## TL;DR
Speech recognition converts audio signals to text. Modern end-to-end approaches — CTC, RNN-Transducer, Attention-based Encoder-Decoder — have replaced traditional HMM-GMM pipelines, with Whisper achieving near-human performance.

## Core Explanation
Traditional pipeline: feature extraction (MFCCs) → acoustic model (phoneme prediction) → pronunciation model (phoneme-to-word) → language model (word sequence scoring). End-to-end models collapse these into a single neural network trained on audio-text pairs.

## Detailed Analysis
Whisper's architecture: encoder-decoder transformer trained on diverse data (different recording conditions, accents, languages). Its multitask training format (transcribe/translate/language identification) improves robustness. Wav2Vec 2.0 (Facebook) uses self-supervised pretraining on raw audio followed by fine-tuning with limited labeled data.

## Further Reading
- OpenAI Whisper GitHub
- Kaldi ASR Toolkit
- Hugging Face: Automatic Speech Recognition