Speech and Language Processing

The Multi-Genre Broadcast (MGB) Challenge is an evaluation of speech recognition, speaker diarization, dialect detection and lightly supervised alignment using TV recordings in English and Arabic. The speech data is broad and multi-genre, spanning the whole range of TV output, and represents a challenging task for speech technology. In 2015, the challenge used data from the British Broadcasting Corporation (BBC). It was an official challenge of the 2015 IEEE Automatic Speech Recognition and Understanding Workshop. 

The 5th CHiME Speech Separation and Recognition Challenge (CHiME-5). The new challenge will consider the problem of distant multi-microphone conversational speech recognition in everyday home environments. Speech material was elicited using a dinner party scenario with efforts taken to capture data that is representative of natural conversational speech.

This 4th CHiME Speech Separation and Recognition Challenge (CHiME-4) challenge revisits the datasets originally recorded for CHiME-3, i.e., Wall Street Journal corpus sentences spoken by talkers situated in challenging noisy environments recorded using a 6-channel tablet based microphone array. CHiME-4 increases the level of difficulty by constraining the number of microphones available for testing.

One year ago the 2011 PASCAL CHiME Speech Separation and Recognition Challenge considered the problem of recognising speech mixed in two-channel nonstationary noise typical of everyday listening conditions. Following the success of this challenge we are now organising a new challenge that, while keeping the same setting, extends the difficulty along two independent tracks: a larger vocabulary size and a more realistic mixing process that accounts for small head movements made while speaking.

In 2006 the PASCAL network funded the 1st Speech Separation challengewhich addressed the problem of separating and recognising speech mixed with speech. We are now launching a successor to this challenge that aims to tackle speech separation and recognition in more typical everyday listening conditions. The challenge employs noise background that has been collected from a real family living room using binaural microphones.