learn.ifs.name

Automatic Speech Recognition

Jurafsky & Martin, Speech and Language Processing, 3rd ed. draft, ch. 16

15 cards in “Historical notes” · whole deck

Historical notes

  1. Early Bell Labs digit recognizer (Davis et al., 1952)?
    Recognized the 10 digits from a single speaker using 10 speaker-dependent stored patterns roughly representing the first two vowel formants; 97–99% accuracy by choosing the highest correlation.
  2. What was notable about Fry and Denes's 1959 recognizer?
    A phoneme recognizer at University College London (4 vowels, 9 consonants) — the first to use phoneme transition probabilities to constrain recognition.
  3. What three paradigm shifts came in the late 1960s–early 1970s?
    (1) Feature extraction: FFT, cepstral processing, LPC. (2) Handling warping with dynamic programming. (3) The rise of the HMM.
  4. What is warping in early ASR?
    Stretching or shrinking the input signal to handle differences in speaking rate and segment length when matching stored patterns.
  5. What is dynamic time warping, and who developed it for speech?
    Non-probabilistic dynamic programming to match a template against incoming speech. First by Vintsyuk (1968), reinvented by Velichko & Zagoruyko (1970) and Sakoe & Chiba (1971); Itakura (1975) combined it with LPC features.
  6. Who applied HMMs to speech around 1972?
    Independently: James Baker at CMU (from Baum's work at the Institute for Defense Analyses), and Frederick Jelinek's group at IBM (from information theory).
  7. Baker's DRAGON vs IBM: decoding algorithms?
    DRAGON used Viterbi (dynamic programming) decoding; IBM used Jelinek's stack decoding. Baker later founded Dragon Systems.
  8. What was the dominant ASR paradigm by the 1990s?
    HMMs with Gaussian Mixture Models (GMMs) as the phonetic component.
  9. What was ARPA's 1971 program and its winner?
    A five-year program for 1000-word, constrained-grammar, few-speaker speech understanding; CMU's Harpy (a simplified version of Baker's DRAGON) was the best.
  10. What is a bakeoff, and why did it matter?
    An approximately annual competitive evaluation in ARPA tasks. It spurred wide borrowing of techniques between labs and helped spread the HMM paradigm.
  11. Name early neural ASR architectures (~1990).
    Time-delay neural network (TDNN) — the first convolutional networks for speech; RNNs; and the hybrid HMM/MLP architecture.
  12. What was the hybrid HMM/MLP architecture and its limitation?
    A feedforward network trained as a phonetic classifier whose outputs serve as probability estimates for an HMM (Morgan & Bourlard). Too slow to train on CPUs of the era — the largest had a 4000-unit hidden layer and needed special hardware.
  13. When did deep hybrid systems surpass HMM/GMM, and what mattered?
    By 2012. By 2013 it was clear what mattered was lots of data and enough layers, plus log mel features instead of MFCCs, dropout, and ReLUs.
  14. Timeline of CTC, RNN-T, and attention-based ASR?
    CTC proposed 2006 (Graves et al.); RNN-T defined by 2012; attention-based encoder-decoders ~2014–2016 (Chorowski et al.; Chan et al.'s LAS); Transformers in encoder-decoders by 2018.
  15. Name popular speech processing toolkits.
    Kaldi and ESPnet.

Source