Automatic Speech Recognition

Introduction

1. What was Radio Rex and how did it “recognize” speech?
   A 1920s toy celluloid dog, the first machine that recognized speech. A spring released it when it detected 500 Hz acoustic energy — roughly F1 of the vowel [eh] in “Rex” — so it seemed to come when called.
   • Jurafsky & Martin, Speech and Language Processing, 3rd ed. draft, ch. 16, Introduction
2. What is the task of automatic speech recognition (ASR)?
   To map any acoustic waveform to the appropriate string of words (or graphemes).
   • Jurafsky & Martin, Speech and Language Processing, 3rd ed. draft, ch. 16, Introduction
3. List the main practical applications of ASR mentioned in the chapter.
   Natural interfaces for appliances, digital assistants and chatbots (esp. on phones); general transcription such as captions for audio/video; dictation (e.g. in law); and augmentative communication for people with difficulties typing or hearing.
   • Jurafsky & Martin, Speech and Language Processing, 3rd ed. draft, ch. 16, Introduction
4. What are the two families of ASR methods introduced in Chapter 16?
   (1) The attention-based encoder-decoder paradigm (e.g. Listen Attend and Spell, Whisper, OWSM). (2) Self-supervised speech models (SSL) like wav2vec 2.0 or HuBERT, paired with the CTC loss for decoding.
   • Jurafsky & Martin, Speech and Language Processing, 3rd ed. draft, ch. 16, Introduction
5. What is OWSM?
   Open Whisper-style Speech Model (Peng et al., 2023): an open system using Whisper's architecture, with a fully open-source toolkit and publicly available training data.
   • Jurafsky & Martin, Speech and Language Processing, 3rd ed. draft, ch. 16, Introduction