How to Transcribe Audio to Text: Free Methods and Accuracy Tips

A transcription engine never hears words the way you do. It slices the waveform into overlapping frames about 25 milliseconds long and turns each slice into a spectrogram, a map of which frequencies are loud at that instant. The acoustic model looks at that map and estimates which speech sounds are present: the hiss of an s, the burst of a t, the vowel shapes that tell ship from sheep. On its own that produces a stream of likely sounds, not readable text.

The language model fixes much of that. It knows which word sequences are likely in real sentences, so when the sounds could be recognize speech or wreck a nice beach, context tips the balance. Older systems kept these two parts separate and trained them on limited, carefully read material, which is why dictation software of the 2010s stumbled on accents and casual talk.

In September 2022 OpenAI released Whisper, an open model trained on about 680,000 hours of multilingual audio collected from the web. It folds the acoustic and language parts into one network that reads 30-second windows and writes text with punctuation directly. Models in this family made accented speech, mixed vocabulary and messy real-world recordings far more usable, and most free transcription services today, including browser tools, run something similar. They still guess, though, and when the audio is unclear they guess fluently, which makes their mistakes harder to spot.

Accuracy is usually measured as word error rate: the share of words that were swapped, dropped or invented compared with a careful human transcript. A 5% error rate means one mistake every twenty words, which sounds small until you notice that the mistakes cluster on the words that matter, like a surname, a product code or a figure in a budget discussion.

The factors below are listed roughly in the order they hurt. Notice that almost all of them are decided when the recording is made, not when it is transcribed. Moving a phone from the middle of a table to within arm's length of the speaker often does more than any setting you can change afterwards.

Most recordings can go in exactly as they are. Preparation pays off in three situations: the file is too long or too large, the speech is buried under silence or music, or the level is so low that you have to turn the volume all the way up to hear it. A long silent stretch at the start is more than wasted time, because auto-detect decides the language from the first 30 seconds and may guess from room noise or an intro jingle. Cutting the recording down to the part with speech, or setting Start at to skip it, avoids that.

Format conversion is rarely needed. iPhone Voice Memos are M4A (AAC audio), Android recorders produce M4A, AAC or 3GP, and call recorders often produce AMR; all of these can be uploaded directly. Converting a compressed file to WAV does not restore detail that was thrown away when it was recorded, as the comparison in our guide to MP3, WAV, FLAC and AAC explains, so only convert when a tool refuses the format. For a recording that runs for hours, split it at natural breaks such as a coffee break or a change of topic, so each part stays under the upload limit and is easier to proofread.

Pick the output by what happens next. If you are going to read, quote or search the transcript, plain text or Word is right; Word keeps paragraphs, timestamps and speaker labels laid out for editing, and PDF is for sending a fixed copy someone should not change. If the words will appear on screen over a video, you need a subtitle file: SRT or VTT break the speech into short cues, each with a start and end time, and the line length decides how much fits on screen. Forty-two characters per line is the common broadcast standard.

SRT and VTT carry the same information with small differences. SRT, from the old SubRip program, numbers each cue and writes times with a comma (00:01:15,000). WebVTT is the W3C format for HTML5 video; it starts with a WEBVTT header, uses a period in the times and allows styling and positioning. YouTube, Vimeo, Premiere Pro and DaVinci Resolve accept both. JSON with word-level timings is for developers building search, karaoke-style highlighting or their own subtitle layout.

Before uploading anything, check what your own devices already do. Apple added transcripts to the Voice Memos app in iOS 18, so a recording made on a recent iPhone can show its text right in the app. Google's Recorder app on Pixel phones has transcribed while recording since 2019, entirely on the device, and can tell speakers apart. In Microsoft 365, Word for the web has a Transcribe command next to Dictate that accepts uploaded MP3, M4A, WAV and MP4 files with a monthly allowance of upload minutes. Google Docs Voice typing (under Tools, in Chrome) only listens to a live microphone; it cannot read an existing file.

Meeting platforms are another source. Zoom, Microsoft Teams and Google Meet can produce a transcript of a call when transcription is switched on and your plan or organization allows it, and assistants such as Otter.ai join meetings to take notes. These are convenient for calls you host, but each works only inside its own app, needs everyone's consent to record, and exports in its own formats. For a file you already have, such as an old interview, a lecture recording or a voice note forwarded by someone else, a file-based transcriber is usually the simpler route.

A browser tool covers what the built-in options leave out: any file from any device, subtitle outputs, and languages your phone does not support. On the audio to text converter, add the file, leave Spoken language on Auto-detect or pick the language yourself, choose an Output, and turn on Timestamps or Speaker labels if you need them. Voice memos from an iPhone have their own M4A to text page, and podcast or recorder files have MP3 to text; both run the same engine. Each file can be up to 95 MB, roughly an hour and a half of 128 kbps MP3. For anything longer, use Start at and End at to transcribe in parts, or cut the recording first.

Set honest expectations. A single speaker on a decent microphone in a quiet room usually comes out 90–95% correct or better, often good enough to publish after a light edit. A phone on a table in a busy restaurant, a conference call with dropouts, or a panel where people interrupt each other can fall much lower, with whole sentences missing. Speaker labels are reliable when voices differ clearly and people take turns, and they mix up two similar voices more often than you might expect.

Proofread with the audio open, not from the text alone, because modern models write grammatical sentences even when they are wrong. Check every proper name, since a model has no way of knowing that your colleague spells it Caitlin or Katelyn. Check every number: fifteen and fifty, dates, prices and percentages are easy to mishear. Watch for homophones such as their/there, affect/effect and complement/compliment, and for confident filler where the audio was unclear. Timestamps make this fast, because you can jump to the exact moment a doubtful line was spoken.

A recording often carries more personal information than a document: voices, names, health details, salary figures. Before you upload, know where the file goes and how long it stays. On UltKit, uploads are processed on the server and deleted automatically 30 minutes after processing; there is no account, so nothing is tied to your name. Pixel Recorder works offline, so the audio never leaves the phone. For anything covered by a confidentiality agreement or medical or legal privilege, check your organization's policy first; some require approved, contracted services.

Language support is broad but uneven. The transcriber covers 33 languages, among them English, Spanish, French, German, Portuguese, Arabic, Hindi, Japanese, Korean, Chinese and Ukrainian, and the best results come in the languages with the most training data. Auto-detect picks one language per file. If a meeting moves between two languages, transcribe each stretch separately with the language set by hand. The transcript is always written in the language that was spoken; translation is a separate step you do afterwards on the text.

Most failed transcripts come down to a handful of causes, and the fix is usually in the file or one setting rather than in trying another service. If the audio itself is too poor for you to follow comfortably, the machine will struggle just as much, and a short manual transcript of the key passages may be quicker than correcting a long automatic one.