Simulating realistic multiparty speech data: for the development of distant microphone ASR systems

Automatic speech recognition has become a ubiquitous technology integrated into our daily lives. However, the problem remains challenging when the speaker is far away from the microphone. In such scenarios, the speech is degraded both by reverberation and by the presence of additive noise. This situation is particularly challenging when there are competing speakers present (i.e. multi-party scenarios) Acoustic scene simulation has been a major tool for training and developing distant microphone speech recognition systems, and is now being used to develop solutions for mult-party scenarios. It has been used both in training -- as it allows cheap generation of limitless amounts of data -- and for evaluation -- because it can provide easy access to a ground truth (i.e. a noise-free target signal). However, whilst much work has been conducted to produce realistic artificial scene simulators, the signals produced from such simulators are only as good as the `metadata' being used to define the setups, i.e., the data describing, for example, the number of speakers and their distribution relative to the microphones. This thesis looks at how realistic metadata can be derived by analysing how speakers behave in real domestic environments. In particular, how to produce scenes that provide a realistic distribution for various factors that are known to influence the 'difficulty' of the scene, including the separation angle between speakers, the absolute and relative distances of speakers to microphones, and the pattern of temporal overlap of speech. Using an existing audio-visual multi-party conversational dataset, CHiME-5, each of these aspects has been studied in turn. First, producing a realistic angular separation between speakers allows for algorithms which enhance signals based on the direction of arrival to be fairly evaluated, reducing the mismatch between real and simulated data. This was estimated using automatic people detection techniques in video recordings from CHiME-5. Results show that commonly used datasets of simulated signals do not follow a realistic distribution, and when a realistic distribution is enforced, a significant drop in performance is observed. Second, by using multiple cameras it has been possible to estimate the 2-D positions of people inside each scene. This has allowed the estimation of realistic distributions for the absolute distance to the microphone and relative distance to the competing speaker. The results show grouping behaviour among participants when located in a room and the impact this has on performance depends on the room size considered. Finally, the amount of overlap and points in the mixture which contain overlap were explored using finite-state models. These models allowed for mixtures to be generated, which approached the overlap patterns observed in the real data. Features derived from these models were also shown to be a predictor of the difficulty of the mixture. At each stage of the project, simulated datasets derived using the realistic metadata distributions have been compared to existing standard datasets that use naive or uninformed metadata distributions, and implications for speech recognition performance are observed and discussed. This work has demonstrated how unrealistic approaches can produce over-promising results, and can bias research towards techniques that might not work well in practice. Results will also be valuable in informing the design of future simulated datasets.