openai /

Whisper

Live map
View repository
Snapshotmain · 5f86d1d

Log-Mel windows pass through one encoder-decoder that detects language, transcribes or translates, and aligns timestamps.

Click a building or route
Choose a path

Prepare audio, decode tokens, align them, and advance the segment loop.

System map7 components · 9 connections
130%
Whisper: audio becomes multitask decoder tokens architectureFollow audio preprocessing across the top, then descend through windowed decoding and timestamped text output.file path or waveformlog-Mel framesaudio sample, language…Mel window, prompt tok…encoded audio and cand…token probabilities an…decoded tokens and cro…word timestamps, segme…CLI &Python APIAudio frontendWhisper modelLanguage &tokenizerWindow loopDecoding taskTimestampalignment
Connections
BuildingsEntryServiceComputeStorageRuntimeToolingExternal

Terms, in plain English

Log-Mel spectrogramA time-frequency representation of audio used as the encoder input.

How to study this repo

  1. 1
    Prepare audio

    Trace a file through resampling and log-Mel feature construction.

  2. 2
    Inspect decoding

    Follow one token through logits, filters, ranking, and candidate updates.

  3. 3
    Align words

    Connect decoder attention weights to word-level timestamps and segment seeking.