Speech is a domain with access to ridiculously large amounts of data and a very large number of categories.

Judge as of 2014-05-15. True: speech recognition at the time had very large training datasets available (thousands of hours or more, especially at large tech companies) and many output categories (e.g. thousands of context-dependent phone states/senones or large vocabularies) compared with other ML domains. Mostly true: both hold, but "ridiculously large" overstates the publicly available data, while industrial data was large. Mixed: only one of the two (large data, many categories) holds. Mostly false: speech data or categories were modest compared with other domains, and the claim overstates one notable case. False: both are contradicted.

Set before the research began.

1. E1 Large Scale Distributed Acoustic Modeling With Back-off N-grams arXiv (Chelba, Xu, Pereira, Richardson; Google)

   “We use 87 000 hours of training data (speech along with transcription) obtained by”

2. E2 Large Scale Distributed Acoustic Modeling With Back-off N-grams arXiv (Chelba, Xu, Pereira, Richardson; Google)

   “Typical amounts of training data used for the acoustic model (AM) in ASR vary from 100 to 1000 hours.”

3. E3 Deep Neural Networks for Acoustic Modeling in Speech Recognition (IEEE Signal Processing Magazine proof) IEEE Signal Processing Magazine (Hinton, Deng, Yu, Dahl, Mohamed, Kingsbury, Jaitly, Senior, Vanhoucke, Nguyen, Sainath) · November 2012

   “Even when many of the states of these triphone HMMs are tied together, there can be thousands of tied states.”

4. E4 Increasing Deep Neural Network Acoustic Model Size for Large Vocabulary Continuous Speech Recognition arXiv (Maas, Hannun, Lengerich, Qi, Jurafsky, Ng; Stanford)

   “There are 8,986 output classes which results in the output layer being the largest single layer in any of our networks.”

The claim mostly holds. A Google paper on arXiv (1302.1123) reports training on 87,000 hours of transcribed speech (E1). The same paper says typical acoustic-model training sets were 100 to 1,000 hours (E2). So very large data existed in industry, but "ridiculously large" overstates what was typical outside it. On categories, Hinton et al. (2012) note that triphone HMMs have thousands of tied states (E3), and a 2014 Stanford paper uses 8,986 output classes (E4). That fits the "mostly true" criterion: both parts hold, with industrial data large but typical data more modest. None of the sources directly compares speech with other machine-learning domains.