Ahmed Ali from the QCRI (Qatar Computing Research Institute) will give a talk on Friday 20 at 10:00 in room C005.
He will present his work on Automatic speech recognition developed on 1200 hours of speech.

Abstract: 

Dialectal Arabic speech is suffering from a lack of labelled resources and a lack of orthography. There are three main challenges in dialectal Arabic speech recognition: (1) finding labelled dialectal speech, (2) building robust dialectal speech recognition with limited labelled and (3) evaluating speech recognition for dialects with no standard orthographic rules.
My talk is divided into three parts:

• The first section addresses dialect identification using speech data, and introduces our novel approach in combining acoustic, lexical and phonetic features using deep learning.
• The second section presents our effort in the Multi-Genre Broadcast challenges. In the MGB-2, more than 1,200 hours lightly supervised speech data has been used for training multi-dialect broadcast system. The MGB-3 and MGB-5 will highlight the effort in building dialectal challenge in the wild, where limited dialectal multi-general YouTube data has been used for transfer learning.
• The third part of my talk addresses our effort in evaluating dialectal speech with no standard orthographic rules. Our methods learn from multiple transcribers and align the speech hypothesis to overcome the non-orthographic aspect. We have also automated this process by learning from twitter data different writing and propose a new evaluation metric. Finally, we tried to estimate the word error rate with no reference transcription using decoding and language features. We show that our word error rate estimation is robust for many scenarios with and without the decoding features.

Biography : 

Ahmed Ali is a Principal Engineer at the Arabic Language Technologies group (ALT) at the Qatar Computing Research Institute (QCRI). His work interests are in the area of speech recognition and natural language processing. Ahmed completed his PhD at the University of Edinburgh and BSc at Cairo University. He joined QCRI in 2011.
Prior to joining QCRI, Ahmed worked for Nuance in Cambridge, UK from 2006-2011. Before Nuance, Ahmed worked for six years at different IBM labs, including two years at Watson, USA, labs. His current research on speech recognition has led to the development of state-of-the-art multi-dialect Arabic speech recognition: QCRI advanced transcription system (QATS). QATS is used by Al-Jazeera to transcribe their online content. DW and BBC have been using QATS. Ahmed’s team has won first place in the worldwide MGB-2 speech recognition competition. He also won the prestigious Qatar Foundation Best Innovation Award in 2018.