How to Fix Whisper SRT Files: Tags, Overlaps and Encoding

OpenAI's Whisper and the tools built on top of it have become the fastest way to generate subtitle files from audio. The transcription quality is genuinely impressive. The SRT output, depending on how you exported it, often is not.

Three problems appear consistently in Whisper-generated files: rogue HTML tags injected by GUI wrappers and API integrations, character encoding errors that turn accented or non-Latin characters into garbled boxes, and millisecond-level timing overlaps at cue boundaries that cause flickering on playback. Each problem has a different cause and a different fix.

This guide walks through all three in the order you should apply them, using free browser-based tools with no uploads required.

Why Whisper SRT Files Need Cleaning

Whisper itself — the core model from OpenAI — produces reasonably clean output when run directly from the command line. The issues come from the ecosystem built around it.

HTML tags from wrappers and apps. Tools like Descript, Kapwing, WhisperX, and other Whisper-based applications post-process the model's output and inject formatting tags: <i> for music or off-screen audio, <c.white> color tags from VTT-to-SRT conversions, <u> for word-level highlights in some timestamp modes, and speaker labels like [SPEAKER_00]: from diarization workflows. YouTube auto-captions, which use a Whisper-like model under the hood, download as VTT with color cue tags that do not translate cleanly to SRT. These tags cause downstream problems in video editors, players, and delivery pipelines that do not know how to render them.

Encoding errors on non-Latin scripts. Whisper outputs UTF-8 by default from its CLI, but many Windows-based tools and some API integrations default to ANSI or Windows-1252. The result is mojibake: Arabic, Hebrew, Chinese, Japanese, Korean, or accented Latin characters render as question marks, boxes, or wrong characters. This happens silently — the file looks fine in the tool that generated it but breaks the moment it opens anywhere else.

Millisecond cue overlaps. Whisper segments audio in chunks and assigns timestamps to each chunk. At segment boundaries, the end time of one cue and the start time of the next can overlap by anywhere from one to twenty milliseconds. This is a known and well-documented issue in Whisper's output, especially prevalent in files from faster-whisper, whisper-timestamped, and WhisperX. A 5ms overlap is invisible to the human eye but causes subtitle flickering on some players and upload warnings on platforms that require strictly sequential cues.

The Three-Step Cleanup Workflow

Fix the problems in this order:

1. Strip tags first — before anything else, so the encoding fixer and overlap fixer work on clean text
2. Fix encoding second — normalise the character set before adjusting timing
3. Fix overlaps last — with clean, correctly encoded text, the timing pass is accurate and nothing is missed

Each step is independent. Not every Whisper file needs all three — skip any step where the problem is not present in your file.

Step 1: Strip HTML and Formatting Tags

What tags Whisper-based tools inject

The most common are:

• <i> and </i> — italic tags added around music cues, sound effects, or off-screen speech
• <c.white> and similar — color tags from VTT files converted to SRT without stripping cue settings
• <u> — underline tags from word-level highlight modes
• [SPEAKER_00]:, [MUSIC], [APPLAUSE] — speaker labels and sound annotations from diarization workflows
• <font color="#ffffff"> — color tags from older Whisper GUI tools

These look harmless in a text editor but break rendering in professional video players, subtitle editors, and streaming platform ingest pipelines that expect plain text cues.

How to strip the tags

Paste or upload your SRT or VTT file into the Subtitle Tag Stripper and run it. The tool removes HTML and formatting tags from every cue in a single pass while leaving the timing, cue numbers, and dialogue text completely untouched. You choose which tag types to strip, so if you want to keep italics but remove color codes, you can do that.

Download the cleaned file before moving to Step 2.

Step 2: Fix Character Encoding Errors

When you need this step

You need Step 2 if your file shows any of the following symptoms:

• Question marks or boxes where accented characters should be (caf? instead of café)
• Reversed, broken, or scrambled Arabic, Hebrew, or Persian text
• Chinese, Japanese, or Korean characters replaced with Latin strings or symbols
• Random character sequences that look nothing like the original language

If your file contains only standard English with no accented characters, this step is usually safe to skip.

