Skip to content
CAI
Software that uses CAICheck a score

openai/whisper

57.8

Adequate · 23 October 2026

3.4k

lines of production code

Python

primary language

6

measurements over time

CAI band scale
CAI trend line
CAI lens gauges

What this system is

This system is a speech-to-text service built on the Whisper model, providing multilingual transcription with word-level timing and configurable silence handling. It manages model loading, tokenization via tiktoken, and text normalization to ensure accurate output. The codebase includes Jupyter notebooks for demonstration and a comprehensive test suite to validate timing and normalization logic.

Features

Whisper 20250625: New models, word-level timing, and SDPA acceleration

This release introduces the large-v3 and large-v3-turbo models, alongside a new \in\_memory\ option for \load\_model\ to preload weights into host memory. Transcription now supports word-level timestamps via Dynamic Time Warping (DTW) and cross-attention hooks, with configurable silence handling for hallucinations. Performance is improved by using PyTorch's \scaled\_dot\_product\_attention\ (SDPA) when available, and audio loading no longer depends on the \ffmpeg-python\ package. The tokenizer has been migrated from \transformers\ to \tiktoken\, and Hebrew language codes are updated to the IANA standard 'he'.

whisper · high confidence

Behavioural changes

Migrate build system to pyproject.toml and update pre-commit hooks

The project has migrated its package metadata from setup.py to pyproject.toml (PEP 621), removing the legacy setup.py file. To support this and improve code quality, the pre-commit configuration has been updated to pin hook revisions to immutable commits and upgrade black to v25.1.0 and isort to v6.0.0. Additionally, a .flake8 configuration was added to ignore unused imports in \_\init\\_.py files, and .gitattributes was configured to mark Jupyter notebooks as generated to correct GitHub's language statistics.

(repo-wide) · high confidence

Refactored download logic and updated widget IDs in Multilingual\_ASR notebook

The Multilingual\_ASR notebook now extracts the file download logic into a dedicated \download\ function, improving code readability and maintainability. Additionally, all Jupyter widget model IDs and output IDs have been regenerated to resolve GitHub display errors, ensuring that progress bars and interactive elements render correctly when viewing the notebook on GitHub.

notebooks · high confidence

Switched tokenizer from Hugging Face GPT-2 to tiktoken

The asset files in whisper/assets have been updated to use the tiktoken library for tokenization. The previous Hugging Face GPT-2 tokenizer files (merges.txt, vocab.json, special\_tokens\_map.json, and tokenizer\_config.json) have been removed and replaced with gpt2.tiktoken and multilingual.tiktoken files. This change updates the underlying tokenization mechanism used by the model.

whisper/assets · high confidence

Fixes

Fixes to English text and number normalization

The English normalizer now preserves standalone measurements like '20mm' by removing the incorrect mapping of 'mm' to 'hmm' in the spelling dictionary. Additionally, the rule converting the digit '1' to the word 'one' has been tightened to only apply when '1' stands alone, preventing unintended changes in contexts like '$1' or '1%'.

whisper/normalizers · high confidence

Test coverage

Expanded test coverage for timing, normalization, and transcription features

Added new tests for word-level timing logic (DTW and median filter) and updated existing tests to verify English number and spelling normalization edge cases (such as standalone 'one' and 'mm' handling). The transcription tests were also enhanced to validate word-level timestamps and token consistency, and made to run on CPU when CUDA is unavailable.

tests · high confidence

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

How this codebase got here

Score

  • CAI 58 → 58 (-0.3)

Lenses

  • Code Health 91 → 91 (+0.0)
  • Architecture 100 → 100 (+0.0)
  • Maturity 47 → 46 (-0.7)
  • Readiness 50 → 50 (+0.0)
  • Security 79 → 79 (+0.0)

Resolved (2)

  • Off-boarding risk: anonymized user #1
  • Workflow token permissions not restricted

New (8)

  • Off-boarding risk: anonymized user #1
  • Orphaned file (whisper/init.py)
  • Orphaned file (whisper/audio.py)
  • Orphaned file (whisper/timing.py)
  • Orphaned file (whisper/tokenizer.py)
  • Orphaned file (whisper/triton_ops.py)
  • Workflow token permissions not restricted (.github/workflows/python-publish.yml)
  • Workflow token permissions not restricted (.github/workflows/test.yml)

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

Survey your own repository

openai/whisper was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.

About this page

  • The score is its most recent published measurement, taken on 23 October 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
  • Measured at commit 86098128c0b4f24f0e2aa2994de830614b474227 — the exact code this score is about.
  • Scored under rubric-2026.10.5 — the same rubric and the same method as every other entry in this index.
  • Measured by watchdog.canine.dev using codehealth-analyzer measure-2026.10.6+preprod-d36780c2d857.