Audio to Text

Speech recognition has become good enough that a clean recording often needs only light corrections. It saves the slow part of the job, typing everything out, and leaves you with the quicker part: reading through and fixing what was misheard. For notes, research interviews, lecture study, meeting minutes and first drafts of subtitles, that is usually a large time saving.

It is not perfect, and it would be misleading to promise otherwise. The software guesses the most likely words from the sound, and when the sound is unclear, so is the guess. A transcript can read smoothly and still contain a wrong word that changes the meaning, which is more dangerous than an obvious gap. The box below lists what tends to cause errors, so you know where to look when you proofread.

Most of the accuracy is decided before you upload, when the recording is made. A phone placed close to the speaker beats a better microphone across the room. A quiet room beats a café. One person at a time beats a lively discussion. When you can plan the recording, those three points do more than any setting.

When the recording already exists, a few things still help. Set Spoken language yourself if you know it, instead of relying on detection; detection listens to the opening seconds, and an intro with music or a greeting in another language can send it the wrong way. Cut long silences, music and unrelated chatter from the edges with Trim Audio, or simply use Start at and End at to transcribe only the part you need. A very quiet recording may benefit from Normalize volume in the Audio Converter page first.

Video files can be uploaded directly; the sound track is read and the picture is ignored. There is no need to extract the audio first, though if you want to keep the audio as well, Video to Audio saves it as a separate file.

Batches help with series. A semester of recorded classes, a set of research interviews or a season of podcast episodes can go in together, up to 20 files of up to 95 MB each, and every file gets the same language and output settings. If the recordings are in different languages, run them as separate batches with the language set for each, rather than relying on detection for all of them.

Plain text (TXT) opens anywhere and is the easiest to paste into notes, emails or other tools. Word (DOCX) is the practical choice when the transcript will be edited, commented on or shared as a document. PDF suits a final, read-only record. JSON (word timings) is for developers and researchers: it lists each word with its start and end time so that software can search, align or analyze the speech.

Subtitles (SRT) and Web subtitles (VTT) are subtitle files: short lines of text, each with the time it appears and disappears. SRT is understood by nearly every video player and editor. VTT is the format browsers use for captions on web pages and is specified by the W3C. YouTube accepts both, along with several other caption formats, and lists the details on its help page for supported caption files.

Characters per line applies to subtitles and runs from 20 to 80, with 42 as the default. That default matches the line length used in widely followed style guides, such as Netflix's guide for English subtitles. Shorter lines suit phone screens and vertical video; longer lines mean fewer subtitle changes but more reading at once.

If you are unsure, pick Word for anything you will read and edit, and SRT for anything that goes back into a video. You can always run the same recording again with another output; the transcription itself is the same, only the file around it changes.

Timestamps put the time, such as [01:15], in front of each paragraph of a TXT, Word or PDF transcript. They make it easy to jump back to the recording and check a passage, which is exactly what proofreading needs. Subtitle and JSON outputs always contain timings, so the option does not apply to them.

Speaker labels mark changes of voice as Speaker 1, Speaker 2 and so on. They are useful for interviews, panels and meetings, but the software works only from the sound of the voices: two people who sound alike may be merged, and one person may be split in two after a change in microphone distance. Rename the labels to real names while you proofread.

Punctuation is on by default and adds full stops, commas, question marks and capital letters, which makes the text readable. Turn it off only if you need raw words for another program. Remove filler words drops hesitations such as um and uh for a cleaner read; leave it off when you need a faithful record of exactly how something was said, as in research interviews or quotes.

The quickest way to check a transcript is to play the recording at normal or slightly faster speed while reading along, pausing to fix errors as you hear them. Timestamps help you find your place. Pay most attention to the kinds of words listed in the box above; ordinary sentences are usually right, and names and numbers are where errors hide.

For subtitles, also watch them with the video once. Check that each line appears while the words are being spoken and that nothing stays on screen too long. Quotes for an article, evidence for a report and anything published under someone's name deserve a word-by-word check against the audio.

Keep the recording until the transcript is final. Results here are deleted after 30 minutes, so save the transcript right away, and keep your own copy of the audio so you can go back to a passage later if a question about the wording comes up.