Audio Transcription and Subtitles
Typing up a meeting, a lecture or an interview by hand takes several times as long as the recording. A speech recognition model does most of the work, but its output still needs checking. This guide explains how the model reads a recording, which model to choose, how to get fewer mistakes and how to turn the text into subtitle files.
How Speech Recognition Reads a Recording
The Audio to Text tool uses Whisper, a speech recognition model that OpenAI released in 2022 under the MIT License. According to its model card, Whisper was trained on 680,000 hours of audio and transcripts collected from the internet, about two thirds of it in English.
Whisper does not listen to a whole file at once. It reads 30 seconds of audio at a time, turned into a spectrogram (a picture of which frequencies are loud at each moment), and writes the most likely text for that window, word piece by word piece, together with time stamps. A long recording is therefore cut into windows. The tool places each cut at the longest pause near the end of a window, so that sentences are not split, and skips windows that are silent.
Before it writes anything, the model can also tell which language is spoken: the first thing it predicts is a language tag. The tool uses that prediction on the first 30 seconds of speech when you leave the language on "Detect automatically", and then uses one language for the whole file.
Choosing a Model
Whisper comes in several sizes. The tool offers three: base and small run on the processor in any recent browser, and large-v3-turbo runs on the graphics processor where the browser supports WebGPU.
| Model | Parameters | First download | Time for 1 minute of speech* | Use it for |
|---|---|---|---|---|
| Whisper base | 74 million | about 90 MB | about 30 seconds | Clear recordings, phones |
| Whisper small | 244 million | about 250 MB | about 1 minute 35 seconds | Other languages where WebGPU is missing |
| Whisper large-v3-turbo | 809 million | about 750 MB | about 20 seconds | Computers with WebGPU: the most accurate |
* Measured on one desktop computer in a browser: base and small with 8-bit weights on the processor, large-v3-turbo with 4-bit weights on the graphics card. Slower devices take longer. Each model is downloaded once and then saved in the browser until you delete it in the tool.
The difference in quality is largest outside English. In our test with a one-minute Korean recording read by a text-to-speech voice, base wrote 후회 (regret) for 회의 (meeting) and 머료일 for 월요일 (Monday), small got both right but wrote 무늬 (pattern) for 문의 (inquiry), and large-v3-turbo made no mistakes. Use large-v3-turbo when your browser offers it; otherwise start with base and switch to small when the text has too many mistakes.
Large-v3-turbo is an optimized version of large-v3, Whisper's largest model, built for speed. It needs a graphics processor: on the processor alone it took about 13 times the length of the recording, so the tool shows it only where WebGPU is available, for example in recent versions of Chrome and Edge on computers and in Safari on recent Macs. It also needs far more memory than base, which many phones do not have.
Getting a Better Transcript
- Record close to the microphone. Distance, echo and background noise cost more accuracy than anything else.
- Avoid music and overlapping voices. Whisper writes one stream of text and does not tell speakers apart.
- Choose the language when the recording starts with music, silence or a greeting in another language.
- Expect to fix names and terms. The model guesses spellings from sound and prefers common words.
- Watch for invented text. The model card warns that Whisper can produce text that was not spoken and repeat itself, mostly in silent or noisy parts. Read those passages against the recording.
- Split very long files. An hour of audio takes about 220 MB of memory once decoded, and everything runs on your own device.
Subtitle Files: SRT and VTT
A subtitle file is plain text: a list of cues, each with a start time, an end time and the words to show. The two common formats look almost the same.
SubRip (.srt) numbers each cue and writes times with a comma before the milliseconds:
1
00:00:00,000 --> 00:00:02,500
Good morning, everyone.
2
00:00:02,500 --> 00:00:05,860
Let me go over the main points.WebVTT (.vtt) is the W3C format for subtitles in web pages, used by the HTML <track> element. The file starts with the word WEBVTT, times use a full stop, and cue numbers are optional:
WEBVTT
00:00:00.000 --> 00:00:02.500
Good morning, everyone.
00:00:02.500 --> 00:00:05.860
Let me go over the main points.In WebVTT a cue's text may not contain -->, and the characters & and < must be written as & and <; the tool does this for you. SRT has no official specification, but video editors and players read the form shown above. Choose the format your player or platform asks for.
Making Subtitles Readable
Whisper's segments follow its own sense of a sentence, so some are long. Before you publish subtitles, play them with the video, split segments that stay on screen too long or fill more than two lines, and check that each cue appears when the words are spoken. Most subtitle editors can open an SRT file and shift or split cues.
Local and Cloud Transcription
Online transcription services usually upload the recording to their servers, where large models run quickly. A browser tool keeps the recording on your device: the audio is decoded and transcribed in the page, and only the model files are downloaded. The trade-off is speed, which depends on your processor, and model size, which is limited by what a browser can hold.
Whoever does the processing, recordings of other people deserve care. The Whisper model card cautions against transcribing recordings of people made without their consent; ask before you record or transcribe a conversation.
Frequently Asked Questions
Can Whisper tell who is speaking?
No. It writes one stream of text. Telling speakers apart (speaker diarization) needs a different model; add the names by hand.
Why does the transcript contain sentences nobody said?
Whisper predicts likely text, and in silence, music or noise it sometimes continues with words it has seen often, such as a closing phrase. Skipping silent windows helps, but read the transcript against the recording.
Can it translate the recording into English?
Whisper models can, but this tool only transcribes: it writes the text in the language that is spoken.
Which format does a video platform need?
Most video editors and platforms accept SRT, and web players use VTT. When in doubt, download both: they contain the same cues.
Sources
Source: OpenAI, Whisper (README) · Last checked:
Source: OpenAI, Model Card: Whisper · Last checked:
Source: W3C, WebVTT: The Web Video Text Tracks Format (Candidate Recommendation Draft, 20 May 2026) · Last checked: