Transcribe Audio With Whisper, ChatGPT or Gemini

Yes. Whisper, ChatGPT and Gemini can all turn recorded speech into text, but they differ in cost, privacy and limits. Automatic speech recognition is the technology under all three. Whisper is OpenAI's open model, and its code and model weights are released under the MIT License ¹, so you can run it on your own computer without uploading the audio. ChatGPT and Gemini are chat apps. You attach a file and get text back. That is quick for a short clip, but each service sets its own limits, and a chat model may summarize or reword what it hears unless you ask for an exact copy.

Choose the route by where the audio can go and how large the file is:

For a short file where privacy does not matter, a chat app is the quickest start. For private, long or sensitive audio, Whisper on your own computer keeps the recording on your device. That is what on-device speech recognition means.

Whisper is a free, open model: run it on your own computer through an app or the command line

Whisper handles many languages, and its accuracy differs between them, as the language section below shows. To run it, you need Python and a terminal. The steps follow the README in OpenAI's repository. How to use Whisper to transcribe audio covers the same route as a question.

1. Install the package. In a terminal, run pip install -U openai-whisper ⁴.
2. Install ffmpeg. Whisper needs the command-line tool ffmpeg on your system, in addition to the Python package ⁵.
3. Transcribe a file. The README's example transcribes speech in audio files with one command, using the turbo model ⁶. For your own recording, run whisper your-recording.mp3 --model turbo.
4. Switch models if needed. If the run is slow or the computer runs out of memory, choose a smaller model. The table in the next section shows what each one needs.

The README does not cover every operating system, so test the setup on a short file before you start a long recording. If the command line is a barrier, a Whisper app with a window is the alternative. Check whether it runs the model on your computer or sends the audio to a server, since a server-based app sends your recording off the device.

Faster engines and apps built on the same model

Some projects run Whisper's models on faster engines. The faster-whisper project says it is up to 4 times faster than openai/whisper for the same accuracy, while using less memory ⁷. That figure is the project's own benchmark, so measure it on your hardware before you switch. For the setup choices in more depth, read how to run Whisper locally.

Pick the Whisper model size by your hardware and how much accuracy you need

Whisper comes in five sizes, from tiny to large, plus the turbo model. Bigger models need more video memory (VRAM), so the size you can run depends on your hardware. The figures below are the VRAM that the README's model table lists for each size.

If your hardware can give the model about 6 GB of VRAM ¹², the turbo model is the one to test first. The README's table rates it at about 8 times the relative speed of the large model ¹⁴. The table does not measure accuracy for each size, so run your own audio through any model you consider before you rely on it.

Accuracy varies by language, so test your own language

Accuracy depends on the language. On the FLEURS benchmark, the Whisper large-v2 model had a word error rate of 4.2 percent for English ¹⁵ and 8.3 percent for French ¹⁶. For Hindi the figure was 21.5 percent ¹⁷, and for Swahili it was 39.3 percent ¹⁸.

A word error rate of 4.2 percent means about four wrong words in every hundred ¹⁵. These figures are for large-v2, so they do not show results for the turbo model. Run ten minutes of your own audio through the model you plan to use, and read the output against the recording before you rely on it.

ChatGPT and Gemini transcribe uploaded audio in a chat, within file and length limits

Both chat apps accept audio files, and both come with limits you have to plan around. The limits belong to each product, so check which one you are using before you upload.

ChatGPT accepts audio files up to 512 MB, on paid plans

OpenAI's help article sets the limit for an audio file at 512 MB ². That is a file size, not a length, so how much speech fits depends on the file's format and quality. Audio uploads are available with paid subscriptions and workspaces, including Enterprise ¹⁹, so the Free plan does not include them. The API is a separate product with a much smaller limit: 25 MB per file ³. Read can ChatGPT transcribe audio files for the full answer.

Gemini's API takes 9.5 hours per prompt, and the app has its own limits

The Gemini API documentation gives a maximum of 9.5 hours of audio per prompt ²⁰. That figure is for the developer API, not the Gemini app, so it does not tell you the app's limit. Google's help page for uploading files to Gemini Apps sets the app's own limits. Check that page before you upload a long recording.

The steps for a chat upload are short. Check the file against the limit in the table above, open a new chat, and attach the file in the same message as the verbatim prompt from the next section. Ask for the transcript in one reply, then read the end of it against the recording. If the reply is a summary, send a follow-up that asks for the speech word for word, and attach the file again so the model works from the audio rather than from its earlier answer.

Gemini can also summarize audio. Google's Gemini API documentation lists describing, summarizing and answering questions about audio content among its use cases, next to transcription ²¹. The same audio input can therefore produce a summary when you wanted a transcript, which is why the prompt in the next section matters. The full answer is in can Gemini transcribe audio.

NotebookLM accepts audio files as sources in a notebook

Google announced in September 2024 that users could add public YouTube URLs and audio files to a NotebookLM notebook ²². Treat its answers about the audio as a starting point, and check any quote against the recording. The question page on whether NotebookLM can transcribe audio covers the details.

Chat assistants can paraphrase or invent words, so prompt for verbatim and spot-check

A chat model follows instructions, which is useful when you want a word-for-word copy. It also writes text, which is how it comes to summarize. Whisper works differently. OpenAI says Whisper “doesn’t follow instructions like a general-purpose text model,” and its prompt parameter accepts up to 224 tokens for passing correct spellings ²³. With Whisper, a prompt can pass it the correct spelling of a name. With ChatGPT or Gemini, a prompt tells the model what to do, so the wording matters.

