Problem/error with Vietnamese datasets? #961
Replies: 3 comments 15 replies
-
|
Sometimes certain models can go crazy with certain content, I would try another model and see if it makes the same errors |
Beta Was this translation helpful? Give feedback.
-
|
I've trial and error'd multiple models, and the "largest" one that seems to work fine without anomalies like these is base. The results that it gives though is extremely subpar and not at all suitable for actual usage. |
Beta Was this translation helpful? Give feedback.
-
|
it isn't a problem specific to vietnamese, but rather a classic hallucination problem, see #679 |
Beta Was this translation helpful? Give feedback.
-
|
Possibly. Though, this happens with any audio input, and this is the only output it gives (that ad line over and over again), so I think there's also a bit of a problem with the dataset. Have you tried out the proposed fix in the aforementioned issue? |
Beta Was this translation helpful? Give feedback.
-
|
Maybe but this problem seem like happen in less common languages, especially with Vietnamese because the data available is not so much. |
Beta Was this translation helpful? Give feedback.
-
|
@DavidScann certainly it's a problem with dataset, unfortunately we can't do much about it @goctoidongtien it happens to english, many people reported having "music in background" or many times "thanks for watching!" if you read more closely the thread i mentioned |
Beta Was this translation helpful? Give feedback.
-
|
this is an example of combining different techniques to prevent hallucinations: https://github.com/EtienneAb3d/WhisperHallu |
Beta Was this translation helpful? Give feedback.
-
|
Can confirm, I had the similar error in Indonesian with the line "Thanks for watching" in Indonesian. |
Beta Was this translation helpful? Give feedback.
-
Beta Was this translation helpful? Give feedback.
-
|
to answer simply: unless u change model, to improve accuracy, u can train longer and throw more data into it |
Beta Was this translation helpful? Give feedback.
-
|
the thing is nobody know what whisper training curve look like, hard to say whether it's under-/over-fitted but with deep learning u should worry less about overfitting, more about data (quality + quantity) small model maybe sufficient for single language yes, but u know how much more there's english speech vs vietnamese speech in training data ? 400k h vs 691 h ; well if u really have that much data for vietnamese better training from scratch instead of fine tune |
Beta Was this translation helpful? Give feedback.
-
|
u have to request access to vlsp dataset |
Beta Was this translation helpful? Give feedback.
-
|
wait u dont have any leadership/planning yet ? i dont have any AI/ML real world experience yet, i can participate but i dont think im able to lead |
Beta Was this translation helpful? Give feedback.
-
|
u can see email on my profile |
Beta Was this translation helpful? Give feedback.
-
Hello, I'm just started to use Whisper via FreeSubtitles.ai and the result has some problem.
A large part of the resulting subtitle file returned to me is not subtitle lines, but a call for viewers to subscribe to a YouTube channel. The line in the picture means "Please subscribe to the channel Ghien Mi Go so you don't miss interesting videos." Tried many different models but same result.
I assume this is an issue related to the data source imported into the model from the start. Many subtitles on Vietnamese YouTube videos are not actually subtitles. It's just dummy subtitles, with a single line calling viewers to subscribe to the youtube channel with timecode from start to finish. I think OpenAI should double check its multilingual data source, because it greatly affects the results. Hope it gets fixed soon.
Beta Was this translation helpful? Give feedback.
All reactions