What "audio to SRT" actually means
An SRT file is plain text: a cue number, a start and end timestamp, then the line. Nothing in it describes the audio. That means producing one always requires answering two separate questions — what were the words, and when was each word spoken — and almost every tool on the market answers both at once with a single speech recogniser.
That works until the recogniser is wrong. It is wrong most often on exactly the words that matter: a person’s name, a medical term, a brand, a product number, a line of dialogue read with an accent. The result is a file with correct timings attached to incorrect text, and every one of those errors has to be found and fixed by hand.
This tool separates the two questions. You supply the words, because you already have them — the script, the read-through transcript, the approved VO copy. The aligner supplies only the timing, matching your exact characters against the audio waveform. A name the recogniser would have mangled never enters the process, so it never has to be corrected.
How the alignment works
The audio is converted to a phoneme sequence. Your text is converted to the phoneme sequence it would produce if read aloud. The aligner then finds the path through both sequences that minimises total mismatch — a dynamic-programming problem, the same shape as the classic Viterbi pass over a hidden Markov model.
Because the comparison happens at the phoneme level rather than the word level, small differences between script and delivery do not break it. An added "um", a repeated word, a contraction read in full — all of these cost a little in the alignment score and are absorbed. What it cannot absorb is a sentence that was rewritten entirely, and that is exactly what gets flagged.
Each line comes back with a confidence score. High confidence means the audio and the text agreed; low confidence means something diverged there and a human should look. That is the practical difference between an aligner and a transcriber: a transcriber cannot tell you where it guessed, because everything it produced is a guess.
- Accepted input: MP3, WAV, M4A, FLAC, AAC, OGG — and video containers (MP4, MOV) if the audio track is what you need
- Accepted text: plain text, one line per cue, or a single paragraph that gets split on sentence boundaries
- Output: SRT, VTT, or plain text with timestamps
- Boundary: 4 hours of audio per file, no surcharge for length
What the file looks like when it comes back
Cues land on sentence and clause boundaries rather than fixed intervals, because the aligner knows where words begin and end and can choose a cut point that does not split a phrase. A cue is never allowed to run past roughly seven seconds, and a one-word cue is merged into its neighbour — both are reading-speed rules, not alignment rules, and they exist because a technically accurate cue that flashes for 200 ms is unusable.
Timings are frame-agnostic. SRT is expressed in milliseconds, so the same file drops into an editor at 23.976, 24, 25, 29.97, or 30 fps without retiming.
Where this goes wrong, and why
Alignment failures are not random. In practice they come from three sources, and knowing them saves an hour of guessing:
- The transcript is not the transcript. If the file you uploaded is a different draft from the one that was recorded, whole paragraphs have no audio to attach to. They come back flagged, not silently mistimed.
- The audio has music or heavy effects over the voice. A music bed at a low level is fine; a wall of sound where the dialogue sits is not, because there is no clear phoneme boundary to find.
- The recording has long silences, or a second speaker overlapping. Overlapping speech is the hardest case for any aligner, including this one — split the tracks and run them separately.
A real cue, not a mock-up
Three cues, cut on clause boundaries, 3.36 s / 3.48 s / 1.38 s. The word "alignment" starts at 7.48 s — the aligner found that boundary inside the audio, it was not estimated from a words-per-minute constant.
So the thing nobody tells you about forced alignment is that the hard part isn't the algorithm.
1 00:00:04,120 --> 00:00:07,480 So the thing nobody tells you about 2 00:00:07,480 --> 00:00:10,960 forced alignment is that the hard part 3 00:00:10,960 --> 00:00:12,340 isn't the algorithm.
Versus running a transcriber and fixing the output
The default path is to send the audio to a speech-to-text model and correct whatever it got wrong. That is a reasonable path for audio with no script. It is the wrong path when the script exists.
| The usual way | AlignSRT | |
|---|---|---|
| Where the words come from | Model prediction, corrected by hand | Your script, unchanged |
| Proper nouns and jargon | Frequently wrong, always manual fixes | Correct by construction — they are in your text |
| What you get when it is unsure | A confident-looking wrong word | A flag and a confidence score on that line |
| Work left after the tool finishes | Proofread the whole file | Check the flagged lines only |
| Cost basis | Usually per minute, transcription priced | $3.99 per 100 minutes, audio length only |
Questions about this
Do I need a transcript before I can use this?
Yes — that is the whole design. If you have no script, run the audio through any transcription service first, correct the text once, then align it here. You still end up ahead, because you proofread the text as text rather than as subtitles with timings baked in.
Does it work for languages other than English?
The aligner works on phoneme sequences, so it handles any language with a phonetic dictionary. English, Spanish, French, German, Portuguese, Italian, Dutch, and Mandarin cover the overwhelming majority of use. If a language is unsupported the job fails up front and your minutes are returned rather than burned.
Can I convert a video file directly?
Yes, MP4 and MOV are accepted — but you are billed on the audio track length, and the picture is never processed. If you only need subtitles, extracting the audio first is faster to upload and costs exactly the same.
What happens if the alignment is bad?
The job returns the file with low-confidence lines flagged, and you can edit any cue in the browser and re-export. If the whole alignment failed — usually because the script does not match the recording — the job is marked failed and your minutes are returned automatically, not on request.
Is the free tier a watermarked demo?
No. The first 30 minutes produce a complete SRT with no watermark, no truncation, and no account requirement beyond an email. It exists so you can judge the alignment quality on your own material before paying for it.