OpenAI Whisper Transcription: How to Use It and When to Switch

Updated 5 November 2026 · Standard formula, shown below

OpenAI Whisper turns spoken audio into text, and can also identify the language or produce translated text when you choose the appropriate task. You can run it locally with the open-source Whisper package or send audio to an OpenAI transcription API; the best route depends on your privacy, setup, and accuracy needs. Before starting, make sure you have permission to transcribe the recording and know whether you need a verbatim transcript, readable notes, or a translation.

What You’ll Need

An audio file in a format your chosen tool accepts, such as WAV, MP3, M4A, or another supported format. For API uploads, check the current endpoint documentation for supported formats and file-size limits.
For local use: Python, the Whisper package, and FFmpeg installed and available on your system. A compatible GPU can speed up processing, but is not required.
For API use: an OpenAI account with API access, an API key, a working internet connection, and a current supported transcription model and endpoint. API availability, names, limits, and pricing can change, so verify them in the official documentation.
A quiet, clear recording and enough time to review names, numbers, technical terms, and punctuation.

Step-by-Step

Choose local or API transcription. Use local Whisper if you want to process files on your own machine and can install the software. Use the API if you prefer a managed service and can send the recording over the internet. Local processing avoids uploading audio to the API, but the file and transcript still need appropriate protection on your device. Check privacy and retention terms before sending sensitive recordings to any service.
Prepare the recording. Work from the clearest original file available. Reduce background noise only if doing so does not distort speech; aggressive filtering can erase quiet words. If the recording contains long silence, you may trim it, but keep a copy of the unedited original. Confirm that speakers have consented where required and remove or protect personal information as appropriate.
Install local dependencies if needed. Install Python and FFmpeg using the trusted instructions for your operating system, then install the Whisper package in a terminal, for example with pip install -U openai-whisper. Run ffmpeg -version to confirm FFmpeg is available. Installation commands and compatibility can vary, so consult the package’s current documentation if a command fails. Avoid running unknown scripts with administrator privileges.
Run a local transcription. In the terminal, use a command such as whisper recording.mp3 --model small, replacing the filename with your audio file. Whisper offers model sizes with different speed and accuracy trade-offs; larger models generally take more memory and processing time. The command can create transcript files in the working directory. Use whisper --help to see options supported by your installed version.
Set language or task when appropriate. If you know the spoken language, specifying it can avoid language-detection errors; use the language code and option format documented for your installed version. For translation, use Whisper’s translation task where supported. Translation is not the same as a transcript in the original language, so choose transcription when you need the words as spoken.
Use the API alternative. Create an API key through your account’s official interface and keep it secret; do not paste it into public code or share it in logs. Send the audio to the current transcription endpoint with a supported model, following the official API documentation for request format, upload limits, and available output options. Store the returned text securely. Do not assume API model names or limits are permanent.
Review and correct the output. Listen while following the transcript, especially around speaker changes, overlapping speech, acronyms, proper names, dates, and amounts. Whisper can omit or invent words, particularly when audio is unclear. Add speaker labels yourself if needed; do not treat automatically generated labels or timestamps as proof of who spoke.
Export and preserve useful context. Save a clean copy in the format your workflow needs, such as TXT, SRT, or VTT if generated by your tool. Keep timestamps if they matter for editing or evidence, and retain the source recording when permitted. Label machine-generated text as a draft if others might mistake it for a verified transcript.

Common Mistakes

Using the wrong task: Translation produces translated text, not necessarily the original-language transcript.
Assuming the transcript is exact: Fluent-looking text can still contain wrong names, figures, or missing phrases. Verify important details against the recording.
Choosing a model only by speed: A smaller model may be adequate for clear speech, but can struggle with accents, noise, or specialist vocabulary. Compare output quality on a representative sample.
Expecting speaker identification: Transcription does not automatically provide reliable diarization in every setup. Use a tool with explicit speaker-labeling features if that is essential, then check the labels.
Ignoring file limits or dependencies: API uploads have limits, and local conversion depends on working audio tools. Check the current documentation rather than assuming every file or command will work.

Tips

Record close to the speaker, minimize room echo, and avoid placing microphones near fans or traffic. Better input usually helps more than repeatedly changing settings.
Test a short section before processing a large collection. Include difficult conditions in the sample, not just the clearest passage.
For local runs, start with a smaller model to check the workflow; move to a larger one if the errors justify the extra time and resources.
Keep a terminology list for recurring names and technical phrases. Correct these consistently during review rather than assuming the model will learn them from one file.
Use timestamps and speaker labels only when the selected tool actually produces them, and verify them when accuracy matters.

Troubleshooting

The command says Whisper or FFmpeg is not found: Confirm the package is installed in the Python environment you are using and that FFmpeg is on your system’s PATH. Close and reopen the terminal after changing PATH, then check the installation instructions for your operating system.

The audio file will not load: Check its path, spelling, permissions, and format. Convert a copy to a common format such as WAV or MP3 with a trusted media tool, then try again. For API use, also check the current file-size and format restrictions.

The result is in the wrong language or is translated: Confirm you selected transcription rather than translation. Set the known spoken language using the option supported by your version, and retry a short sample.

Words are inaccurate or missing: Listen for noise, clipping, overlapping speakers, or very quiet speech. Try a larger model or a cleaner audio source, then manually correct the transcript. No model setting can reliably recover speech that is absent or unintelligible in the recording.

The API request is rejected: Check the key, account access, endpoint, model name, request format, and upload limits against current documentation. Keep secrets out of shared logs. If the service is temporarily unavailable, retry later rather than repeatedly submitting a large file.

Frequently Asked Questions

Is OpenAI Whisper free?
Whisper is available as open-source software for local use, though your computer, storage, and setup have costs. API transcription is a separate service whose current pricing and access should be checked in official documentation.

Does Whisper work without internet?
Local Whisper can process audio offline after its software and model files are installed. API transcription requires an internet connection to send the audio and receive the result.

Which Whisper model should I use?
Start with a smaller model for speed on clear audio. Try a larger model when accents, noise, or vocabulary produce errors, and compare results on a short representative clip.

Can Whisper transcribe multiple speakers?
It can transcribe speech from recordings with multiple speakers, but plain transcription may not reliably separate speakers. Choose a workflow with diarization if speaker labels are required and review them manually.

Can Whisper translate speech?
Whisper supports translation to English in its translation task. For a transcript in the language originally spoken, use transcription instead; check the current tool’s supported options.

Is the transcript accurate enough to publish?
It can be a useful draft, but accuracy varies with recording quality, language, and vocabulary. Review the audio before publishing, quoting, or relying on names, figures, or sensitive statements.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.