Medium size model (English Only) generates random timeline, and no subtitle in the player

[yiqiaowang-arch]

Describe the bug As described in the title, Memo AI v1.0.10 behaves unexpected after creating subtitle for a 90-min long video.

• Random timeline after 44:00 (e.g., after a sentence at 45:12, the next one has a timestamp of 00:34, see screenshot and srt example)
• no subtitle appeared in the player after 45:12.
• In the output .srt file, the timestamp is again weird (end time < start time): SRT file example:

  
  393
  00:44:39,220 --> 00:44:43,720
  cost potentials we have already set but also production patterns.

394 00:44:44,000 --> 00:00:03,430 You see something like this here for the Swiss rivers.

395 00:44:48,000 --> 00:00:10,630 Technical operational constraints like ramping rates and so, and policy and taxes.

**To Reproduce**

1. Video with clear human voice in English uploaded
2. Prompt: This is a lecture in Energy System Analysis course in ETH Zurich, specifically, a lecture about bottom-up energy optimization. 
3. video playing while Memo AI transcribing
4. See error

**Expected behavior**
All subtitles should be in order and the next sentence should always have a timestamp later than the previous sentence.

**Screenshots**
![image](https://github.com/Makememo/MemoAI/assets/28997207/f1b524cd-fed2-4140-b65a-e3f6d7a5b049)

**Environment:**
 - OS: Windows 11 Pro
 - Memo AI v1.0.10
 - GPU enabled (RTX 3060 Laptop, memory=6GB)

[Makememo]

Thank you for your feedback. We will check the problem and fix it in the next version.

[yiqiaowang-arch]

Update: after regenerating using prompt: This is a lecture in Energy System Analysis course about bottom-up energy optimization. Each identification waits for the discussion to end instead of breaking in the middle. The timeline issue disappears.

[Makememo]

Update: after regenerating using prompt:

This is a lecture in Energy System Analysis course about bottom-up energy optimization. Each identification waits for the discussion to end instead of breaking in the middle.

The timeline issue disappears.

Can you share the original video with us? Let's find this problem.

[yiqiaowang-arch]

Update: after regenerating using prompt: This is a lecture in Energy System Analysis course about bottom-up energy optimization. Each identification waits for the discussion to end instead of breaking in the middle. The timeline issue disappears.

Can you share the original video with us? Let's find this problem.

Here's the video link in OneDrive: https://1drv.ms/v/s!Ai_5pui2MokRk_IQ2Sjg9EzuCJ6W5Q?e=G3w36g Hope it helps!

[Makememo]

We have made some fix, please see if there is an error timeline. https://memo.ac/releases.html

[yiqiaowang-arch]

Unfortunately I still have wrong timelines. This time I used the prompt: "Mathematical Optimization Course. Please add punctuation." I will try again using my previous prompt and see if this time it works.

[Makememo]

@yiqiaowang-arch Is there a problem opening VAD?

[Makememo]

And, what is the transcribe model?

[yiqiaowang-arch]

After the second try, this time the problem remains.

In my case VAD is not enabled. I am using Medium.en 1.53 GB in all these cases. I will try again using VAD with your VAD settings in the screenshot.

[yiqiaowang-arch]

export_log.zip After VAD enabled I am not able to see the transcripted text anymore. Here's the log file for you to check.

[Makememo]

@yiqiaowang-arch Can you refer to this setting and try again. 0.2 or 0.3

[yiqiaowang-arch]

I have set this value to exactly as shown in your previous screenshot (mode=lenient, threshold=0.3) and it doesn't output anything.

[Makememo]

I have set this value to exactly as shown in your previous screenshot (mode=lenient, threshold=0.3) and it doesn't output anything.

Can you send me log? and try 0.2.

[yiqiaowang-arch]

Here's the log file with VAD threshold=0.3. I'll try 0.2 now.

2023-11-05.log

Update: 0.2 doesn't work as well. Log file attached below: 2023-11-05_03.log

[Makememo]

The new version is being tested and will be released recently. Do you have zoom? I want to see the problem remotely.

Please close VAD first.

[yiqiaowang-arch]

Yes, I am happy to share my experience via zoom. I will be available during the weekends, if that also works for you. I will first try to update to the newest version and see if the problem remains.

[yiqiaowang-arch]

Please contact via my email, in case you want me to show how the software behaves.