What causes the problem

When a Windows-based Whisper wrapper saves an SRT file in ANSI or Windows-1252 encoding instead of UTF-8, the file writes correctly in that environment but fails everywhere else. The issue is invisible until you open the file in a media player, editor, or delivery pipeline that expects UTF-8. By that point the character data is already corrupted.

How to fix the encoding

Open the Subtitle Encoding Fixer and paste or upload your tag-stripped file from Step 1. The tool detects the encoding mismatch and re-exports the file as valid UTF-8. Characters that were saved correctly in the original encoding are converted accurately. Download the re-encoded file before moving to Step 3.

Step 3: Fix Overlapping Cue Timings

Why Whisper creates overlaps

Whisper processes audio in segments and assigns a start and end timestamp to each. When two adjacent segments come from a continuous speech stream, the boundary timestamps are sometimes off by a few milliseconds in the wrong direction — the second cue starts before the first one ends.

A typical Whisper overlap looks like this in the raw SRT:

1
00:00:04,280 --> 00:00:06,540
So the key thing to understand is

2
00:00:06,520 --> 00:00:08,900
that timestamps run sequentially.

Cue 2 starts at 6,520ms but cue 1 does not end until 6,540ms — a 20ms overlap. On most players this produces a brief flicker where both cues appear on screen simultaneously. On strict delivery platforms it triggers a validation error that rejects the file.

What happens if you leave overlaps unfixed

The consequences depend on the destination. VLC and most consumer players are tolerant and play through overlaps without visible issues. YouTube accepts overlapping files but the captions flicker on the brief overlap frames. Professional captioning delivery pipelines, Netflix QC tools, and broadcast ingest systems will reject the file outright.

How to fix the overlaps

Upload your cleaned and correctly encoded file from Step 2 to the Subtitle Overlap Fixer and click Fix Overlaps. The tool scans every timing line in a single pass. For each cue whose start time falls before the previous cue's end time, the start is pushed forward to the exact millisecond where the previous cue ended. Cues that would fall below 300ms after the fix have their end time extended automatically so the text stays on screen long enough to read.

Cue numbers, dialogue text, and any remaining formatting are never touched — only the conflicting timestamp lines are rewritten.

Before and After: A Whisper Cleanup Example

Raw output from a Whisper-based GUI — tags present, encoding error, and timing overlap at cue 2:

1
00:00:01,200 --> 00:00:03,400
<i>[Music]</i>

2
00:00:03,380 --> 00:00:05,760
Caf? au lait, s'il vous pla?t.

3
00:00:05,740 --> 00:00:07,900
<c.white>The order was confirmed.</c.white>

After the three-step workflow — tags stripped, encoding fixed, overlaps resolved:

1
00:00:01,200 --> 00:00:03,400
[Music]

2
00:00:03,400 --> 00:00:05,760
Café au lait, s'il vous plaît.

3
00:00:05,760 --> 00:00:07,900
The order was confirmed.

Cue 1: tag stripped, timing unchanged. Cue 2: overlap fixed (start pushed from 3,380ms to 3,400ms), encoding corrected. Cue 3: tag stripped, overlap fixed (start pushed from 5,740ms to 5,760ms).

Which Steps Do You Actually Need?

Run all three if you are unsure — each tool passes clean cues through unchanged and finishes in seconds regardless of file size. If you know the source, use this as a quick guide:

• YouTube auto-captions downloaded as SRT or VTT: Step 1 (color tags) and Step 3 (overlaps)
• WhisperX with speaker diarization: Step 1 (speaker labels), Step 3 (overlaps)
• Windows GUI tools on Arabic, Hebrew, or CJK content: Step 1 and Step 2
• OpenAI Whisper CLI, English-only content: usually Step 3 only

For professional deliverables, run all three regardless. The total time across all three tools for a feature-length subtitle file is under a minute.

Frequently Asked Questions