Baidu Research

Contribution to AI

Two languages, one architecture, no phoneme dictionary: Deep Speech 2, published at the end of 2015, trained the same network on English and Mandarin and reached accuracy competitive with pipelines that had taken decades of language-specific engineering to assemble. It largely settled the argument over whether hand-built acoustic models and pronunciation lexicons were necessary at all. The supporting releases travelled further than the paper — warp-ctc, a fast implementation of the CTC loss, and DeepBench, which measured the primitive matrix and convolution operations that training actually spends its time on, both ended up in other people's stacks. A 2017 study from the same group measured how generalisation error falls as training sets grow and found power laws holding across translation, language modelling, speech and image classification, an empirical scaling result arriving three years before the ones that reorganised industry spending. PaddlePaddle gave China a deep learning framework maintained domestically rather than in California, which mattered more once tooling access became a policy question. ERNIE, from 2019, changed what masked pretraining hides — whole entities and phrases instead of random tokens, a better fit for a language written without spaces — and its successors topped SuperGLUE in 2021. There was also an expensive lesson: a Baidu team barred from the ImageNet evaluation server in 2015 for repeated test-set submissions, which turned benchmark etiquette into enforced rules.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).