How to transcribe audio or video to text with AI (free and paid options)

Pick the right tool for your situation

You may not need a new app at all. Start with what you already have, and move to a specialist tool when you hit a limit.

Get clean audio (it matters more than the tool)

Every tool does well on clear speech and struggles with echo, background noise and people talking over each other. A few habits fix most problems before they start:

• Get the mic close. A phone on the table right in front of the speaker beats one across the room. For interviews, a cheap clip-on mic helps a lot.
• Pick a quiet, soft room. Turn off fans and notifications. Rooms with curtains, carpet and furniture echo less than bare kitchens and meeting rooms.
• One voice at a time. Crosstalk is the biggest cause of garbled text and wrong speaker labels.
• Say names at the start. Ask everyone to introduce themselves. It makes labeling speakers far easier later.
• On calls, wear headphones so the other person's voice doesn't echo back through your mic. Many video-call tools (Zoom, Teams, Google Meet) can also make their own recording or transcript on some plans.

Have a video? Many tools (Word, Gemini, Otter, Descript, Whisper) accept video files directly. If one only takes audio, extract the soundtrack first. With the free tool ffmpeg installed, this one line does it:

ffmpeg -i my-video.mp4 -vn my-audio.mp3

For very long recordings, split the file into 20-30 minute parts if your tool has a length limit, and transcribe them in order.

Transcribe and label the speakers

In dedicated tools this is mostly automatic. Word labels speakers "Speaker 1", "Speaker 2" and so on; edit a label and choose to change every instance to the real name. Otter and Descript also detect speakers and let you assign names. Do it now: correct labels make every later step more accurate.

If you're using an AI assistant that accepts audio (such as Gemini, or ChatGPT on a paid plan), be specific, or you may get a summary instead of a transcript:

Transcribe this audio word for word. Don't summarize or skip anything.
- Label speakers as Speaker 1, Speaker 2, etc., or by name if they introduce themselves.
- Start each new speaker turn on a new line with a timestamp like [12:34].
- If you can't make out a word or phrase, write [inaudible 12:34] instead of guessing.

Using Whisper? After installing Python and ffmpeg, install Whisper and run it on your file. It creates plain text, subtitle (.srt and .vtt) and data files next to each other:

pip install -U openai-whisper
whisper my-audio.mp3 --model turbo --output_format all

Smaller models (such as base or small) run faster on older computers but make more mistakes. Since Whisper doesn't identify speakers, you'll add names in the next step.

Clean up the transcript with AI

Raw transcripts are full of "um", false starts, missing punctuation and misheard names. Paste the transcript into any assistant (ChatGPT, Claude, Grok or Gemini) with a prompt that cleans it up without rewriting what people said:

Below is a raw automatic transcript of a [interview / lecture / meeting / podcast] about [topic].
Speakers: Speaker 1 = [name, role]. Speaker 2 = [name, role].
Clean it up:
- Fix punctuation, capitalization and paragraph breaks.
- Remove filler words (um, uh, you know) and false starts, but don't reword, summarize or skip anything.
- Correct obvious mis-hearings of these names and terms: [names, products, places, jargon].
- Keep the speaker labels and timestamps.
- If a word or phrase looks wrong and you're not sure of the fix, keep it and add [check].
Return the full cleaned transcript.
[paste transcript]

Two quick checks keep this honest:

• Compare lengths. If the cleaned version is much shorter than the original, the assistant summarized instead of cleaning. Work in smaller chunks (roughly 15-20 minutes of audio at a time) and ask again.
• Listen to the risky bits. Play back every [check] and [inaudible] spot, plus any quote, name or number you plan to use. Most tools let you click a line or timestamp to jump to that moment.

Turn it into notes, a summary or captions

A transcript is raw material. Ask for the format you actually need next:

Using only this transcript, create:
1. A 3-sentence summary.
2. The key points, each with the timestamp where it's discussed.
3. Decisions and action items as a table: task | owner | due date (write "unclear" if not stated; don't guess).
4. 3-5 direct quotes worth keeping, word for word, with timestamps.
If something isn't in the transcript, say so instead of filling the gap.
[paste transcript]

For meetings, our meeting notes to action items guide has follow-up email and agenda prompts. For a long lecture or interview, save the transcript as a .txt or .docx file and treat it like any long document: our guide to summarizing long documents covers chunking and checking with references.

Captions (SRT files). An SRT file is plain text: numbered captions, each with a start and end time, which YouTube and most video platforms and editors can import:

1
00:00:01,000 --> 00:00:04,200
Welcome back to the show.

2
00:00:04,400 --> 00:00:07,900
Today we're talking about budgeting apps.

Get your SRT from the tool that did the transcription, because it knows when each word was spoken: Descript exports SRT and VTT captions, Whisper makes an .srt automatically, and Otter offers caption export on some plans. Don't ask a chatbot to invent caption timings from plain text: without the audio timing data, the timestamps are guesses and will drift out of sync. Fixing typos inside an existing SRT is fine, though:

Here is an SRT caption file. Fix spelling, punctuation and these names: [list].
Don't change any caption numbers, timestamps or the order of captions. Return the complete file in the same SRT format.
[paste SRT]

Handle privacy, consent and accuracy

• Get consent before you record. Tell people you're recording and why. Laws differ by place: in some places, including several US states, everyone in a conversation must agree to be recorded. If a meeting assistant like Otter joins a call, let attendees know it's there.
• Know where your audio goes. Most tools here upload your recording to the cloud. Check whether the service may use your content to improve its models and what settings control that (ChatGPT, for example, has an "Improve the model for everyone" setting). Word's Transcribe saves recordings to a Transcribed Files folder in your OneDrive.
• Keep sensitive recordings in approved tools. For medical, legal, HR or client conversations, use your organization's approved tool, or run Whisper locally so the audio never leaves your computer.
• Never trust a quote you haven't heard. Accents, crosstalk, jargon and numbers trip up every tool. Before you publish a quote or act on a figure, listen to that part of the recording.
• Some uses need a human. Court records, official minutes and medical documentation often require certified or professionally reviewed transcripts. Treat AI output as a draft.

Next, turn your notes into action with AI meeting summaries, or edit the recording itself with Descript.