Auto Subtitle Generator Offline: Correctly Timed SRT Without Uploads

What Auto Subtitle Generator Offline actually does

Auto Subtitle Generator Offline is a browser-only subtitle timing studio: it takes a transcript you already own, finds where speech happens in your own audio or video file, and exports a time-aligned .srt or .vtt. It does not transcribe speech to text, and it never uploads your media.

The workflow is deliberately narrow. You drop in one media file (audio or video), paste your script into a textarea, and the page decodes the audio locally, detects speech segments by cutting at silence gaps, then matches your lines to those segments. Alphabetical guessing is not involved — your words stay your words; the tool only supplies in and out timecodes.

Because alignment is a visual match-and-verify job, the layout is a split view: media and script controls on the left, a live cue table and waveform on the right. The waveform ribbon doubles as a scrubber, and clicking any cue seeks the player to that moment. You can open the full tool at Auto Subtitle Generator Offline and see the preview update as you change settings.

How does it compare with automatic captioning?

Automatic captioning tools generate the words. Auto Subtitle Generator Offline times words you already have. That single difference decides which category of tool you should reach for, and it makes this one much faster when your script already exists because there is no transcription pass to pay for or wait on.

The alternatives fall into three practical groups. Cloud captioning services upload your file, transcribe it per minute, and bill you for the privilege — fine for public content, awkward for client footage under NDA. Local speech-to-text engines keep files private but need an install and can be slow on long recordings. Editor caption panels are accurate but tied to that specific application and often run a transcription step anyway. This tool sits in a fourth slot: no transcription, no upload, export in seconds.

The honest limitation is that this is the wrong tool if you have no transcript at all. It also will not fix a script that does not match the audio. But a transcript can come from more places than people expect: a YouTube auto-caption export you want retimed, a teleprompter script, subtitle text from a previous cut, or meeting notes with the speaker's actual sentences.

How the timing controls change your cues

Timing mode decides the unit of each cue. The default is one cue per sentence, which suits most talking-head and course content. Per line treats every newline in your script as its own cue, useful when you have already broken the text exactly the way you want it on screen. Per paragraph produces longer cues that work for read-along narration but usually need a lower wrap width.

The silence-gap control, min_gap_ms, defaults to 350 ms and accepts 100 to 2000. Lower it when a speaker barely pauses between sentences — at 350 ms, two short sentences separated by a 200 ms pause can be merged into a single cue. Raise it for slow narration with long breaths so cues do not fragment into one-word flashes.

Text wrapping is handled by max_chars_per_line, default 42 and adjustable from 20 to 80. At the default, a cue that runs long wraps onto a second line, so roughly 84 characters fit before you should be splitting the cue instead. The offset field shifts every cue from -5000 to +5000 ms; if your recording starts with a 2.5-second slate, entering 2500 moves the whole file into place in one edit.

A five-step workflow for a clean SRT

First, prepare the script as plain text and decide your line breaks before pasting. The tool splits on newlines first, then sentence punctuation, so a clean script with one idea per line gives you the most predictable result. Remove timestamps, speaker labels, and stage directions that are not spoken.

Second, drop in the media file. It must be 100 MB or smaller and decodable by your browser — in practice mp3, wav, m4a, mp4 and webm. Third, choose the timing mode that matches your script shape, then check the waveform strip: detected speech appears as blocks, and cue spans are drawn over them. Gaps between cues and blocks are your first signal that something is off.

Fourth, fix problems at the source rather than cue by cue. A uniform drift is an offset problem; merged cues are usually a min_gap_ms problem; cramped captions are a wrap-width problem. Fifth, download the .srt or .vtt and play it back in a real player against the real video before you publish. The export is a Blob download named from your source file, so there is no project to save or account to log into.

Formats, limits, and what breaks

The 100 MB ceiling is the constraint most people hit first. As a rough guide, a 128 kbps mp3 runs about 0.96 MB per minute, so an hour of audio lands near 58 MB — inside the limit but with little headroom. A high-bitrate video export can blow past it in minutes, which is why stripping the audio track first is usually the fastest fix.

As of 2026, browser decoding covers mp3, wav, m4a, mp4 and webm reliably. Common failures are MKV containers with AC3 audio and DRM-protected downloads; those are rejected or warned about because the browser simply cannot decode them, not because the aligner is picky. Protected or encrypted files will never work, and re-wrapping will not help.

Two smaller guards are worth knowing. A script under five words triggers a warning, since there is nothing meaningful to align. And the tool does no translation, no burn-in, and no editing after export — you re-run alignment instead. If you are organising a batch of recordings, it helps to keep a simple plan; a free online schedule maker generator is enough for slotting recording, captioning and review days.

How accurate is the alignment?

Accuracy depends almost entirely on how closely your script matches what was actually said. When the script and audio agree line for line, alignment is consistent; when the speaker ad-libbed, reordered sentences, or the script contains a paragraph that was cut in editing, those lines have nowhere to land. The diagnostics strip reports detected speech segments and the aligned versus unaligned line count, so you can see the mismatch as a number rather than guessing.

The same strip reports a words-per-minute reading pace. A cue of 90 words over 30 seconds works out to 180 words per minute, which is far too fast to read comfortably on screen; splitting that cue or shortening the text is the realistic fix. Anything that reads uncomfortably in the cue table will read worse on a phone.

Treat the exported file as a strong first pass, not a finished deliverable. Watch the first minute, the last minute, and any section with music or overlapping speech, because those are where silence detection struggles. Then verify against your own playback before publishing, and check that the timecodes survive the import into your editor or hosting platform.

References

• W3C: Web Video Text Tracks Format (WebVTT)
• Wikipedia: SubRip (.srt) subtitle format