Speech recognition shifted from time warping and shallow nets to HMMs then deep networks

Evidence

The most common speech recognition systems are based on HMMs (Rabiner, 1989). The early techniques in speech recognition were dynamic time warping (Vintsyuk, 1968) and neural networks (Kohonen, 1988), but they were largely abandoned in the late 1990s. However, recently, deep neural networks have been proposed for use in speech recognition (Dahl et al., 2012). — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 2269

Categorization

• XREF: Connects to the broader arc of neural-network history (the ‘AI winters’ and deep-learning resurgence) also covered in histories like The AI Con.