SpeechWave

点击翻译按钮获取中文摘要

Speech recognition has made major advances in the past few years. Error rates have been reduced by more than half on standard large-scale tasks such as Switchboard (conversational telephone speech), MGB (multi-genre broadcast recordings), and AMI (multiparty meetings). These research advances have quickly translated into commercial products and services: speech-based applications and assistants such as such as Apple's Siri, Amazon's Alexa, and Google voice search have become part of daily life for many people. Underpinning the improved accuracy of these systems are advances in acoustic modelling, with deep learning having had an outstanding influence on the field.However, speech recognition is still very fragile: it has been successfully deployed in specific acoustic conditions and task domains - for instance, voice search on a smart phone - and degrades severely when the conditions change. This is because speech recognition is highly vulnerable to additive noise caused by multiple acoustic sources, and to reverberation. In both cases, acoustic conditions which have essentially no effect on the accuracy of human speech recognition can have a catastrophic impact on the accuracy of a state-of-the-art automatic system. A reason for such brittleness is the lack of a strong model for acoustic robustness. Robustness is usually addressed through multi-condition training, in which the training set comprises speech examples across the many required acoustic conditions, often constructed by mixing speech with noise at different signal-to-noise ratios. For a limited set of acoustic conditions these techniques can work well, but they are inefficient and do not offer a model of multiple acoustic sources, nor do they factorise the causes of variability. For instance, the best reported speech recognition results for transcription of the AMI corpus test set using single distant microphone recordings is about 38% word error rate (for non-overlapped speech), compared to about 5% error rate for human listeners. In the past few years there have been several approaches that have tried to address these problems: explicitly learning to separate multiple sources; factorised acoustic models using auxiliary features; and learned spectral masks for multi-channel beam-forming. SpeechWave will pursue an alternative approach to robust speech recognition: The development of acoustic models which learn directly from the speech waveform. The motivation to operate directly in the waveform domain arises from the insight that redundancy in speech signals is highly likely to be a key factor in the robustness of human speech recognition. Current approaches to speech recognition separate non-adaptive signal processing components from the adaptive acoustic model, and in so doing lose the redundancy - and, typically, information such as the phase - present in the speech waveform. Waveform models are particularly exciting as they combine the previously distinct signal processing and acoustic modelling components.In SpeechWave, we shall explore novel waveform-based convolutional and recurrent networks which combine speech enhancement and recognition in a factorised way, and approaches based on kernel methods and on recent research advances in sparse signal processing and speech perception. Our research will be evaluated on standard large-scale speech corpora. In addition we shall participate in, and organise, international challenges to assess the performance of speech recognition technologies. We shall also validate our technologies in practice, in the context of the speech recognition challenges faced by our project partners BBC, Emotech, Quorate, and SRI.