Systems

The whisper package decomposes into a handful of internal modules, each with a clear scope. This section documents them one by one. Most are a single file.

At a glance

graph TD
    audio[audio<br/>whisper/audio.py] --> transcribe
    tokenizer[tokenizer<br/>whisper/tokenizer.py] --> transcribe
    tokenizer --> decoding
    model[model<br/>whisper/model.py] --> decoding
    decoding[decoding<br/>whisper/decoding.py] --> transcribe
    timing[timing<br/>whisper/timing.py] --> transcribe
    triton[triton_ops<br/>whisper/triton_ops.py] --> timing
    transcribe[transcribe<br/>whisper/transcribe.py] --> writers
    writers[output writers<br/>whisper/utils.py]
    norm[normalizers<br/>whisper/normalizers/]

normalizers does not sit on the inference path — it is used for post-hoc evaluation (paper-style WER scoring).

Pages

Subsystem	Source	Page
Audio frontend	whisper/audio.py	Audio processing
Transformer model	whisper/model.py	Model
Decoding and sampling (window-level)	whisper/decoding.py	Decoding
Long-form transcription loop	whisper/transcribe.py	Transcribe
Word-level timing via DTW	whisper/timing.py, triton_ops.py	Timing
Tokenizer	whisper/tokenizer.py	Tokenizer
Output writers (txt/vtt/srt/tsv/json)	whisper/utils.py	Output writers
Text normalization (eval only)	whisper/normalizers/	Normalizers

Reading order

If you are new to the codebase, the most useful order is:

1. Audio processing — what the model eats.
2. Tokenizer — how text is represented.
3. Model — the actual neural network.
4. Decoding — how a single 30 s window is decoded.
5. Transcribe — how an arbitrary file is processed.
6. Timing — how word timestamps are produced.
7. Output writers — how results are serialized.
8. Normalizers — only relevant if you are doing paper-style evaluation.