# How to Transcribe Audio Locally with Whisper

> Install OpenAI Whisper and ffmpeg, run whisper file.mp3 --model turbo, and get SRT and TXT subtitle files with timestamps out on your own machine.

- Canonical: https://guides-ai.pages.dev/guides/whisper-transcribe-audio-locally/
- Plate 13.07 · Topic: Everyday tasks (https://guides-ai.pages.dev/topics/tasks/)
- Published: 06 Sept 2026 · 3 min read
- Source site: guides-ai — https://guides-ai.pages.dev/

Whisper transcribes audio offline — nothing is uploaded, and it handles about 100 languages. There are three common builds; this guide uses the reference `openai-whisper` package because it is one `pip install` and it writes SRT and TXT files itself. `faster-whisper` runs the same models roughly 4x quicker but is a Python library you have to format output with, and `whisper.cpp` is the choice when you want a single C++ binary and no Python at all.

## 1. Install

```bash
pip install -U openai-whisper
```

Installs the CLI and the Python package.

Whisper shells out to **ffmpeg** to decode audio, so install that too:

```bash
sudo apt update && sudo apt install ffmpeg   # Debian/Ubuntu
brew install ffmpeg                          # macOS
```

```powershell
choco install ffmpeg   # or: scoop install ffmpeg
```

Puts `ffmpeg` on your PATH. Without it, every transcription fails at the decode step.

## 2. Transcribe

```bash
whisper interview.mp3 --model turbo --output_format srt --output_dir out
```

Transcribes the file and writes `out/interview.srt` with timestamps. Use `--output_format txt` for plain text, or omit the flag entirely to get every format at once (txt, vtt, srt, tsv, json).

The first run downloads the model into `~/.cache/whisper`.

## 3. Pick a model size

| Model | Parameters | VRAM | Speed |
| --- | --- | --- | --- |
| tiny | 39M | ~1 GB | ~10x |
| base | 74M | ~1 GB | ~7x |
| small | 244M | ~2 GB | ~4x |
| medium | 769M | ~5 GB | ~2x |
| **turbo** | 809M | ~6 GB | ~8x |
| large | 1550M | ~10 GB | 1x |

`turbo` is the default and the best trade for most people — near `large-v3` accuracy at a fraction of the time. Drop to `small` on a machine with little memory. No GPU is fine; it just runs on CPU and takes considerably longer.

## 4. Non-English audio

```bash
whisper podcast.m4a --language German --output_format srt
```

Skips language auto-detection, which is faster and avoids the occasional wrong guess on a quiet opening.

To translate speech into English instead of transcribing it, add `--task translate` and use a multilingual model — `turbo` is not trained for translation and will hand back the original language:

```bash
whisper podcast.m4a --model medium --language German --task translate
```

Produces English text from German audio.

---

Next: [summarize a YouTube video with AI](/guides/summarize-youtube-video-with-ai/) or [run a local model to summarize the transcript privately](/guides/run-local-llm-ollama-macos-linux/).