openai / whisper

This is a Python repository: platform metadata records Python as the primary language, with 160526 Python language bytes.

The GitHub description reads "Robust Speech Recognition via Large-Scale Weak Supervision" — the author's own wording, not an analyzed fact.

It is MIT licensed; the latest recorded release is v20250625 (published 2025-06-26), and the metadata snapshot counted 109056 stargazers.

Code sits in a top-level whisper package directory next to a tests directory; the analyzer's summary reports 45 analyzed files, 18 of them under whisper.

The one detected script entrypoint is in whisper/transcribe.py:

if __name__ == "__main__":
    cli()

The static core flow is short: cli() (lines 517-619) builds an argparse parser through repeated parser.add_argument calls and checks torch.cuda.is_available.

All six bounded execution paths from __main__ end at these external or unresolved boundaries, so nothing beyond argument setup is traced.

Role analysis labels whisper/transcribe.py an entry/orchestration candidate (3 incoming, 18 outgoing static calls).

• Relation totals: 1388 static edges — 1207 calls, 181 imports; 155 resolved locally, 1233 external or unresolved.
• whisper/model.py: 22 incoming / 20 outgoing — the most connected module, labeled a service/core candidate.
• whisper/timing.py (16/9), whisper/utils.py (14/2), whisper/tokenizer.py (12/2), whisper/decoding.py (11/13), whisper/audio.py (8/5): also service/core candidates.
• whisper/__init__.py (3/4) and whisper/normalizers/english.py (3/3) are smaller service/core candidates; tests/test_timing.py (0 in / 6 out) is an orchestration candidate.
• whisper/transcribe.py (3 in / 18 out) is the entry/orchestration candidate.
• The pattern suggests the often-called modules (model, timing, utils, tokenizer) form a library core that transcribe.py drives — an inference from call degree, not runtime behavior.

• Dependency evidence: requirements.txt lists seven runtime dependencies — numba, numpy, torch, tqdm, more-itertools, tiktoken, triton (>=2.0.0, the only version floor).
• The repository holds two manifests (pyproject.toml, requirements.txt), but only requirements-style files are parsed, so pyproject.toml dependency tables are invisible to this analysis.
• Tests: 7 test or test-like files, including tests/conftest.py; tests/test_timing.py alone defines test_dtw, test_dtw_cuda_equivalence, test_median_filter, and test_median_filter_equivalence.
• Install and run guidance here is partial and installation commands are unverified — inspect the manifests before installing.

1. README.md — orientation.
2. pyproject.toml and requirements.txt — intended setup and dependencies.
3. LICENSE — terms.
4. whisper/transcribe.py — read transcribe() (lines 38-514), then the cli() parser (517-619).
5. whisper/__init__.py — load_model(), available_models(), _download().
6. whisper/model.py — the Whisper class, to connect model construction with transcribe().

The first three steps follow the analyzer's project-level important-file list; whisper/transcribe.py is flagged as a likely entrypoint.

• Static only: reflection, runtime dependency injection, dynamic imports, monkey-patching, generated code, framework runtime wiring, and dynamic dispatch are not resolved.
• 1233 of 1388 static relations are external or unresolved.
• Install/run/build inference is partial, and installation commands were not verified.
• README extraction is line-based and may capture code lines instead of prose, so README-derived statements are author-claimed, not established.
• Manifest fields describing the analyzer's own capabilities are excluded from repository evidence.
• No runtime flow — CLI transcription, model loading, GPU use — is established end to end.

• whisper/model.py: Whisper (259-352), AudioEncoder (181-211), TextDecoder (214-256), MultiHeadAttention (81-146), ResidualAttentionBlock (149-178).
• whisper/decoding.py: detect_language, DecodingOptions, DecodingResult, GreedyDecoder, BeamSearchDecoder.
• whisper/audio.py: load_audio, pad_or_trim, mel_filters, log_mel_spectrogram; whisper/tokenizer.py: Tokenizer, get_encoding, get_tokenizer.
• whisper/timing.py: median_filter, dtw_cpu, dtw_cuda, find_alignment, add_word_timestamps.
• whisper/utils.py: ResultWriter, WriteTXT, SubtitlesWriter, WriteVTT, plus format_timestamp and str2bool.

The AudioEncoder/TextDecoder naming suggests an encoder-decoder split between audio input and text output — a structural inference, not confirmed by execution.