What Is Whisper? Transcribe Audio to Text: Setup, Turkish Support, API Pricing and Choosing a Model

You install Whisper with Python, give it an audio recording and get the transcribed text back; it works in many languages including Turkish, detects the language and translates speech into English. It is OpenAI's open-source model and Python package; the software is free (MIT license) and runs on your computer, needs the command line and Python, and has no official desktop app. Hardware costs and updates are on you. OpenAI also offers a paid API under the same name: it is billed in dollars per minute and files are limited to 25 MB.

What it does: Transcribes audio to text, detects the language
Free: Locally yes; the API is paid
Turkish: Transcribes Turkish
Where: On your own computer; OpenAI API
License: MIT, commercial use allowed
Latest release: 20250625 (June 2025)
Price: Software $0 · API $0.0045/min

What does Whisper do?

Whisper is a general-purpose speech recognition model trained by OpenAI: it listens to an audio or video recording and writes down what is said. The same model can identify the language and translate speech in other languages into English. The model was trained on 680,000 hours of audio and text collected from the internet.

This page covers two separate things, because both are called “Whisper”: the open-source package (runs on your own computer, free) and the paid speech-to-text service in the OpenAI API. For most people one of the two is enough; see the first box to tell them apart.

Transcribe audio and video recordingsWhat you get: txt, vtt, srt, tsv and json filesNeeds a separate setup Generate a subtitle file for a videoWhat you get: srt or vtt fileNeeds a separate setup Translate speech into English textWhat you get: English textNeeds a separate setup Transcribe audio files with the OpenAI APIWhat you get: Text (and subtitles or timestamps if requested)Needs a connection

Who it helps: teams and content creators who want to transcribe meetings, lectures and interviews, don't want to upload recordings to the cloud, and can use the command line. Who it doesn't help: people who don't want to use the command line (there is no official interface) and anyone who needs real-time, live transcription: the model card says Whisper is not designed for real-time transcription out of the box.

On your own computer (command line and Python)Turn audio and video recordings into text, subtitles and JSON without them leaving your computer.Terminal, Desktop · Requires Python and ffmpeg; the software is free, hardware costs are on you · details

OpenAI API (cloud, paid)Have audio files transcribed on OpenAI's servers; for connecting it to your own software.Web, Terminal · Requires an API key and credit; billed in dollars per minute · details

How do you install Whisper? Windows, Mac, Linux, ffmpeg

Can I use it without code?

The official Whisper repository offers only the command line and Python; OpenAI has no desktop Whisper app. There are third-party interfaces, desktop apps and Google Colab notebooks built on Whisper, but they are not from OpenAI; this guide does not recommend them and has not vetted them all. If you choose a third-party app, you will be uploading your audio to it, so read its privacy terms before deciding.

Setup has three parts: the Python package, ffmpeg (the program that reads audio files) and, if needed, Rust. According to the official README, the code was written to be compatible with Python 3.8–3.11; the PyPI page also lists 3.12 and 3.13, and which ones run without issues is not clear in the official sources.

Latest published release› pip install -U openai-whisper

Newest code in the repository (including changes after the release)› pip install git+https://github.com/openai/whisper.git

The latest release on PyPI is 20250625 (June 26, 2025). New changes landed in the repository through August 2026 (for example, the JSONL output format in July 2026), but no new release came out; the second command gets them.

Model files are downloaded from the internet on first run and saved to the ~/.cache/whisper folder by default (stated in the official code). First use requires an internet connection.

Whisper reads audio files with ffmpeg, so it must be installed on your computer. The commands from the README:

Windows (Chocolatey)› choco install ffmpeg

Windows (Scoop)› scoop install ffmpeg

macOS (Homebrew)› brew install ffmpeg

Ubuntu or Debian› sudo apt update && sudo apt install ffmpeg

Arch Linux› sudo pacman -S ffmpeg

If there is no prebuilt package of the fast tokenizer tiktoken for your computer, the pip install fails and Rust is required. In that case the README tells you to install the Rust development environment, set your PATH variable, and, if you see “No module named setuptools_rust”, run:

On a setuptools_rust error› pip install setuptools-rust

Source: Whisper README “Setup” and the PyPI page, October 14, 2026. Python version and Windows setup details may vary with your computer.

