Dr Adnan Haider
Dr Adnan Haider is a Senior ML Research Scientist and Engineer at Apple.
Adnan joined Apple as an ML Research Scientist and Engineer in 2019. Throughout his research career, Adnan primarily focused on the challenge of automatic speech recognition (ASR). ASR technology seamlessly transforms spoken words into text or commands, revolutionising hands-free device operation, enhancing accessibility for individuals with disabilities, and streamlining tasks like dictation and system control. At Apple, Adnan’s research is aimed at developing novel algorithmic frameworks to further improve the robustness of state-of-the-art machine learning ASR models. In particular, he recently developed the ‘Focused Discriminative Training’ framework, an algorithm specifically tailored to identify and improve the recognition accuracy of audio segments that a streaming ASR model finds challenging.
Adnan is an alumnus of Darwin College, and prior to joining Apple, he was a post-doctoral researcher in Division F of the Cambridge University Engineering Department. It is also within the same department that Adnan completed his PhD under the supervision of Professor Phil Woodland, a distinguished fellow at Peterhouse and the Royal Academy of Engineering (FREng). Adnan’s doctoral work was fully funded by the Cambridge Trust. His doctoral research focused on integrating concepts from Differential Geometry to improve the training of Deep Neural Models in large-scale settings.
Adnan’s primary interest and focus lies in incorporating and leveraging techniques from mathematical disciplines like Differential Geometry, Topology, and measure theory to develop pragmatic engineering solutions for the large-scale training of machine learning models. Currently, his research is focused on LM-Aware training, exploring the nuances of transferring knowledge from Large Language Models (LLMs) to improve the training of ASR models.
Outside of work, Adnan is a keen practitioner of Brazilian Jiu-Jitsu and trains under the Brazilian Top Team in Cambridge.
Select publications
Thesis:
[1] A. Haider, “Optimisation Methods for Training Deep Neural Networks in Speech Recognition.”, PhD Thesis, University of Cambridge, 2019.
Journals:
[1] A. Haider, C.Zhang, F.Kreyssig and P.C. Woodland, “A Distributed Optimisation Framework Combining Natural Gradient with Hessian-Free for Discriminative Sequence Training”, in Neural Networks, Elsevier, vol 143, pp-537-549, 2021.
Conferences:
[1] A. Haider, X. Na, Z. Huang, T. Ng and X. Zhuang, “Focussed Discriminative Training”, in Proc. InterSpeech, Greece, 2024.
[2] A. Haider and P.C. Woodland, “Comparison between Hessian Free and Natural Gradient training for DNN Acoustic Models”, in Proc. UK Speech, Cambridge, UK, 2017
[3] A. Haider and P.C. Woodland, “Sequence Training of DNN Acoustic Models With Natural Gradient”, in Proc. Automatic Speech Recognition and Understanding (ASRU), Okinawa, Japan, 2017
[4] A. Haider and P.C. Woodland, “Combining Natural Gradient with Hessian Free Methods for Sequence Training”, in Proc. InterSpeech, Hyderabad, India, 2018.
[5] A Haider, T. Ng, Z. Huang, X. Na and A. Rosti, “A Treatise On FST Lattice Based MMI Training”, in Proc. Sane, Boston, USA, 2022.
Select awards
- Cambridge International Scholarship | Issued by Cambridge Trust