Subtitle Refinement

Task Objectives

Clean the provided raw

srt

into a deliverable complete

clean.srt

Do not overwrite the original subtitle file.
Only perform subtitle-level correction, not polishing, summarization, or expansion.
Final verification: Run
```
scripts/check_clean_srt.py
```
directly on the original
```
srt
```
and the complete
```
clean.srt
```
.

Original Requirements

Text Cleaning Requirements

Correct recognition errors, including proper usage of Chinese structural particles "de" (的, 地, 得) and appropriate choice of pronouns "he/she/it" (他, 她, 它).
Delete meaningless filler words such as "en, a, e, ha" (嗯、啊、呃、哈); cleaned subtitles should not contain punctuation.

Add necessary space pauses between clauses. Example:

来那你就先开始

来 那你就先开始

Example:

你是第二个是吧

你是第二个 是吧

Delete obvious repeated words; do not rewrite into formal text, summarize, or expand, only perform subtitle-level cleaning and correction.
Each subtitle entry should not exceed 14 characters (English words are counted as one character by default), no line breaks within a single subtitle entry.
If the original subtitle exceeds 14 characters, split it with accurate sentence segmentation. Example:
```
还得谢谢各位母亲对家里的付出
```
->
```
还得谢谢各位母亲
```
+
```
对家里的付出
```
Do not cause overall advance or delay of subsequent subtitles due to text correction, filler word deletion, repeated word deletion, or subtitle splitting. All adjustments must be limited to the current subtitle entry and its local scope.

Timeline Requirements

Processed subtitles must be strictly synchronized with the original audio, with no global timeline drift.
If only subtitle-level cleaning and correction are performed without changing the semantic boundary of the original sentence, the original start and end time of the entry will remain unchanged by default.
If splitting is required due to sentence segmentation or single-entry exceeding 14 characters, all split sub-entries must fall completely within the original subtitle time range; there should be no overlap between sub-entries, prioritize end-to-end connection, and the total coverage duration must be consistent with the original entry.
Time allocation for split entries should prioritize semantic pauses and speaking rhythm; if precise judgment is not possible, allocate according to the proportion of valid characters in each clause, do not split mechanically evenly.
When deleting filler words or obvious repeated words in a sentence, usually do not adjust the original entry's start and end time; if an entry is an independent and meaningless filler word, it can be directly deleted, but do not stretch the coverage of preceding or following subtitles without basis.
Each subtitle should fit the audio segment as much as possible, with no obvious early display or retention; if fine-tuning is needed, the deviation of the start and end time of a single subtitle relative to the audio boundary should be controlled within ±100 to 150 milliseconds as much as possible.
After splitting and modification, readability must still be ensured, avoiding overly short flashes, obvious missing displays, meaningless gaps, or segmentation that breaks natural semantic boundaries.

Additional Requirements

Do not accidentally delete meaningful subtitle entries.
Double check the timeline.
If the subtitle file is too long, parallel processing is only allowed if the user explicitly requests or permits sub-agents; even if parallel processing is used, treat it only as an editing aid, and the main agent must finally organize a complete
```
clean.srt
```
manually.
The script should not assume that subtitles must be split; it must be compatible with both split and non-split scenarios.
Conservative priority. Unless it is clearly identified as an ASR recognition error, meaningless filler word, obvious repeated slip of the tongue, lack of necessary pause space, or must be split due to exceeding 14 characters, do not rewrite the word order, syntactic structure, collocation relationship of the original sentence, or replace it with a smoother expression.
Do not rewrite the original sentence just because it is "more fluent", "more concise", or "more like formal written language". For example, changing "给我支持" to "支持我", "跟我说" to "对我说", "就是也是" to "也是" are not allowed unless they are clear recognition errors or required by rules.
Allow subtitles to be less formal or less polished; only correct errors, do not polish. If a sentence already meets the rules in its current form, keep it as is.
Strict deletion rules. Independent entries of pure filler words can be directly deleted; in addition, only allow deleting entire entries when they are clearly judged as broken fragments, repeated fragments, or pure noise misrecognition. All non-filler deletions must be organized into an allowlist and listed one by one in the delivery.
If the user specifies a separate word count standard, such as "English words are counted as one character", follow the user's requirements; if the default script standard is inconsistent with the user's requirements, explain the script's limitations and the actual review method used in the delivery.
If the user specifies a separate word count standard, such as "English words are counted as one character", follow the user's requirements; you can add
```
--latin-word-as-one-char
```
when running the verification script.
The default output complete
```
clean.srt
```
prioritizes verifiability and re-listenability. Unless the user explicitly requests renumbering, prioritize retaining the original
```
block_id
```
; use temporary numbers like
```
123a
```
,
```
123b
```
when splitting a single entry.
Do not borrow words from other entries, add words, or rewrite to "smooth out" the current entry. If the current entry is unclear or suspected to be ASR garbled but cannot be clearly corrected based only on this entry, prioritize retaining the original sentence and marking it for manual re-listen, do not subjectively splice new text from preceding or following subtitles.

Output Specifications

The default output file name is in the form
```
xxx.clean.srt
```
.
Do not overwrite the original file.
The final deliverable is a complete
```
clean.srt
```
; do not use "block files to be merged" as the formal product.
If it is necessary to retain the original number after splitting for verification or editing convenience, you can directly use
```
123a
```
,
```
123b
```
in the complete
```
clean.srt
```
; do not rely on an additional merging script to assemble them a second time.

Recommended Process

Read the original
```
srt
```
, confirm the total number of entries, whether there are overlong entries, and whether block processing is needed.
Clarify the word count constraint standard for this task; if there is no additional explanation from the user, check according to the script's default standard.
Perform cleaning, correction, necessary splitting, and timeline control entry by entry, directly organizing into a complete
```
clean.srt
```
.
If the file is very long and the user allows parallel processing, assign to sub-agents by subtitle number or time range, but the main agent must finally manually integrate into a complete
```
clean.srt
```
, do not rely on merging scripts.
First organize an allowlist for non-filler deleted entries.
Run
```
scripts/check_clean_srt.py
```
for rule checks and timeline review.
Manually review the alarms, focusing on residual filler words, residual repeated slips of the tongue, possible missing pause spaces, overly short flashes, text misalignment, and the allowlist.

Script Resources

scripts/check_clean_srt.py

Used to check whether the complete

clean.srt

meets requirements such as length, punctuation, timeline, deleted entries, local splitting boundaries, etc. This script should be independently executable.

The script must be compatible with two types of input:

A complete but non-split
```
clean.srt
```
.
A complete
```
clean.srt
```
containing locally split entries, allowing temporary numbers like
```
123a
```
,
```
123b
```
.

Common usage:

bash

python3 scripts/check_clean_srt.py raw.srt clean.srt
python3 scripts/check_clean_srt.py raw.srt clean.srt --allowed-deletions '284,415,450'
python3 scripts/check_clean_srt.py raw.srt clean.srt --allowed-deletions '284,415,450' --fail-on-warnings
python3 scripts/check_clean_srt.py raw.srt clean.srt --latin-word-as-one-char

Delivery Requirements

At least explain the following during delivery:

Output file path
Whether the original file is retained without overwriting
Whether there are allowed deleted entries
Whether the verification script has been executed
Whether there are still alarm points that require manual re-listen review

subtitle-refine

NPX Install

Tags

SKILL.md Content (Chinese)