Multilingual speech recognition library adding word-level timestamps, confidence scores, and VAD to OpenAI Whisper.
Extract word alignments and confidence scores from audio without extra neural alignment models.
whisper-timestamped is a Python wrapper for OpenAI's Whisper speech recognition models. It extracts word-level timestamps and confidence scores directly from cross-attention weights. Technical professionals and automated workflows can use it to build timestamp-accurate transcriptions for meetings, interviews, and media archives.
This package modifies Whisper's transcription pipeline to compute exact start and end times for every spoken word. It includes integrated Voice Activity Detection to prune silence before decoding, preventing repetitive text hallucinations on silent audio. Additionally, it outputs confidence scores per word and handles speech disfluencies without requiring separate alignment networks.
This tool suits software engineers, technical executive assistants, and media specialists who need word-level audio alignment for subtitles, indexing, or automation scripts. It does not suit non-technical users looking for an out-of-the-box graphical desktop application.
Uses Dynamic Time Warping on cross-attention weights to calculate start and end times for individual spoken words.
Filters out silence using Silero or Auditok VAD prior to transcription, preventing hallucinated text output.
Assigns explicit reliability metrics to each word and segment to highlight potential transcription errors.
Computes alignments on the fly during decoding rather than running separate alignment networks, saving memory.
The package is free under the AGPL-3.0 licence. Running costs depend entirely on your local GPU hardware or cloud server deployment.
880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
These reviewed AI tools do the same job instantly — from our scored directory.
Standard Whisper provides approximate 1-second segment timestamps. whisper-timestamped extracts exact timing for each individual word.
No. It computes alignment from cross-attention layers during inference, avoiding external wav2vec models.
Yes. CPU execution is supported via PyTorch CPU packages or the provided CPU Dockerfile, though processing will take longer.
See Zekai first in Google