whisper

Install openai-whisper and ffmpeg, then load a model with whisper.load_model("base") and call model.transcribe("audio.mp3"). The result dictionary contains the full text plus timestamped segments you can iterate over.

Use turbo for the best speed-to-quality balance, base for quick prototyping, and large for maximum accuracy on low-resource languages. Tiny and base models need about 1 GB VRAM, while large requires roughly 10 GB.

Yes, Whisper supports translation to English from any of its 99 supported languages by passing task="translate" to the transcribe call. The output is English text regardless of the source audio language.

The standard openai-whisper package is not designed for low-latency live captioning. For streaming use cases, the documentation recommends faster-whisper, which runs about 4x faster with reduced latency.

Whisper can hallucinate or repeat text, accuracy degrades on audio longer than 30 minutes, and it provides no speaker diarization. Background noise and strong accents also reduce transcription quality.

Run the CLI command whisper video.mp4 --output_format srt to produce a subtitle file directly. Whisper handles common media formats through ffmpeg, so no separate audio extraction step is required.