A decade of deep learning cut ASR word error rates by more than 50% relative and, along the way, made classical HMM pipelines look like a relic. This Part 1 breakdown of the end-to-end ASR survey by Prabhavalkar, Hori, Sainath, Schluter, and Watanabe (a lineup that basically is the field) sets up the taxonomy of E2E ASR models and grounds it against the HMM heritage most working engineers grew up on. It's the framing episode that makes the rest of the survey click into place.

The overview focuses on modeling — how CTC, attention-based encoder-decoders, and transducer variants differ in inductive bias, what they buy you in terms of reduced dependence on ASR domain-specific expertise, and why the industry consolidated around all-neural stacks that learn more consistently from data. It sets the stage for later parts on training, decoding, external language model integration, deployment, performance analysis, and future directions for the field. If you're onboarding onto a modern ASR team or trying to decide between Kaldi-style hybrids and something like Whisper, NeMo, ESPnet, or a transducer-based recognizer, this is the map you want before walking the terrain.