What computer does Whisper run on? Does it need a GPU?

The official README gives the graphics card memory (VRAM) each model needs (see the table in the Which model section): about 1 GB for tiny and base, 2 GB for small, 5 GB for medium, 6 GB for turbo and 10 GB for large.

Is a GPU required? If --device is not given, the command-line tool uses a CUDA graphics card if one is present, otherwise the processor (CPU) (source code). So it runs without a graphics card; how slow it is on a CPU is not stated in the official sources.
Operating system: the README gives ffmpeg commands for Windows, macOS, Ubuntu/Debian and Arch. The README does not cover Apple silicon or AMD graphics card support, and nothing is guessed here.
Python: the README says 3.8–3.11; PyPI lists 3.8–3.13. Torch and numba are installed as dependencies.

If speed and memory are a concern, you have two options: pick a smaller model, or look at separate projects built for speed (see Faster Whisper and WhisperX).

Source: Whisper README, PyPI, transcribe.py.

How do you use Whisper? Command line and Python

After setup, open a terminal (Command Prompt or PowerShell on Windows) and run the whisper command in the folder that contains the audio file. The default model is turbo; output files are written next to the file.

Open the terminal in the folder that contains your audio file.
Run the whisper command with the file name and, optionally, a model or language (examples below).
When it finishes, .txt, .vtt, .srt, .tsv and .json files appear in the same folder. You choose which one you want with --output_format; if you don't, all of them are written.
Always check the output by listening: the model can write sentences that were never said during silent parts (see Common problems).

Transcribe a file (default model: turbo)› whisper audio.mp3 --model turbo

Set the language manually (Turkish speech)› whisper speech.wav --language Turkish

Subtitle file only (srt)› whisper video.mp4 --language Turkish --output_format srt

Translate speech into English (medium or large, not turbo)› whisper speech.wav --model medium --language Turkish --task translate

From Python

› import whisper

model = whisper.load_model("turbo")
result = model.transcribe("audio.mp3")
print(result["text"])

Type whisper --help to see all options. If you don't pass --device, it uses a graphics card (CUDA) if available, otherwise the processor.

Commands from the README “Command-line usage” section; the --output_format and --device options from the transcribe.py source (October 14, 2026). Whisper processes the file with a sliding 30-second window (README).

Transcribing audio files with the OpenAI API: file limit, pricing, privacy

If you don't want to install anything on your own computer, or you want to connect it to your software, you can send audio files to the OpenAI API. This is a separate service from the open-source package: it is paid and the file goes to OpenAI's servers.

File limit: 25 MB. Formats: mp3, mp4, mpeg, mpga, m4a, wav, webm. If it's larger, convert it to a compressed format or split it into chunks without cutting mid-sentence.
Recommended model: gpt-transcribe (for recorded files). You can improve results with prompt (information about the recording), keywords (words likely to appear in the recording) and languages (expected languages).
Timestamps, subtitles (srt/vtt) and translation into English: the guide points to the whisper-1 model for these tasks. For speaker labeling there is gpt-4o-transcribe-diarize.
Price: gpt-transcribe $0.0045/minute, whisper-1 $0.006/minute.

With curl (example from the OpenAI guide)

› curl --request POST \
  --url https://api.openai.com/v1/audio/transcriptions \
  --header "Authorization: Bearer $OPENAI_API_KEY" \
  --header 'Content-Type: multipart/form-data' \
  --form file=@/path/to/file/audio.mp3 \
  --form model=gpt-transcribe

Four older speech-to-text models in the API, including whisper-1, were announced for deprecation on August 26, 2026 and will be removed from the API on February 26, 2027; OpenAI recommends gpt-transcribe or gpt-live-transcribe instead. But according to the guide, timestamps and translation are only available in whisper-1; which API model will handle these tasks after its removal is not stated in the official sources. If you have software that depends on these tasks, note the date.

What happens to your data in the API?

Since March 1, 2023, data sent to the API is not used for model training unless you explicitly opt in (OpenAI data controls page).
According to the table on the same page, for the speech-to-text (/v1/audio/transcriptions) and translation endpoints: training use no, abuse monitoring retention no, application state retention no.
The file still goes to OpenAI's servers. Before uploading customer or employee audio, check your own privacy obligations and those under KVKK (Turkey's Personal Data Protection Law). This is not legal advice.

