Whisper on Windows: Install, Transcribe a File and Check SRT Output

You have a recording on your Windows PC and need editable text or subtitles. Start with one short file: check the installation, generate the output, then listen back before processing the full recording. This guide uses the open-source Python implementation of Whisper, not an OpenAI API or an IliciLabs file-transcription feature.
Quick answer: what you need
- A local audio or video file you are allowed to process. For the first check, use a short recording without confidential material.
- Python and the
openai-whisperpackage in a dedicated environment, plus the FFmpeg command-line tool. - Internet access to install dependencies and download the chosen model initially. The file-transcription path described here loads media through local FFmpeg and runs Whisper on the PC; downloading a model is a separate operation.
- A place to save the transcript, and time to check words and subtitle timings. A generated file is a draft, not a verified record.
Whisper’s code and model weights are MIT-licensed. That does not make hardware, electricity, third-party applications or distribution services free. Local processing also does not protect a file from a synced folder, backups, other software or someone with access to the PC.
1. Prepare Python and FFmpeg on Windows
The official Whisper README documents installation and expects compatibility with Python 3.8–3.11. The example below explicitly selects Python 3.11; it does not claim that every newer Python/PyTorch combination works. If you do not have that version, obtain it from Python’s official downloads before using these commands.
Open PowerShell in a working folder. Check Python, create an environment, and use its interpreter directly:
py -3.11 --version
py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -U openai-whisper
.\.venv\Scripts\python.exe -m whisper --help
Using the interpreter path avoids needing to activate a PowerShell script. The Python venv documentation explains this arrangement. Keep the environment and the recording in your own working folder, not a Windows system directory.
Whisper also needs FFmpeg available on PATH. If you already use Chocolatey or Scoop, the Whisper README gives these alternatives; choose the one you actually have:
# Existing Chocolatey installation:
choco install ffmpeg
# Or an existing Scoop installation:
scoop install ffmpeg
These commands do not install the package manager itself. Alternatively, start from FFmpeg’s official download page, follow its Windows build links and the chosen distributor’s instructions. Open a new terminal after a PATH change and check:
ffmpeg -version
Do not continue to transcription until both whisper --help through the environment’s interpreter and ffmpeg -version work. A download button from an unrelated site is not a substitute for checking which project or distributor supplies the program.
2. Run one small, explicit transcription
Put a short English recording named sample.wav in the working folder. In the same PowerShell window, run this deliberately conservative CPU baseline:
.\.venv\Scripts\python.exe -m whisper ".\sample.wav" --model tiny --language English --task transcribe --device cpu --fp16 False --output_dir ".\output" --output_format all
Replace the quoted input path and spoken language to match your file. A path containing spaces must remain quoted. tiny is a small multilingual starting model, not a promise of acceptable accuracy. --device cpu --fp16 False makes this first run independent of a working CUDA installation; it does not promise a particular duration or support for every PC.
The selected model may download on its first use. The implementation then reuses its model cache; keep the same model available before expecting an offline run. Do not assume that installing the Python package alone has downloaded every model.
With --output_format all, the CLI requests all available output formats, including TXT, SRT and VTT. For this example, look for sample.txt, sample.srt and sample.vtt inside output. Use TXT for editable prose, SRT or VTT when the destination accepts timed subtitles. Check the target editor or player rather than assuming it accepts every format. On later runs, --output_format srt requests just SRT.
3. Decide whether the output is usable
Before running a long interview or publishing subtitles:
- Listen to the beginning, a difficult section and the end while reading the output.
- Check names, numbers, negations and speaker changes against the recording. Do not treat text alone as proof of what was said or who said it.
- Open the subtitle file in the intended editor/player and check its timing. Saving SRT does not prove synchronization with an edited video.
- Keep the original recording and make corrections in a separate copy of the output.
For a comparison, transcribe the same short extract with another model, keeping language, task and device unchanged. Record elapsed time and concrete errors you found; choose from that evidence rather than claiming that a GPU or larger model always wins. The official README’s relative speeds were measured on an A100 with English speech and are not a timing guarantee for a Windows laptop.
Transcription keeps the spoken language. Whisper’s --task translate means translation into English, not any chosen target language; the README also states that turbo is not trained for translation. This file workflow is separate from live audio capture, a floating overlay or two-way conversation translation.
4. Fix the failing step before changing models
| Symptom | What to check next |
|---|---|
py -3.11 cannot find Python | Check that Python 3.11 and the Windows launcher are installed. Do not silently run a different interpreter and assume compatibility. |
No module named whisper | Use the same .venv\Scripts\python.exe for installation and execution. Check the installation’s actual error before retrying. |
ffmpeg is not recognized, or media loading fails | Run ffmpeg -version in the same terminal, verify PATH and reopen the terminal after changes. A media-loading error also requires checking the input path and format. |
| The input cannot be found | Check the real filename, extension, working folder and quoted path. A missing executable and a missing recording are different problems. |
FP16 is not supported on CPU | The implementation falls back to FP32 on CPU. The explicit baseline already uses --fp16 False; this warning alone is not a failed transcript. |
| CUDA or GPU-memory error | Return to the CPU baseline to isolate installation/file loading from GPU configuration. For GPU use, follow the current PyTorch installation guidance; do not install random CUDA packages. |
| Output is inaccurate or repetitive | Check the audio and spoken language first, compare a short extract with another model, and review the text manually. No model choice makes an unchecked transcript authoritative. |
If package installation fails, keep the exact error and interpreter version. The Whisper README has specific notes for missing tiktoken wheels, Rust and setuptools_rust; apply the relevant fix rather than installing unrelated dependencies.
5. Choose the next step for your actual need
For a saved recording, finish the TXT/subtitle check above first. If you need captions while listening instead, read Aurora’s Windows live-caption information and verify its current workflow and requirements. The commands in this guide do not demonstrate importing recordings into Aurora.
For help scoping a repeatable workflow, contact IliciLabs with the file type, language, Windows/Python versions, the exact failing step and the output you need. Do not send a confidential recording just to describe the problem. You can also explore the current product catalogue.
Sources and scope
Checked on 26 October 2026 against the Whisper README, CLI implementation, output writers, local audio loader and Python environment documentation. Commands were checked against those sources; this editorial review did not execute a Windows installation, measure transcription speed or validate a real recording.