How is this different from the auto-captions in my editor?

An editor re-transcribes the finished cut, and the text and timings are locked together inside the software, so every correction is a separate click. Here the SRT comes from the same pass as the transcript: fix the words in bulk, then export, and the timings are untouched. Audio generated here skips recognition entirely — the timings were computed at synthesis.

How precise are the timings?

Word level, to 0.01 seconds. That means line breaks land on word boundaries rather than an estimated midpoint, so splitting a long line in two never makes the first half vanish early or the second half arrive late.

Can long lines be broken automatically?

The exported SRT splits on sentence boundaries. If a cue is still too long, the word-level timings let you split at any word and read off the exact start and end seconds rather than estimating them.

If I use your voice generation, do I still need to transcribe?

No, and you should not. Synthesis emits a word-level caption track directly, and those timings are exact. Running recognition over the result replaces exact values with estimated ones and can introduce transcription errors. Use the track produced at synthesis.

Subtitle Generator: Timings That Line Up

Generate SRT subtitles from audio or video. Timings are measured per word during transcription, not estimated afterwards, so they align with no second pass.

Last updated: 2026-08-26

Subtitles drift because the timings were guessed

The usual automatic-caption pipeline recognises the words, then distributes the time evenly across them based on speaking rate. Real speech has pauses, held vowels and filler — so even distribution drifts. The first half of a line disappears early; the second half arrives half a second late.

The timings here are not distributed. They are measured per word during transcription, to 0.01 seconds.

How precise that actually is

Real word timings from a Mandarin storytelling clip:

Voice: 老者说书人Chinese17 s

Transcript:上回书说到,那少年提剑立于城门之下,身后是三千追兵,身前是万丈深渊。他回头笑了一声——诸位,且听下回分解。

The hard part of subtitles is timing: two speakers alternating, pauses between lines, and every cue has to land on the exact in and out point. Listen to this exchange, then see how the transcript splits into cues:

Voice: 霸道总裁 / 冷艳女王Chinese16 s

Transcript:你知道我为什么让你留下来吗? 知道。因为整个公司,只有我不怕你。错。是因为只有你,敢在会上说我错了。

上[0.42→0.72]  回[0.72→0.88]  书[0.88→1.06]  说[1.06→1.48]
到[1.48→1.84]  ，[1.84→2.62]  那[2.62→2.96]  少[2.96→3.18]
年[3.18→3.40]  提[3.40→3.64]  剑[3.64→3.96]  立[3.96→4.22]

Look at the sixth entry — the comma. It occupies 0.78 seconds, longer than any word before it. That is the storyteller's pause.

Distribute time evenly and those 0.78 seconds get smeared across the surrounding words, and every cue after that point drifts. Measured timings record the pause as it happened, so everything downstream stays put.

Two paths — take the right one

Where your audio came from	Path	Why
Generated here	Use the caption track from synthesis	Timings are exact — synthesis knows when each word is emitted
Recordings, outside material	Upload and transcribe, export SRT	Timings are measured by recognition, word level

The first row is easy to miss: if the voiceover was generated here, do not run recognition over it. That swaps exact values for estimated ones and can add transcription errors on top. The caption track already exists.

Correcting words never moves the timings

Recognition is never perfect — homophones are a hard limit. But the words and the timings are two separate pieces of data.

Fix a word in the text and its start and end times are unchanged. So the right proofreading order is bulk-correct the text, then export the SRT — not click through cues one at a time inside an editor.

That is the practical difference from editor auto-captions, where text and timing are welded together.

How to use it

Open the voice studio → transcribe
Upload audio or video (200MB total per upload)
Read it through, concentrating on names, jargon and code-switching
Export the SRT and drop it into CapCut, Premiere or Final Cut

One ordering trap: lock the edit before transcribing. Re-cutting afterwards leaves the timings misaligned against the new picture.

FAQ

Different from my editor's auto-captions? There, text and timings are locked together and corrected one cue at a time. Here you bulk-edit text and export; timings are untouched.

How precise? Word level, 0.01s. Breaks land on word boundaries, not estimated midpoints.

Can I split long cues? Yes — split at any word and read the exact seconds.

Do I transcribe my own generated voiceover? No. Synthesis already emitted exact timings.

Start here

Cost: 20 credits per file (up to 200MB per upload), refunded automatically if it fails. The $3.99 trial pack has 150 credits, enough for 7 files, and they never expire.

Upload material and generate subtitles →