Whisper Timestamped

Built by the LinTO team, this Python package wraps OpenAI's Whisper so that every transcribed word comes back with its own start time, end time and confidence value. Rather than loading a second alignment network, it reads the decoder's cross-attention maps and runs a DTW pass over them, keeping memory overhead low on long recordings and avoiding a phoneme model per language. Call transcribe() much as in stock Whisper, or use the whisper_timestamped command, which writes JSON, CSV, SRT, VTT and TSV output. Greedy decoding is the default for speed; an --accurate switch restores Whisper's beam search settings. Optional extras include a Silero or Auditok speech detector to stop hallucinated text on silent stretches, a flag marking fillers and hesitations, alignment plots, and fine-tuned Hugging Face checkpoints via a transformers backend. Install with pip alongside ffmpeg; a CUDA GPU helps, a CPU-only torch build works, and Dockerfiles cover both. AGPL-3.0 permits commercial use, but a modified copy run as a network service must have its changes published. A good fit for subtitle tools, captioning and speech research.