Subtitles: local speech recognition and SRT export

What the "Subtitles" service can do

The service accepts an audio file up to 20 MB in size and duration up to 5 minutes or a microphone recording. Processing takes place locally in the browser: audio is not sent to external servers. As a result you get subtitles with timestamps in SRT format, which can be copied or downloaded. This reduces the risk of data leakage but limits the volume of material processed.

The interface is minimalist: a field for text instructions (up to 8000 characters), output language selection and a Generate button. The result is displayed in a separate panel. There are no style settings, sliders or preview - only basic functions. Such an approach is convenient for quick tasks, but is not suitable for fine-tuning timings or formatting.

The service does not perform batch processing, does not save history and does not integrate with video editors. This is worth considering when planning your workflow.

Upload the file interview.mp3 (3 minutes), select the language "Russian", press Generate. After a few seconds text with timestamps will appear in the panel. Copy it to a .srt file.

Which stages can be delegated to automation

Initial speech recognition is an ideal task for automation. The service quickly converts audio to text and sets timestamps. This saves hours of manual work, especially for recordings with clear diction and neutral accent.

Rough splitting into subtitles is also automated: the algorithm divides the speech stream into segments, relying on pauses. This is a convenient foundation that can then be edited. Do not expect ideal splitting for fast speech or dialogues with overlaps - such places will have to be corrected manually.

If the recording contains only one speaker without strong background noise, the result may be quite suitable for publication after a light proofread.

For a 4-minute lecture the service output 32 subtitles. After checking it was necessary to merge 5 short phrases and fix 3 typos.

What must be checked manually

Punctuation and capital letters. Automation often ignores commas, dashes and question marks, especially in complex sentences. Without manual editing subtitles look sloppy and make reading difficult.

Terms, proper names and numbers. The service may make mistakes in specialised vocabulary, abbreviations, dates and amounts. For example, "twenty five" may turn into "25" or vice versa. Always verify such places against the original audio.

Duration and readability. Subtitles should not flash too quickly or stay on screen longer than the phrase sounds. It is recommended that one line contain no more than 42 characters and remain on screen from 1 to 6 seconds. Automatic timestamps sometimes require correction.

In an interview "We invest in R&D" was recognized as "We invest in ar en de". Had to be corrected manually.

When the service is not suitable

Recordings longer than 5 minutes or larger than 20 MB are not accepted. For podcasts, webinars and movies preliminary cutting into fragments is required, which increases processing time and risks losing context at the junctions.

Strong background noise, music, multiple speakers speaking at the same time or a pronounced accent reduce recognition quality. In such cases automatic subtitles may contain many errors, and manual editing will take more time than typing from scratch.

If video synchronization, color-coding of speakers, ASS/SSA format support or embedding into a player is required - the service does not provide such capabilities. This is a tool for basic SRT files only.

For a recording with a strong accent and air conditioner noise the service gave 40% errors. Had to retype manually.

Practical workflow

Step 1: prepare the audio. Make sure the file meets the limits (up to 20 MB, up to 5 minutes). If the recording is longer, cut it into parts of 3-4 minutes with a small overlap of 2-3 seconds so as not to lose words at the boundaries.

Step 2: run recognition. Select the language, if necessary add an instruction (for example, "this is a dialogue between two people"). Press Generate and wait for the result.

Step 3: edit the SRT. Open the file in a text editor or special program. Check punctuation, fix errors, adjust timestamps. Remove unnecessary pauses, combine short replicas.

Step 4: check synchronization. Play the video with subtitles and make sure the text appears on time. If necessary shift the marks manually.

For a 4-minute video: cut into 2 parts, recognition, manual editing 15 minutes, final check 5 minutes. Total about 25 minutes instead of 2 hours of manual typing.

Limitations and typical errors

The service does not guarantee perfect accuracy. The model can confuse words that sound similar, skip endings, incorrectly determine sentence boundaries. Always leave time for proofreading.

Timestamps may "drift" on long pauses or when the speaker changes. If there are musical insertions in the recording, the algorithm may try to recognize them as speech. Such fragments are better removed manually.

Do not use the service for legal, medical or financial documents where a term error may have consequences. In such cases professional transcription with expert verification is required.

In a medical podcast the service replaced "arrhythmia" with "arithmetic". Without checking such an error is unacceptable.

How to evaluate the result quality

Compare several random subtitle fragments with the audio. Pay attention to word accuracy, punctuation and timing. If errors exceed 10% or timestamps are off, the result requires serious improvement.

Check readability: subtitles should be easily perceived from the screen. Too long lines or rapid replacement make understanding difficult. If necessary break long phrases into several subtitles.

Make sure the SRT format is correct: sequential numbers, timestamps in HH:MM:SS,mm format, empty line between blocks. Errors in structure can cause the player not to display subtitles.

After editing check the file in VLC: open the video, add subtitles and scroll through several moments.