Sources: OpenAI API documentation (File transcription, model pages, pricing, deprecations, data controls), October 14, 2026. Tax and payment terms for API payments from Turkey and the account setup steps are not clear in the official sources.

Which model? tiny, base, small, medium, large and turbo compared

Whisper comes in six sizes; the larger the size, the higher the accuracy, but speed and memory requirements get worse. Four of them also have an English-only .en version. The table is from OpenAI's README.

Size	Parameters	English-only version	Required VRAM	Speed (large = 1x)
tiny	39 M	tiny.en	~1 GB	~10x
base	74 M	base.en	~1 GB	~7x
small	244 M	small.en	~2 GB	~4x
medium	769 M	medium.en	~5 GB	~2x
large	1,550 M	none	~10 GB	1x
turbo	809 M	none	~6 GB	~8x

According to the README, speed was measured with English speech on an A100 card; in practice it varies a lot with language, speaking rate and hardware. The model card gives turbo's parameter count as 798 M, the README as 809 M: the two official pages differ. VRAM = graphics card memory.

Which one to pick

Transcribing Turkish: start with large (officially large-v3); if you need speed, turbo. The README gives a per-language accuracy chart only for large-v3; Turkish accuracy for the smaller models is not in the official sources. Compare the two models on your own recording.
Translating into English: medium or large. The README states that turbo was not trained for translation and returns text in the original language even with --task translate.
Low-powered computer: small or medium; tiny and base are fast but less accurate.
English-only recordings: the .en models (especially tiny.en and base.en) usually do better.

Language support

The OpenAI guide says Whisper supports 98 languages and that accuracy varies by language. The language list is in the tokenizer.py file in the source code. Turkish is on the list (code: tr).

Translation

Whisper translates only into English (for example Turkish → English). It does not translate from English into Turkish or between two other languages.

Versions

Per the model card: original series September 2022, large-v2 December 2022, large-v3 November 2023, large-v3-turbo September 2024. In the code, the name large maps to large-v3 and turbo to large-v3-turbo.

Turkish accuracy (README chart)

The chart in the README shows the word error rate (WER, lower is better) for large-v3: Turkish 12.0% on Common Voice 15 and 6.7% on FLEURS. For comparison, English is 9.3% and 4.1%. Because these are two different datasets, the numbers differ from each other; noise and accent in your own recording will change the result.

This section was checked on October 14, 2026 (Whisper README and model card, __init__.py). The Turkish values were read from the chart image in the README.

Is Whisper free? Cost locally and in the API

Short answer: the software is free, the machine you run it on is not. Open-source Whisper (code and model weights) is released under the MIT license: downloading and using it is free. But you cover the computer, graphics card, electricity and maintenance, and updates are your responsibility too. OpenAI's API, on the other hand, is billed in dollars per minute.

On your own computer$0

MIT license, commercial use included
Audio stays on your own machine
Command line and Python

OpenAI API · gpt-transcribe$0.0045 / minute of audio

Officially recommended model
Files up to 25 MB

OpenAI API · Whisper (whisper-1)$0.006 / minute of audio

The model the guide names for timestamps, subtitles and translation into English
Will be removed from the API on February 26, 2027

View as a table (OpenAI API speech-to-text pricing)

Model	Used for	Price
gpt-transcribe	Transcribing recorded files and real-time sessions (recommended)	$0.0045 / minute
Whisper (whisper-1)	Transcription, timestamps, translation into English	$0.006 / minute
gpt-live-transcribe	Live transcription (real time)	$0.017 / minute

What is the real cost?

Software: $0. Update with pip install -U openai-whisper; for the newest code in the repository, install from the git URL.
Hardware: a graphics card (GPU) is not required but has a big effect on speed; the memory you need depends on the model size (see What computer). If you run it on a server, rental or electricity costs are not in the official sources and are not estimated here.
API: dollars per minute; a Turkish lira price and whether VAT (KDV) or other taxes apply to API payments are not stated in these sources. Go by the amount on the checkout screen.
Time: a large model can take a long time on a CPU; speed depends on your computer.

Prices from the OpenAI API pricing page and model pages on October 14, 2026. Prices and model names change often; check the current page before paying.

