[BUG] youtube 字幕与原声不同步

[m4rsquake]

Before you report

• I have searched existing issues to avoid duplicates
• I have tried with the latest available version (or can reproduce on latest)

Current vs. Expected behavior

目前问题：
youtube字幕存在播放着播放着就出现字幕与画面不同步的问题。
起初我以为是youtube原字幕的问题，我关闭插件后测试发现原字幕是单个单词单个单词精准弹出的，所以不存在不同步的问题。
并且我关闭了播放器速度控制，使其在1倍速情况下进行测试，问题仍然存在。

预期行为：
字幕与实际语音保持同步。

To Reproduce

1. 打开任意youtube 英文原声视频，并开启插件。
2. 播放一会视频，起初字幕与原声是同步的。
3. 播放一段时间后出现字幕与原声不同步问题。有概率过几段原声后恢复。

Which area(s) are affected? (Select all that apply)

Website

Extension version (if applicable)

No response

Provide environment information

- Brave Browser
- youtube.com

Additional context

No response

[ananaBMaster]

@frogGuaGuaGuaGua try to analyze this

[frogGuaGuaGuaGua]

I traced the relevant subtitle pipeline. This looks more like cue normalization losing word-level timing than cumulative playback-clock drift:

• The scheduler reads video.currentTime directly and handles seeks/rate changes, so 1× playback should not accumulate a separate timer error.
• YouTube word-timed/scrolling ASR cues are deliberately coalesced for translation: the scrolling-ASR parser accumulates text across events until a sentence/length boundary, while the standard path's optimizer merges consecutive fragments and retains only the first start time and final end time.
• The UI then displays that entire coalesced original/translated fragment from the first word's timestamp. As a result, it can appear ahead of the currently spoken word and line up again at the next chunk boundary. That matches the reported “out of sync, then sometimes recovers” pattern.

I ran the focused parser/optimizer suite on current main: 23 relevant tests passed. The tests also confirm that cross-event and short-line merging is the current intended behavior. This code is already in v1.42.3, and I found no later subtitle-timing fix on main.

This is a likely mechanism, but I cannot confirm that it is the reporter's exact case without the affected track. Could you provide the exact YouTube video URL, the Read Frog version, whether AI subtitle segmentation is enabled, and whether the Read Frog subtitle is ahead of or behind the audio (roughly by how much)? With that URL, we can capture the actual timed-text format and turn it into a regression fixture.

[mengxi-ream]

如果是不太好避免的问题，那应该可以用 语音识别字幕那个新功能解决。就不用管这个了。 @taiiiyang 后面你可以评估一下