Make SRT and VTT Subtitles From Audio

Subtitles used to mean a transcriber, a day of waiting and a bill. Now a small speech model can run inside a web page, on your own device, and write the words for you. It is not perfect, and a tool that pretends otherwise would waste your time, but it gets you from a recording to a timed subtitle file in a few minutes, and most of the work left is a read-through.

The short version: drop an audio or video file into Audio to Subtitles, press Make subtitles, correct any wrong words in the boxes while the file plays, and download an SRT or VTT file. The recording never leaves your device.

What it does, and what it does not

Audio to Subtitles turns English speech in an audio or video file into timed subtitles. You can download them as SRT, as WebVTT (VTT), or as plain text. A small English speech model runs inside your browser, which is why the recording is not uploaded anywhere.

It has firm limits, and it states them. It handles English only. A file can be up to 60 minutes long, and longer recordings must be split and done in parts. It works best on clear speech with little background noise. It does not identify speakers, and it does not translate.

The first use downloads a model

The speech model is about 100 MB. The first time you use the tool, the page says so before you press anything, and the button reads "Download model (about 100 MB) and make subtitles". The page suggests using Wi-Fi, and that is good advice. The model is kept in your browser afterwards, so later uses start at once, and the button changes to a plain "Make subtitles". If you need the space back, a small link, Remove the saved model, clears it.

During the download the page shows how many megabytes have arrived out of the total. On later visits it shows "Loading the saved model...".

What happens when you press the button

The tool works through four stages, and the page names each as it goes.

1. Reading the sound in your file. The sound is decoded and reduced to 16 kHz mono, which is what the model expects. A video file works too, so you can drop an MP4, MOV or WebM directly. If a video has no sound track, you get a clear message that there is nothing to turn into subtitles. If the file is longer than 60 minutes, the message tells you its length and the limit.
2. Finding the speech. The tool measures loudness in 30 millisecond steps and cuts the recording into spoken segments, using the quiet floor of your own file to decide what counts as speech. A silence of at least 0.6 seconds ends a segment, a segment can run to about 28 seconds because the model reads 30 seconds at most, and a few tenths of a second of padding are kept at each end. Because of this, subtitle times come from where the speech really starts and stops. If no speech is found, the tool says "No speech found in this file."
3. Loading the model.
4. Writing the subtitles. The model transcribes one segment at a time, and the page shows "part 7 of 42" with a bar. After the first part, it measures the speed of your device and shows an estimate, such as "About 3 minutes left". Before that, it gives an honest warning that it can take longer than the recording, especially on a phone. Words appear in the list as each part finishes, so you can start reading early. A Stop button cancels the job.

How the cues are built

A spoken segment is not always the right size for a subtitle. A 25-second segment would put a paragraph on screen. So the tool cuts long segments into cues of up to two lines, with at most 42 characters per line and up to about 7 seconds, cutting first at the ends of sentences, then at commas, and only then wherever it must. The time of the segment is shared between the pieces in proportion to the length of the text. Two-line cues are balanced so the lines are about equal.

Here is a worked example with a typical segment. The model hears, from 1.2 to 6.8 seconds: "Welcome to the quarterly update. Today we will look at sales, hiring, and the plan for the next two quarters." The tool builds three cues from it:

• 00:00:01,200 to 00:00:02,875: Welcome to the quarterly update.
• 00:00:02,875 to 00:00:04,759: Today we will look at sales, hiring,
• 00:00:04,759 to 00:00:06,800: and the plan for the next two quarters.

A second, shorter segment from 8.0 to 9.5 seconds becomes a single cue, Let's begin. That is why the preview feels natural while the file plays.

Sound labels and invented words

Speech models sometimes write labels for sounds that are not speech, such as [MUSIC PLAYING] or (applause), and occasionally invent a closing "Thank you." on silence. The tool removes anything in square brackets or round brackets, and music symbols, and drops any segment that is left empty. In the example above, "Let's begin. [MUSIC PLAYING]" becomes "Let's begin." and a segment that contained only "(applause)" disappears. If the model finds no words at all, the tool says so and suggests checking that the file has clear speech.