Does Whisper support Turkish?

Transcribing Turkish speechYes	Turkish is on the language list (code: tr). The README chart shows the Turkish word error rate for large-v3: 12.0% on Common Voice 15 and 6.7% on FLEURS (lower is better). Turkish figures for the smaller models are not in the official sources.
Translating Turkish into EnglishYes	With `--task translate`, using the medium or large model (turbo does not translate).
Translating into Turkish or another languageNo	Whisper translates only into English. To translate from English into Turkish you need another tool.
Turkish interfaceNo	There is no official interface: command line and Python. Output file formats are language-independent.
Language detectionYes	If no language is given, the model tries to detect it; if you know the recording's language, passing `--language Turkish` is safer.

In Turkish recordings, errors in names, brands and abbreviations are normal: read the output and correct it. The quality of Turkish spelling and punctuation varies with the recording and the model; the official sources do not give a figure for it.

Source: Whisper README and tokenizer.py, October 14, 2026. The Turkish figures were read from the README chart.

What happens to my data? Privacy locally and in the API

If you run it locally

Your audio file never leaves your computer: the model runs there. The only outside connection is downloading the model on first use (from OpenAI's server). The security of your audio and text files is the security of your machine.

If you use the OpenAI API

The file is sent to OpenAI's servers. According to OpenAI, since March 1, 2023 API data has not been used for training without permission; for the speech-to-text endpoints, abuse monitoring retention and application state retention are listed as “no” in the table. If you use a third-party app, its terms apply.

Whose voices?

The people speaking in a recording are personal data. The model card warns against transcribing recordings made without the speakers' consent and mentions the risk of use for surveillance.

KVKK

If you transcribe conversations with customers, patients or employees, you may have notice and consent obligations. Running it locally does not send data abroad; sending it to the API may count as a transfer. This is not legal advice.

Before transcribing a recorded conversation, make sure the speakers know and have agreed. Don't use Whisper output on its own for high-risk decisions (hiring, credit, medical decisions): the model card advises against use in these areas.

Source: Whisper model card, OpenAI API data controls page, October 14, 2026.

Can I use Whisper's output in my business? License

Whisper's code and model weights are released under the MIT license. MIT allows you to use, modify, distribute and use it in commercial products; in return it asks you to keep the copyright notice and the license text. It gives no warranty.

Allowed

Using text you transcribed with Whisper in your business (reports, subtitles, content)
Building Whisper into your own product or service
Modifying and distributing the code

Not allowed

Removing the license text and copyright notice
Expecting a warranty: the software is provided “as is”
Transcribing recordings made without consent (the model card advises against it)

The MIT license sets no separate condition on rights to the output. If you use the OpenAI API, the OpenAI Services Agreement (January 1, 2026) states that the customer owns the output; the individual Terms of Use carry the same wording.
The model card recommends not using Whisper in high-risk decision-making or to infer personal characteristics.

Source: Whisper LICENSE and model card, OpenAI Services Agreement, October 14, 2026. This is not legal advice; for a commercial project, read the license yourself.

Common problems and fixes

What happens	What to do
“ffmpeg not found” error, or the audio file won't open	ffmpeg must be installed and on your `PATH` (README). Install commands are in the Setup section.
`tiktoken` build error during `pip install`, or “No module named setuptools_rust”	Install Rust, add it to `PATH`, then run `pip install setuptools-rust` (README).
Sentences that were never said appear during silence (hallucination)	The model card says this can stem from noisy training data and may be worse in low-resource languages. Check the output by listening. Reported by users The same sentence being written over and over during silent parts (sometimes in another language) is frequently reported on GitHub; there is no official fix.
The same sentence is written over and over	Model card: beam search and temperature scheduling reduce it but don't fully solve it.
Translating with turbo leaves the text untranslated	turbo was not trained for translation; use `--model medium` or `large` (README).
It runs very slowly	Without a graphics card the processor is used. Pick a smaller model or turbo (README speed table). There is no official speed figure for CPUs.
API: file too large	The API file limit is 25 MB: convert to a compressed format or split into chunks (OpenAI guide).
Long wait on first run	The model file is downloaded from the internet on first use and saved to the ~/.cache/whisper folder; with a large model the download takes time.

