Speech recognition is audio classification at every timestep, glued together by a sequence model that knows English and silence. CTC, RNN-T, and attention are the three ways to do it. Pick one and understand why.
You have a 10-second 16 kHz clip. You want a string: "turn on the kitchen lights". The challenge is structural: audio frames do not align one-to-one with characters. The word "okay" might take 200 ms or 1200 ms. Silence punctuates the utterance. Some phonemes are longer than others. The number of output tokens is not known in advance.
Three formulations solve this:
CTC (Connectionist Temporal Classification). Emit per-frame token probabilities including a special blank. Collapse repeats and blanks at decode time. Non-autoregressive, fast. Used by wav2vec 2.0, MMS.
RNN-T (Recurrent Neural Network Transducer). Joint network predicts next token given encoder frame and previous tokens. Streamable. Used by Google's on-device ASR, NVIDIA Parakeet.
Attention encoder-decoder. Encoder compresses audio to hidden states, decoder cross-attends to generate tokens autoregressively. Used by Whisper, SeamlessM4T.
In 2026, SOTA WER on LibriSpeech test-clean is 1.4% (Parakeet-TDT-1.1B, NVIDIA) and 1.58% (Whisper-Large-v3-turbo). The differences are tiny; the deployment differences are huge.
Full Access lesson
Keep going. The rest of this lesson is one step away.
Unlock all 503 lessons, video, code, quizzes and every future update. Pay in your own currency with Paystack.