Tutorials Video

Transcribing Speech with AI: Whisper, Built-In Tools and What to Proofread

By Infinite Creation · Beginner · ~14 min

A 1960s mid-century comic book style editorial illustration showing a person with headphones at a laptop where a speech waveform becomes lines of caption bars, while a colleague proofreads the lines with a red pencil and magnifying glass.

Overview

Automatic transcription turns recorded speech into text with timing, so you search it, edit from it, or use it as captions. It is the AI job with the clearest free route and the most dependable first draft. This guide is for creators and editors who need a transcript or captions. It explains what Whisper, the open source speech recognition model, needs to run, how to choose its size, what built-in tools offer, and which errors a person must still fix before anything is published. We did not test any tool on this page. Each statement about a tool comes from the maker's own page as read on 8 October 2026, and the Sources section says what each page established.

Kit and Prerequisites

A recording of speech, preferably with one speaker at a time
A computer where you are allowed to install software, or a built-in feature in your editor
A person to proofread names, numbers and technical terms
A caption format your platform accepts, which the caption pages cover

Working Through It

Choose where the transcription runs

Whisper runs on your own computer. Its page states the code and model weights are under the MIT licence, and that it needs Python and the command-line tool ffmpeg. Built-in features in editors keep you in one program. Live broadcast captioning is a separate market, sold by contract, and is outside this guide. Run local tools when the material is confidential.

Match the Whisper model to your graphics memory

Whisper's page lists approximate GPU memory by model: tiny and base about 1 GB, small about 2 GB, medium about 5 GB, turbo about 6 GB and large about 10 GB. Larger models are slower. The page states the turbo model is not trained for translation. Pick the largest model your memory allows for names and accents, and a smaller one for a quick draft.

Run it on a copy of your file

The project's documented command to transcribe with the default model is whisper audio.mp3 --model turbo. To set the spoken language, add --language and the language name. To translate speech into English, the page shows a multilingual model with --task translate, for example --model medium --language Japanese --task translate. whisper --help lists every option.

Proofread the parts that go wrong

Read the transcript while the audio plays. Check proper names, brand names, numbers, places and technical terms first, because a model guesses at words it has not met. Check speaker changes, because a plain transcript sometimes merges two speakers. Fix punctuation only where it changes meaning.

Check the timing before it becomes captions

A transcript becomes captions only after its timing is checked. Each caption should appear with the words and stay long enough to read. The caption quality page covers line length and reading speed. The site's caption checker tools test a caption file against those rules.

Worth knowing:

Record clean audio first. A better microphone helps more than a larger model.
Feed the model one language per file. The page lists language identification as a task, but mixed-language audio needs a person.
Keep the raw machine output and the corrected version as separate files.

Knowledge Base

Reference Notes

Two points explain most disappointments: the model is a guesser, and its first draft is not a finished caption.

What Whisper is and what it does

The project's page describes Whisper as a general-purpose model trained to do multilingual speech recognition, speech translation and spoken language identification, and it lists voice activity detection among its tasks. Open weights mean anyone runs it, which is why many other transcription apps are built on it. Audacity's OpenVINO effects, for example, use Whisper.cpp for transcription and translation, according to the Audacity page.

Where machine transcripts fail

Crosstalk, heavy accents, music under speech, and uncommon names are the usual failures. A raw machine draft is a starting point, not a finished caption. The page on how auto-captioning works explains the mechanism, and the caption quality page explains what a finished caption needs.

Where This Fits

A machine transcript is a first draft, and Publishing Accessible Video covers the captions, transcripts and conformance checks a release needs after it.

FAQ

Q: Is Whisper free to use commercially?
A: The project page says the code and model weights are released under the MIT licence. Read the licence text before commercial use.

Q: Does Whisper need a graphics card?
A: The page lists memory needs by model for GPU use. We did not test running it without one, and the page does not say. Expect smaller models to be the practical choice on weak hardware.

Q: Does Whisper subtitle a video?
A: It produces text and timing. Turning that into a caption file or burned-in subtitles is a separate step.

Vocabulary

Speech recognition Software that converts spoken audio into written text. Modern systems are neural networks trained on large amounts of recorded speech. Accuracy depends on audio quality, accents and vocabulary.

Model size How many parameters, the learned numbers inside a neural network, a model holds. Larger models are more accurate on hard audio and need more graphics memory and time.

Language identification A task where software decides which language a recording contains before transcribing it. Mistakes here send the whole file to the wrong vocabulary and produce nonsense text.

Sources

Whisper's own project page and two maker pages were read on 8 October 2026. We ran no transcription. Checked 8 October 2026.

OpenAI, Whisper project page. States the MIT licence for code and weights, GPU memory by model, the tasks supported, that turbo does not translate, the Python and ffmpeg requirements, and the documented command-line examples.
Audacity, OpenVINO AI effects. States transcription and translation use Whisper.cpp and run locally, on a Windows download.

Avid's Pro Tools Speech-to-Text and AI-Media's live captioning are not described from their own pages because both pages were unreachable. Trade coverage says Pro Tools 2026.4, released 30 April 2026, extended Speech-to-Text. We did not read Avid's release. No accuracy figures are quoted because we ran no tests and found no independent figures.