Records · kv:2610.00042v1

Reproduction of SYSTRAN/faster-whisper: large-v3 WER, speed and memory benchmarks

K-Veritas Team

Submitted by K-Veritas Team

Published 2026-10-18

Independent reproduction · Benchmark or leaderboard entry · Natural language processing

Reproduction of: original code

Abstract

Ran the three scripts in the repository's benchmark/ folder at the pinned commit, unchanged, with the Whisper large-v3 model on CUDA: wer_benchmark.py over all 2703 utterances of LibriSpeech clean validation (streamed from the Hugging Face hub, normalized with the bundled English normalizer), speed_benchmark.py on the bundled benchmark.m4a (3 repeats of 10 transcriptions, minimum time reported) and memory_benchmark.py with --gpu_memory (peak GPU memory and power while transcribing benchmark.m4a). The tests/ and docker/ folders are withheld as hash-only entries. CTranslate2 4.8 with cuBLAS and cuDNN 9 from PyPI, datasets 3.6. One NVIDIA RTX 5060 Ti.

Sealed report

Cite this record

@misc{kveritas261000042v1,
  title={Reproduction of SYSTRAN/faster-whisper: large-v3 WER, speed and memory benchmarks},
  author={K-Veritas Team},
  year={2026},
  howpublished={K-Veritas Records},
  note={kv:2610.00042v1},
  url={https://kveritas.org/records/2610.00042v1},
}

References

1. [1] Mamadou K. Keita and Christopher Homan. Computer Science Conferences Should Require Nonrepudiable Experimental Results. NeurIPS 2026 Position Paper Track. arXiv:2605.08586.

Sealed with K-Veritas [1]. Listed, not endorsed. A record shows these results came from the sealed code, unchanged. Whether the work is sound is for the reader to judge.