Writing is also where invented content can enter. In a 2025 study of clinical notes written by large language models, clinicians observed a 1.47 percent hallucination rate, checking each sentence against the source conversation ²⁴. That study covered notes, not transcripts. It still shows that a model asked to write, rather than copy, can add content that was never said.

Prompt the model to copy, then check the copy

Use a prompt that asks for an exact copy and forbids the edits you want to avoid:

Transcribe this audio word for word, in the language spoken. Do not summarize, translate, shorten, reorder or correct grammar. Mark unclear words as [inaudible]. Output only the transcript, with no introduction and no headings.

The word-for-word rule keeps every filler word and false start, which is what a verbatim transcription requires. For a clean verbatim copy, replace the first rule with one that removes filler words, and say so in the prompt. The difference is covered in verbatim vs clean verbatim.

A prompt does not guarantee the result, so check the output:

1. Pick three 30-second passages. Take one near the start, one in the middle and one near the end.
2. Compare word by word. Play each passage and read the transcript against it. Listen for words the transcript adds, drops or changes.
3. Check names, numbers and terms first. A wrong name, figure or technical term changes the meaning, so start there.
4. Check the end. A transcript that stops early, or that closes with a neat conclusion the speaker never gave, shows the model summarized.
5. Split and retry if a passage fails. Cut the file into shorter parts, run each part again and check it the same way.

If the check fails, do not edit the summary into shape. Re-run the copy with a shorter file or in parts, because a corrected summary is still a summary. Keep the original recording as the reference, and mark any passage you could not resolve as [inaudible] in the final text. A marked gap is more useful than a confident guess, because a reader can see where the transcript is uncertain.

Whisper's known failure: invented text during silence and music

Whisper can invent text. In a 2024 study of recordings from speakers with and without aphasia, roughly 1 percent of Whisper's transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio ²⁵. That study tested Whisper as it stood in 2023, so newer versions may behave differently ²⁶.

The trigger to watch for is audio with no speech. In lab tests with artificial sounds, a 2025 study found that a set of hallucinated phrases appeared frequently when non-speech sounds were played to Whisper ²⁷. Whisper's own code checks each 30-second window it decodes. It treats the window as silent, and skips it, when its no-speech probability is above one threshold and its average log probability is below another ²⁸. That rule can also drop quiet speech ²⁸, so it does not replace reading the output.

What to do about it:

1. Trim the silence and music first. Cut long silent stretches and music-only passages from the start and end of the recording before you transcribe.
2. Read the output beside the audio at every long pause. Invented phrases show up where no one is speaking, so those gaps are where to look. Look especially for a phrase repeated back to back, or one unrelated to the topic.
3. Re-run only the doubtful section. When you find an invented passage, transcribe that section again on its own, with the trimmed audio, and compare the two versions against the recording.
4. Check long files in sections. Transcribe and read one section at a time, so each quiet gap gets a look before you move on. Some tools run a voice activity detection step first, which marks where speech is. Check which parts that step removed, and read the removed parts yourself if they matter.

For YouTube, Zoom and long files, extract the audio first and transcribe that

Two things decide the route for video: the size of the audio, and your right to use the video. Start by pulling out the audio track.

Long files: stay under 25 MB or split the audio between sentences

The OpenAI transcription API accepts files up to 25 MB ³. For a larger recording, compress the audio or split it into chunks of 25 MB or less. Cut between sentences, so that no word is divided across two files. Then transcribe each part and join the text in order.

YouTube: check the terms before you download

The yt-dlp README says its option that converts video to audio-only files requires ffmpeg and ffprobe ²⁹. With ffmpeg and ffprobe installed, yt-dlp -x --audio-format mp3 "URL" extracts the audio as an MP3 file.

Before you download any video, read the platform's rules. YouTube's terms say you are not allowed to download any part of the service or its content unless the service expressly authorizes it ³⁰. For a video that already has captions, start with the captions, as covered in video to text. Transcribe recordings you made yourself, or files you have the right to use.

Zoom, Teams and Meet: use the saved file

Meeting platforms can produce their own transcripts, which the guide to Zoom, Teams and Google Meet transcription explains. If you saved a recording as a file, transcribe the file with the steps above. For speaker names, see speaker diarization and timestamps and speaker labels.

A checklist before you transcribe

Work through these four checks before you start, and the transcript is more likely to be right the first time:

• Format and size. Confirm the file is under the limit for the route you chose, and convert it to a format the tool accepts.
• Rights. Confirm you may use the recording, and any consent the people speaking need, before it leaves your device.
• Silence and music. Trim long gaps and background tracks, so the model is not given a chance to invent text in them.
• Output check. Plan the three-passage spot check now, so you do not skip it when the transcript looks clean.

Which route to choose

Use a chat app for one short file, and use Whisper for private or long audio. Use the API when you need to transcribe many files in code. For more options, compare human vs AI transcription and transcription software, or start with how to transcribe audio to text. To transcribe a file in your browser for free, use audio to text. For the whole subject, see our overview of transcription.

If your goal is to speak text into a document rather than transcribe a recording, Yap, which we make, is a dictation app for Mac and Windows. See the download page.