Possible bug with the Amazon Transcribe API

Question

We believe we found an inconsistency in output for batch mode when transcribing chinese:

- Items in `results.items` array should contain objects of type "pronunciation" or "punctuation".
- However, at least for zh_CN, we find "pronunciation" types sometimes including punctuation marks!

Example:

The transcription in full read: "羊猫狗猪鸡。蛇、牛、长颈鹿、大象、鲨鱼、猴子、熊。猎豹、狮子老虎。一。虫甲壳虫。七星瓢虫。蝴蝶蜻蜓于。海豹,海狮、海龟。小鸟。喜鹊、麻雀、朱环、剃壶。蝉唐王。蚊子苍蝇。菲尔谢弗。苍山甲野猪。".

When mapping the array to exclusively the contents of the "pronunciation" items, the array looks like this:
[ "羊猫", "狗", "猪", "鸡。", "蛇、", "牛、", "长颈鹿、", "大象、", "鲨鱼、", "猴子、", "熊。", "猎豹、", "狮子", "老虎。", "一。", "虫", "甲壳虫。", "七星", "瓢虫。", "蝴蝶", "蜻蜓", "于。", "海豹", "海狮、", "海龟。", "小鸟。", "喜鹊、", "麻雀、", "朱环、", "剃壶。", "蝉", "唐王。", "蚊子", "苍蝇。", "菲尔", "谢弗。", "苍山甲", "野猪。" ]

Accuracy of the transcription aside, these include punctuations such as "。" and "、"! Pronunciations should not contain punctuactions such as commas and periods. I would understand if punctuation such as "don't" were included, but this is not the case in point, since these are commas and periods.

Please revise if there is anything to be corrected on your end regarding this inconsistency on this or any other languages. We suspect both zh_CN and zh_HK are affected by this inconsistency, but we have not checked others.

Thank you for your time.

re:Post Agent · Answer

You've identified an interesting behavior in Amazon Transcribe's output for Chinese (zh_CN) transcriptions. In standard Amazon Transcribe output, items are typically categorized as either "pronunciation" or "punctuation" types, with punctuation marks like periods and commas being separated into their own "punctuation" type items.

However, based on your example, it appears that for Chinese transcriptions, punctuation marks like "。" (Chinese period) and "、" (enumeration comma) are sometimes included within "pronunciation" type items rather than being separated into distinct "punctuation" type items.

This behavior differs from the expected pattern seen in English transcriptions, where punctuation is typically separated. The documentation examples for English transcriptions show punctuation marks like periods being properly categorized as "punctuation" type items, separate from the words.

This inconsistency in how punctuation is handled between different languages could indeed be considered a bug or at least an inconsistency in the API's behavior. Since you've observed this specifically with Chinese (zh_CN and potentially zh_HK), it may be a language-specific implementation issue.

I would recommend reporting this to AWS Support so they can investigate this behavior further, as it appears to be inconsistent with how punctuation is typically handled in the transcription results for other languages.  
**Sources**  
[Enabling speaker partitioning in batch transcriptions - Amazon Transcribe](https://docs.aws.amazon.com/transcribe/latest/dg/conversation-diarization-batch-med.html)  
[Transcribing an audio file of a medical conversation - Amazon Transcribe](https://docs.aws.amazon.com/transcribe/latest/dg/batch-medical-conversation.html)  
[Transcribing an audio file of a medical dictation - Amazon Transcribe](https://docs.aws.amazon.com/transcribe/latest/dg/batch-medical-dictation.html)

Possible bug with the Amazon Transcribe API

Add your answer

Relevant content