Neural Foundations · 2015

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

Dario Amodei, Rishita Anubhai, Eric Battenberg, Carl Case, Jared Casper, Bryan Catanzaro, Adam Coates, Greg Diamos, Erich Elsen, Awni Hannun, Andrew Ng, Sanjeev Satheesh, Shubho Sengupta, Zhenyao Zhu

Primary source

Showed that one end-to-end deep recurrent network trained with CTC on tens of thousands of hours of audio, backed by HPC-style systems work on multi-GPU training and batched deployment, could approach human transcription accuracy in both English and Mandarin, replacing hand-engineered speech pipelines.

Editorial record

Plain-language summary

Speech recognisers used to be assembled from separate acoustic, pronunciation and language models built by specialists. Deep Speech 2 trains a single network from spectrograms to characters and shows that the same recipe works for two very different languages. Much of the paper is about engineering at scale: synchronous training across many GPUs, custom kernels, and a batching scheme that made the model cheap to serve. It was an early, influential demonstration that data, compute and end-to-end learning could displace expert-built systems.

Source record

Provenance

Record ID: P-681
Record created: 2026-09-26
Last reviewed: 2026-09-26
Record version: 1

https://arxiv.org/abs/1512.02595
arXiv:1512.02595

Citation caveat: Citation metadata is approximate and marked unverified in the source dataset.