The ASR Pipeline

Automatic speech recognition (ASR) is sequence-to-sequence with a twist: the input isn't tokens — it's thousands of audio samples per second.

The pipeline:

• Features — the waveform becomes a mel spectrogram
• Encoder — reads the spectrogram and compresses it into audio embeddings
• Decoder — emits text tokens one at a time, attending back over the audio

Thirty seconds of speech go in; 'the cat sat on the mat' comes out. The model never sees a waveform — it reads a picture of sound and writes words.