Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Return result with best log prob when all temperature fallbacks failed #356

Merged
merged 7 commits into from
Jul 20, 2023
Merged

Return result with best log prob when all temperature fallbacks failed #356

merged 7 commits into from
Jul 20, 2023

Conversation

hoonlight
Copy link
Contributor

@hoonlight hoonlight commented Jul 17, 2023 •
edited
Loading

Related to openai/whisper#1377

@hoonlight
Copy link
Contributor Author

hoonlight commented Jul 19, 2023 •
edited
Loading

The improvements suggested in openai/whisper#1377 (comment) have been applied.

I'm still testing, but there seems to be a noticeable reduction in hallucinations and an overall improvement in transcription quality.

guillaumekln
guillaumekln reviewed Jul 19, 2023
View reviewed changes
Copy link
Contributor

@guillaumekln guillaumekln left a comment •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Since this PR only changes the temperature fallback final result, which would be very different than openai/whisper anyway, we could merge it before it is accepted in openai/whisper.

Here are some comments:

faster_whisper/transcribe.py Outdated Show resolved Hide resolved
faster_whisper/transcribe.py Outdated Show resolved Hide resolved
faster_whisper/transcribe.py Outdated Show resolved Hide resolved
@hoonlight
Copy link
Contributor Author

Thank you, I have changed the code according to your advice.

guillaumekln
guillaumekln reviewed Jul 19, 2023
View reviewed changes
faster_whisper/transcribe.py Outdated Show resolved Hide resolved
faster_whisper/transcribe.py Outdated Show resolved Hide resolved
@hoonlight
Copy link
Contributor Author

hoonlight commented Jul 19, 2023 •
edited
Loading

Please review again. Also, I think it would be nice if we renamed the PR to be more intuitive, please rename it to something you think is appropriate.

@hoonlight
Copy link
Contributor Author

hoonlight commented Jul 20, 2023 •
edited
Loading

There seems to be another benefit to this change.

Before this change, there were quite a few cases where a sentence was not bad even if it had avg_logprob < -1.
After this change, along with an overall improvement in transcription quality, fewer sentences with avg_logprob < -1 were output, and sentences with avg_logprob < -1 were much more likely to be hallucinations.

So I was able to filter out the segments with avg_logprob < -1.5 and further reduce the hallucinations.
For users who want to reduce hallucinations at the cost of missing some sentences, this might be helpful.

Of course I haven't tested all cases, so there may be many different cases and this won't always guarantee better results.

guillaumekln
guillaumekln approved these changes Jul 20, 2023
View reviewed changes
Copy link
Contributor

@guillaumekln guillaumekln left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks good to me. Thanks!

@guillaumekln guillaumekln changed the title Resolve Inference Selection Bug Return result with best log prob when all temperature fallbacks failed Jul 20, 2023
@guillaumekln guillaumekln merged commit e786e26 into SYSTRAN:master Jul 20, 2023
3 checks passed
@hoonlight hoonlight deleted the return-best-result branch July 20, 2023 14:23
@Purfview Purfview mentioned this pull request Dec 5, 2023
Merged
nguyendc-systran pushed a commit that referenced this pull request Dec 13, 2023
@Purfview
Fix broken prompt_reset_on_temperature (#604)
ebcfd6b 
* Fix broken prompt_reset_on_temperature

Fixing: #603

Broken because `generate_with_fallback()` doesn't return final temperature.

Regression since PR356 -> #356
@TheoBoyer TheoBoyer mentioned this pull request Dec 18, 2023
Open
metame-none pushed a commit to metame-ai/faster-distil-whisper that referenced this pull request Jan 2, 2024
@Purfview @metame-none
Fix broken prompt_reset_on_temperature (SYSTRAN#604)
18e53ad 
* Fix broken prompt_reset_on_temperature

Fixing: SYSTRAN#603

Broken because `generate_with_fallback()` doesn't return final temperature.

Regression since PR356 -> SYSTRAN#356
Aiurus added a commit to Aiurus/faster-whisper that referenced this pull request May 25, 2024
@Aiurus
Fix broken prompt_reset_on_temperature (#604)
4d7db56 
* Fix broken prompt_reset_on_temperature

Fixing: SYSTRAN/faster-whisper#603

Broken because `generate_with_fallback()` doesn't return final temperature.

Regression since PR356 -> SYSTRAN/faster-whisper#356
coolCatalyst added a commit to coolCatalyst/faster-whisper that referenced this pull request Jun 1, 2024
@coolCatalyst
Fix broken prompt_reset_on_temperature (#604)
a7eb543 
* Fix broken prompt_reset_on_temperature

Fixing: SYSTRAN/faster-whisper#603

Broken because `generate_with_fallback()` doesn't return final temperature.

Regression since PR356 -> SYSTRAN/faster-whisper#356
AIXerum added a commit to AIXerum/faster-whisper that referenced this pull request Oct 18, 2024
@AIXerum
Fix broken prompt_reset_on_temperature (#604)
57cbea9 
* Fix broken prompt_reset_on_temperature

Fixing: SYSTRAN/faster-whisper#603

Broken because `generate_with_fallback()` doesn't return final temperature.

Regression since PR356 -> SYSTRAN/faster-whisper#356
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@guillaumekln guillaumekln guillaumekln approved these changes

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@hoonlight @guillaumekln