problem hiding this comment.
You mention the word boundary symbol from Wav2Vec2, but the page you linked only uses |. Can you explain why these other symbols are needed?
And imo an else-clause should be added as a fallback in case of unforeseen issues.
There was a problem hiding this comment.
In the original code, symbols beyond "a-z" were ignored, and word boundaries were added to the string for alignment purposes. However, some of these ignored symbols (like numbers or the dollar sign $) may have pronunciations and require timestamps. To address this, a wildcard placeholder is needed to ensure these symbols are properly processed.
On the other hand, certain symbols (like !"',.:;<=>?[]^_`{|}) can be absorbed into word boundaries without needing a wildcard. The condition in the code filters out these absorbable symbols and only adds wildcard placeholders for symbols that have explicit pronunciations.
As for the else-clause, it simply skips the symbol, so it’s not explicitly implemented in the current logic.
Let me know if you’d like further clarification or adjustments.
Fix #1016
On the wav2vec alignment tutorial page, word boundary symbols are present at the beginning and end. During the whisperX alignment process, the tailing word boundary might be missing, causing the tailing silence to be absorbed into the last word. Adding the tailing word boundary would resolve this issue.