Deep networks in speech recognition
Era 3 · From the expert boom to the statistical turn (1980–2011) › Neural networks
- Lineage
- Neurons and networks
- Kind
- Application
Who
Microsoft Research
What happened
Frank Seide, Gang Li, and Dong Yu of Microsoft Research present a deep network at Interspeech 2011 that transcribes telephone conversations with a third fewer errors than the best earlier systems.
Why it mattered
Deep learning's first major industrial success, a year before AlexNet. Within a few years every major speech recognizer switches to deep networks.
The full story
For two decades, speech recognizers combined hidden Markov models, which track the sequence of sounds, with Gaussian mixture models, which estimate which sound matches each slice of audio. Improvements came year after year in small steps.
In 2009 Hinton's students in Toronto showed that a deep network, pretrained layer by layer like the 2006 deep belief nets, recognized phonemes better than those mixtures. Li Deng and Dong Yu of Microsoft Research took an interest in the result, and together with George Dahl, one of Hinton's students, they carried the idea over to large-vocabulary recognition.
The system built by Seide, Li, and Yu replaced the Gaussian mixtures with a network of up to nine hidden layers trained on more than 300 hours of conversations. On the standard telephone conversation benchmark, the word error rate fell from 27.4 percent to 18.5 percent: a third less in a single stroke, in a field that had spent years advancing a point at a time.
In 2012 the Toronto, Microsoft, Google, and IBM groups published a joint overview that confirmed the result on their own systems. In the years that followed, deep networks went into commercial recognizers, and speech became the first area where deep learning changed a mass-market product.
Builds on
Hidden Markov models in speech (February 1989)
Documented direct influence
The deep network replaces the Gaussian mixtures inside a recognizer built on hidden Markov models
Deep belief nets (July 2006)
Documented direct influence
The recognizer's networks are pretrained layer by layer as deep belief networks
Sources
Original source
- Conversational speech transcription using context-dependent deep neural networks. Interspeech 2011, pp. 437-440 (opens in a new tab). Frank Seide, Gang Li, and Dong Yu, 2011
- Context-Dependent Pre-Trained Deep Neural Networks for Large-Vocabulary Speech Recognition. IEEE Transactions on Audio, Speech, and Language Processing, 20(1), pp. 30-42 (opens in a new tab). George Dahl, Dong Yu, Li Deng, and Alex Acero, 2012
Scholarship and context
- Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups. IEEE Signal Processing Magazine, 29(6), pp. 82-97 (opens in a new tab). Geoffrey Hinton, Li Deng, Dong Yu, George Dahl, Abdel-rahman Mohamed et al., 2012
Further reading
How to cite
“Deep networks in speech recognition.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/deep-networks-speech-recognition
This entry was last reviewed on November 13, 2026.