Whisper

OpenAI trained Whisper on a large and varied audio collection using weak supervision, then published the code and weights for local use. One encoder-decoder Transformer handles transcription in many languages, translation of foreign speech into English, language detection and voice activity detection, steered by special task tokens fed to the decoder. Six checkpoints are published: tiny (39M parameters, about 1 GB of VRAM), base, small, medium, large (1.55B, about 10 GB) and turbo, a faster derivative of large-v3 that should not be used for translation; the four smallest also come in English-only form. Install the openai-whisper package from PyPI; ffmpeg must be on the PATH, and Rust may be needed if tiktoken has no prebuilt wheel for your platform. The whisper command writes txt, srt, vtt, tsv, json or jsonl files with optional word timings, and Python code can call language detection and decoding directly. CPU works for small models, while a CUDA GPU makes larger ones practical. Code and weights are MIT licensed, so commercial use is allowed. A sensible baseline for offline transcription or subtitles.