By the numbers

A quantitative snapshot of the codebase. Data collected on 2026-04-30 against commit 04f449b8a437f1bbd3dba5c9f826aca972e7709a on branch main.

Size

The Python package is small. Most of the repository's bytes are in pretrained tokenizer assets and a couple of notebooks.

File / area	Lines
whisper/decoding.py	826
whisper/transcribe.py	623
whisper/normalizers/english.py	550
whisper/tokenizer.py	395
whisper/timing.py	388
whisper/model.py	345
whisper/utils.py	318
whisper/__init__.py	161
whisper/audio.py	157
whisper/triton_ops.py	119
whisper/normalizers/basic.py	77
All whisper/*.py	~3,400
All tests/*.py	~210
Total Python	4,267

xychart-beta horizontal
    title "Python lines per file (top 10)"
    x-axis ["decoding.py", "transcribe.py", "english.py", "tokenizer.py", "timing.py", "model.py", "utils.py", "__init__.py", "audio.py", "triton_ops.py"]
    y-axis "Lines" 0 --> 900
    bar [826, 623, 550, 395, 388, 345, 318, 161, 157, 119]

Category	Count
Python source files (whisper/)	14
Python test files (tests/)	6
GitHub workflows	2
Notebooks (notebooks/)	2
Top-level docs (README.md, CHANGELOG.md, model-card.md, data/README.md)	4

The repository carries some sizable non-source assets:

Asset	Size	Purpose
whisper/assets/gpt2.tiktoken	836 KB	English-only BPE merges
whisper/assets/multilingual.tiktoken	817 KB	Multilingual BPE merges
whisper/assets/mel_filters.npz	4.2 KB	Precomputed 80- and 128-channel mel filterbanks
tests/jfk.flac	1.2 MB	Test fixture (JFK inaugural clip)
approach.png	904 KB	Architecture diagram embedded in README.md
language-breakdown.svg	273 KB	Per-language WER chart referenced from README.md
notebooks/Multilingual_ASR.ipynb	5.7 MB	Demo notebook

Activity

Commits per calendar year on main (commits authored, including merges):

xychart-beta horizontal
    title "Commits per year"
    x-axis ["2022", "2023", "2024", "2025", "2026 (YTD)"]
    y-axis "Commits" 0 --> 90
    bar [57, 79, 14, 16, 2]

Cumulative as of 2026-04-30: 168 commits across 167 entries in git log. Activity peaked in 2023 and has been mainly maintenance (CI, dependencies, security fixes) since 2024.

The most-changed Python files in the repository's history (by total commits touching the file):

File	Commits touching it
whisper/transcribe.py	~30
whisper/decoding.py	~15
whisper/__init__.py	~15
whisper/tokenizer.py	~12
whisper/model.py	~12
whisper/timing.py	~10

Bot-attributed commits

git log on main shows 0 commits with dependabot[bot] or factory-droid[bot] co-authorship. GitHub Actions for Dependabot is configured (.github/dependabot.yml schedules weekly action updates), but those PRs land as human-authored commits in this history. This number is a lower bound on AI-assisted changes since inline AI tools leave no git trace.

Complexity

Largest functions and classes (qualitative; based on file scan):

• transcribe() in whisper/transcribe.py: ~280 lines. Implements the file-level loop, fallback, segment slicing, and (when enabled) the word-timestamp + hallucination-silence heuristics. The longest single function in the repo.
• cli() in whisper/transcribe.py: ~80 lines, mostly argparse setup.
• DecodingTask in whisper/decoding.py: ~250 lines spread across init, _main_loop, run, and helpers.
• BeamSearchDecoder in whisper/decoding.py: ~120 lines.
• EnglishNumberNormalizer in whisper/normalizers/english.py: ~400 lines, dominated by hard-coded number-word and currency tables.

Test-to-code ratio: 6 test files / 14 source files (~0.43). Coverage is depth-oriented (one or two strong tests per module: tokenizer, audio, timing, normalizer, transcribe) rather than breadth-oriented.

Languages

The repository is essentially mono-language. ~98% of source-tree lines are Python. The remaining lines are YAML (CI / pre-commit), TOML (pyproject.toml), Markdown, and one .flake8 config.

xychart-beta horizontal
    title "Source lines by language"
    x-axis ["Python", "Markdown", "YAML", "TOML"]
    y-axis "Lines" 0 --> 4500
    bar [4267, 320, 130, 50]

(Line counts for Markdown / YAML / TOML are approximate.)

Dependencies

Runtime dependencies declared in pyproject.toml:

Dependency	Notes
more-itertools	Used in whisper/normalizers/english.py (windowed)
numba	JIT for CPU DTW / backtrace
numpy	Array math
tiktoken	BPE tokenizer
torch	The model
tqdm	Progress bars in download and transcribe
triton>=2	x86_64 Linux only; CUDA kernels for DTW + median filter

Dev extras: black, flake8, isort, pytest, scipy.

The transitive footprint is dominated by torch. There is no networking client, web framework, or database driver — everything is local computation plus an ffmpeg subprocess.