How to transcribe audio to text locally with Whisper

You can transcribe audio to text on your own computer, for free and without uploading anything, with OpenAI's open Whisper models. We set it up on October 14, 2026 and timed it: on a desktop processor with no GPU involved, the small Whisper model transcribed a minute of English in about 11.5 seconds with no word errors, even with heavy background noise, and Spanish dialogue with a 3.2% word error rate. This guide shows the setup, which model to choose, what our test does and does not prove, and when a paid cloud service makes more sense.

Local or cloud: the quick comparison

Table: ways to transcribe audio to text (prices and limits from each vendor, checked October 14, 2026)

The local route is the one to choose for interviews, medical or legal recordings and anything you would rather not send to a third party. For the same reason, our guide to whether ChatGPT is safe explains what happens to what you upload to a cloud assistant.

How to transcribe audio locally with faster-whisper

faster-whisper is a reimplementation of Whisper on the CTranslate2 engine. Its README says it is "up to 4 times faster than openai/whisper for the same accuracy while using less memory", and unlike the original it does not need FFmpeg installed, because it decodes audio with a bundled library.

1. Install Python 3.9 or newer, then run pip install faster-whisper.
2. Save this as transcribe.py next to your audio file:

from faster_whisper import WhisperModel

Model = WhisperModel("small", device="cpu", compute_type="int8")
segments, info = model.transcribe("audio.mp3", beam_size=5)
print("Language:", info.language, round(info.language_probability, 2))
for s in segments:
    print("[%.2fs -> %.2fs] %s" % (s.start, s.end, s.text))

3. Run python transcribe.py. The first run downloads the model; after that it works offline. The language is detected automatically, and each line comes with timestamps you can use for subtitles.

int8 is 8-bit quantization, which cuts memory use with little loss (our guide to quantization explains the trade-off). To use an NVIDIA GPU, change the device to "cuda" and the compute type to "float16". That needs NVIDIA's cuBLAS for CUDA 12 and cuDNN 9 installed separately: our first GPU attempt failed with "Library cublas64_12.dll is not found", exactly the requirement the README lists. Installing them is the step most people miss.

If you prefer a command line, the original openai-whisper package works with pip install -U openai-whisper plus FFmpeg, and then whisper audio.mp3 --model turbo.

Which Whisper model to choose

OpenAI trained Whisper on 680,000 hours of audio in many languages, according to its 2022 paper, and released it in six sizes. Bigger models make fewer mistakes and run slower.

Table: Whisper model sizes (OpenAI's README; speed relative to large)

OpenAI says the English-only models do better on English, especially tiny and base, and that turbo is an optimized large-v3 "with a minimal degradation in accuracy", but it does not translate: for translating speech into English use medium or large. Accuracy also "varies widely depending on the language". Small is a sensible start on a CPU, as our test below shows; turbo, which needs about 6 GB of VRAM, is the default model of the openai-whisper command line. Our guide to how much VRAM you need covers the hardware side for larger models.

Our test: speed and accuracy on a CPU

How we tested it. faster-whisper 1.2.1 (CTranslate2 4.8.2), the small model, int8 on the CPU, beam size 5 and automatic language detection, on an Intel Core i9-12900KF. We compared each transcript with the script it was read from and counted word errors after removing punctuation and capitals.

Table: our results (October 14, 2026)

Loading the model took 1.4 seconds, and it identified both languages with a probability above 0.99. The single English error was "share channel" for "shared channel". In Spanish, two errors were real ("Ah" for "Ay" and "que" without its accent) and the third was formatting: "dos" written as a word where the script had the digit "2".

What this does not prove. The voices were synthetic: Windows' built-in English voice and Spanish voices from a commercial text-to-speech service, recorded cleanly. That is close to a best case. Real recordings with accents, people talking over each other, phone-quality audio or reverberant rooms will produce more errors, and we did not test those. For a reference on bigger hardware, faster-whisper's own benchmark transcribes 13 minutes of audio with the small model in 1 minute 42 seconds on an 8-thread Intel Core i7-12700K, and with the large-v2 model in about a minute on an RTX 3070 Ti.

When to use a cloud service instead

• You need speaker labels. Whisper's output has timestamps but no speaker names. OpenAI's gpt-4o-transcribe-diarize model returns segments with a speaker, and Word's Transcribe separates each speaker too.
• You transcribe little and want no setup. At $0.0045 a minute, an hour on gpt-transcribe costs about $0.27. Files must be 25 MB or smaller, so long recordings need compressing or splitting.
• Your computer is slow. On an older laptop, use the tiny or base model, or the batched pipeline in faster-whisper, before paying for a service.

If you are new to running models on your own computer, our guide to running an LLM locally covers the same idea for chat models, and NPU vs GPU explains which chip in a new laptop will actually do the work.