Source: Whisper README, model card and source code; GitHub Discussions (#2856, #2608, #1873) Reported by users reports. October 14, 2026.

Faster Whisper, WhisperX and ChatGPT: how are they different from Whisper?

Two tools built on Whisper and developed as separate projects are widely searched for. Details are in their own guides; here is only the difference:

The difference from ChatGPT: ChatGPT is a chat assistant; Whisper is only a model that converts audio recordings into text. Both come from the same company but are separate tools; ChatGPT's own voice features are not the topic of this page.

Source: the faster-whisper and whisperX repository READMEs and GitHub information (October 14, 2026). Setup and hardware details for the two are not covered in this guide.

Frequently asked questions

What is Whisper and what is it used for?

OpenAI's open-source speech recognition model and Python package. It transcribes audio or video recordings, detects the language and translates speech into English. The software is MIT-licensed and free.

How do you use Whisper?

After setup you type a command such as "whisper audio.mp3 --model turbo" in the terminal; the output is written as txt, srt, vtt, tsv and json files. It can also be used from Python with whisper.load_model and transcribe.

How do you install Whisper on Windows and Mac?

Install the Python package with pip install -U openai-whisper, plus ffmpeg (choco or scoop on Windows, brew on Mac). If there is no prebuilt package for tiktoken, you may need Rust and setuptools-rust.

Does Whisper support Turkish?

Yes, it transcribes Turkish speech. The README chart shows a Turkish word error rate for large-v3 of 12.0 percent on Common Voice 15 and 6.7 percent on FLEURS. It can also translate Turkish and other languages into English; it cannot translate from English into Turkish.

Is Whisper free?

Running it locally is free (MIT), but hardware costs and updates are on you. The speech-to-text service in the OpenAI API is paid: gpt-transcribe costs 0.0045 dollars and the older model called Whisper 0.006 dollars, per minute.

Which models does Whisper offer, and which should I choose?

tiny, base, small, medium, large and turbo. The bigger the model, the higher the accuracy, and the worse the speed and memory use. Use medium or large for translation and turbo for speed; turbo does not translate.

Does Whisper need a GPU to run?

Not required: the command-line tool uses the processor if there is no graphics card. But it can be much slower, and there is no official figure for CPU speed. Required graphics memory ranges from 1 to 10 GB depending on the model.

What is the OpenAI API audio file size limit?

Files can be up to 25 MB; mp3, mp4, mpeg, mpga, m4a, wav and webm are accepted. If the file is larger, compress it or split it into chunks.

Is the Whisper API being discontinued?

OpenAI announced that four older speech-to-text models, including whisper-1, will be removed from the API on February 26, 2027. gpt-transcribe or gpt-live-transcribe are recommended instead. The open-source package is not affected.

Does Whisper send my data to OpenAI?

If you run it locally, your audio file never leaves your computer; the only exception is the initial model download. If you use the OpenAI API, the file goes to OpenAI's servers; since March 1, 2023 API data has not been used for training without permission.

Whisper says “ffmpeg not found”. What should I do?

ffmpeg must be installed and on your PATH. Install commands are in the README: choco or scoop on Windows, brew on Mac, apt on Ubuntu.

What are Faster Whisper and WhisperX, and how do they differ from Whisper?

Separate projects built on Whisper. According to its README, Faster Whisper runs the same model up to 4 times faster with less memory; WhisperX adds word timestamps and speaker diarization. They have separate guides.

Tasks and setup kits that use Whisper

Adding automatic captions to a Reels video

Related guides

faster-whisper WhisperX

Harfiyen

VoiceStudio

Cloudflare Workers AI DaVinci Resolve MCP

Sources

Not clear in the official sources: Speed on a CPU (no official figure) · Apple silicon and AMD graphics card support (not covered in the README) · whether it runs without issues on Python 3.12 and 3.13 (the README says 3.8–3.11, PyPI lists up to 3.13) · Turkish accuracy of the smaller models (the README only gives a large-v3 chart; values were read from the image) · turbo's parameter count (README 809 M, model card 798 M) · which API model will handle timestamps and translation after whisper-1 is removed · taxes and a Turkish lira price for API payments from Turkey · the security of third-party interfaces (not vetted). Each is covered in its section along with what you can do.