Install the original openai-whisper Python package, pick the right model, and fix the classic errors: ffmpeg not found, CUDA out of memory, slow CPU runs, hallucinated text.
pip install -U openai-whisper — and ffmpeg is a hard requirement: brew install ffmpeg / apt install ffmpeg.
whisper input.mp3 --model turbo --output_format srt
turbo ≈ large accuracy at 8× the speed; small for CPU-only machines.
Same Whisper-class accuracy, no install, no model downloads, no GPU questions. Speaker labels, word timestamps and every export (TXT, SRT, VTT, DOCX) included. Free instant preview; credits from $2 and they never expire.
Whisper shells out to ffmpeg for decoding. Install it and make sure it's on PATH — on Windows that means restarting the terminal after install.
large-v3 wants ~10 GB VRAM. Use --model medium, or --fp16 False off GPU, or switch to faster-whisper with int8.
Whisper hallucinates on long silences and music. Pre-trim silence, or pass --condition_on_previous_text False — the loop usually stops.
The reference implementation is not optimized for CPU. Use whisper.cpp or faster-whisper for the same weights at several times the speed — or a hosted GPU run.
--word_timestamps True works but drifts on long files; WhisperX-style forced alignment is tighter if SRT precision matters.
More on OpenAI Whisper: the full OpenAI Whisper page · OpenAI Whisper alternatives · all transcription tools