OpenMOSS-Team

/

MOSS-Transcribe-Diarize

like 363

Follow

OpenMOSS 1.12k

Hi everyone,

First off, thank you so much for the amazing model! I've been testing the BF16 version for Japanese, and the speed and basic grammar are honestly really impressive.

I did notice a few areas where the model gets a bit confused with context, so I just wanted to share some examples in case it helps with training the next version.

Here is what I'm seeing:

• Words that sound the same: It sometimes picks the wrong kanji for words that have the exact same pronunciation. For example, it wrote "一情報" (Number 1 data) instead of "位置情報" (Location data).
• Vowel lengths: It sometimes misses elongated vowels, which changes the word entirely. For example, it heard "Teki" (enemy) instead of "Teiki" (commuter pass), which completely throws off the context of the sentence.
• Fast, casual speech: When people talk really fast or argue, it sometimes mishears the sounds and pieces together sentences that don't make sense. In one clip, it heard "watashi no kawanai de uttenai no" instead of "watashi ga kawanai to ikenai no" (Why do I have to be the one to buy it?).

Basically, the core sentence building is awesome! It just seems to lack a bit of contextual awareness to pick the right words when things are spoken quickly or sound similar.

Hope this is helpful! Thanks again for all the hard work you put into building all these amazing Multimodal models 🙌.