The problem
How it works
Whisper is a speech recognition model that handles multilingual transcription, translation into English and language detection. You install the openai-whisper package and ffmpeg, then run the `whisper` command on audio files with a chosen model size, an optional `--language`, and `--task translate` for English output, or call `transcribe()` from Python.
Results and limitations
The README publishes a table of model sizes with approximate VRAM and relative speed, and says the turbo model trades little accuracy for speed.
Before you try it
Turbo is not trained for translation, so use medium or large for that. Larger models need more GPU memory, and accuracy varies with audio quality and language. Results depend on your setup and were not independently reproduced by Top 200 Repos.
Tools and repositories
- Tools
- Python · ffmpeg · NVIDIA GPU (recommended)
- Prerequisites
- No API key; runs locally. A GPU speeds up larger models.
Try it yourself: Transcribe a batch of recordings
Fill the fields to adapt the prompt. Empty fields stay as visible [placeholders].
Transcribe every audio file in [Audio folder] with Whisper. The speech is mostly in [Spoken language]; [Translate to English?]. Check that ffmpeg and a GPU are available and recommend a model size that fits this machine, explaining the speed and accuracy trade-off; wait for my choice before downloading model weights. Save a .txt and an .srt for each file in [Transcript folder] and never overwrite existing transcripts without asking. Afterwards, list any file where the detected language differed from what I expected or the output looked empty, so I can review it.Get consent before transcribing other people's recordings and check transcripts before quoting them.
Sources
- openai/whisper README (opens in a new tab)
Project documentation · github.com · Undated · Source checked 9 Nov 2026, 22:22 UTC
What this source supports
- Multilingual transcription, translation to English and language identification
- Model sizes from tiny to large and turbo with approximate VRAM
- Requires ffmpeg; CLI flags --model, --language and --task translate
- Turbo model is not trained for translation
- Code and weights under the MIT License