“Deep Speech: Scaling up end-to-end speech recognition,” submitted to arXiv on December 17, 2014 by Awni Hannun, Andrew Ng, and colleagues at Baidu Research, argued that a single deep neural network could replace the elaborate hand-engineered stack that traditional speech systems relied on. Conventional recognizers combined acoustic models, phoneme dictionaries, and hidden Markov models, each separately tuned. Deep Speech learned the mapping from audio to text directly.
Trained on large amounts of data, including synthetically noised audio, the system proved notably robust to background noise and acoustic variation without the specialized noise-handling components used by commercial products of the era. It reported competitive results on the Switchboard Hub5’00 benchmark and handled challenging real-world conditions well.
Why business readers should care: Deep Speech was an early, influential demonstration that end-to-end deep learning could simplify and improve speech recognition. The approach it championed, fewer hand-built parts and more learned-from-data scale, became the dominant pattern for the speech systems people now use every day.