That is useful, but remember it also removes real words that appear in brackets. If someone says something that the model writes in brackets, you will need to add it back by hand.

The edit step matters most

The tool's own advice is to read the text once while the file plays, and it is built for that. The recording plays in the page, with the current subtitle shown under it, and a video shows the subtitles on the picture itself. Under the player is the list of segments. Each has a time you can click to jump to that point and play, and a text box you can edit. Names, product words, numbers and unusual terms are the usual mistakes, so listen for those.

Your edits stay in this browser tab, so a reload does not lose them, and closing the tab erases them. If you drop a new file while you have unsaved edits, the tool asks you to confirm before replacing them. Download before you leave.

The three downloads

SRT is the most widely accepted subtitle format, used by video editors, many players and most social platforms. The tool writes it with Windows-style line endings, which suits older editors and televisions. VTT, short for WebVTT, is the format for web pages and the HTML video element, and it begins with the word WEBVTT. Text gives you only the words, with a blank line between segments, which is handy for notes, minutes or a blog post. A Copy text button puts the same text on your clipboard.

Files are named after the recording, so interview.mp4 gives interview.srt, interview.vtt and interview.txt. The two subtitle formats differ in one small way you may notice if you open them: SRT writes times with a comma before the milliseconds, 00:00:01,200, and VTT writes a full stop, 00:00:01.200.

Getting a good result

Clear speech is the biggest factor. A recording made close to the microphone beats a recording of a room. Remove long music intros before you start, because loud music can be picked up as speech, which adds parts to the job and mistakes to the text. A single voice at a steady pace works best, and overlapping talkers will be run together.

If the recording is long, cut it into parts of a sensible size first. Audio Trimmer shortens a recording to the section you need. If your file is in a format the browser cannot read, the tool suggests trying an MP3, WAV or MP4, and Audio Converter changes the format. If your recording is the sound of a video that you edit elsewhere, Extract Audio saves it as an audio file, and our guide to extracting an audio track from a video to MP3 shows how. Once you have the file, our subtitle adder guide explains how to add subtitles to a video once you have them.

A note on timing

Each subtitle is timed from where the speech starts and stops, so it can sit a fraction of a second off. That is normal for this method and is good enough for most uses. If you need frame-accurate timing for broadcast, finish the file in a subtitle editor. For web video, training clips and meetings, the result is usually fine as it stands once the words are right.

Steps

1. Open Audio to Subtitles and drop in an audio or video file with English speech, up to 60 minutes.
2. On the first use, connect to Wi-Fi and press the button that downloads the model, about 100 MB. It happens once.
3. Wait while the page reads the sound, finds the speech and writes each part.
4. Play the file and read along. Fix wrong words, names and numbers in the boxes.
5. Click any time to jump to that point and check a doubtful line.
6. Download the SRT or VTT file, and the text version if you want a transcript.
7. Download before you close the tab.

The tool is a Professional Tool with 10 trial uses a day, then a $1 day pass or the $4.99 monthly plan. Pressing Make subtitles uses one trial use.

Common mistakes

Starting on mobile data and being surprised by a 100 MB download. Dropping in a recording with several speakers talking at once. Skipping the read-through, so product names and numbers go out wrong. Forgetting that the tool only does English. Closing the tab before downloading. And expecting broadcast-grade timing from a method that times each phrase by where the speech starts and stops.

Summary

A speech model on your own device turns a recording into a timed subtitle file without uploading a word. Audio to Subtitles finds the speech, writes it in cues that fit on screen, removes sound labels, lets you correct the text while the file plays, and gives you SRT, VTT or plain text. More tools for sound are in the audio tools category.

Open Audio to Subtitles→

What language should we add next? Tell us on the contact page.