| Zusammenfassung: | Recently, end-to-end architectures have dominated the modeling of Automatic Speech Recognition (ASR) systems. Conventional systems usually consist of independent components, like an acoustic model, a language model and a pronunciation model. In comparison, end-to-end ASR approaches aim to directly map acoustic inputs to character or word sequences, which significantly simplifies the complex traini... Recently, end-to-end architectures have dominated the modeling of Automatic Speech Recognition (ASR) systems. Conventional systems usually consist of independent components, like an acoustic model, a language model and a pronunciation model. In comparison, end-to-end ASR approaches aim to directly map acoustic inputs to character or word sequences, which significantly simplifies the complex training procedure. Plenty of end-to-end architectures have been proposed, for instance, Connectionist Temporal Classification (CTC), Sequence Transduction with Recurrent Neural Networks (RNN-T) and attention-based encoder-decoder, which have accomplished great success and achieved impressive performance on a variety of benchmarks or even reached human level on some tasks.
However, although advanced deep neural network architectures have been proposed, in adverse environments, the performance of ASR systems suffers from significant degradation because of environmental noise or ambient reverberation. To improve the robustness of ASR systems, in this thesis, we address the research questions and conduct experiments from the following perspectives:
Firstly, to learn more stable visual representations, we propose LipSound and LipSound2 and investigate to what extent the visual modality contains semantic information that can benefit ASR performance. The LipSound/LipSound2 model consists of an encoder-decoder with an location-aware attention architecture and directly transforms mouth or face movement sequences to low-level speech representations, i.e. mel-scale spectrograms. The model is trained in a crossmodal self-supervised fashion and does not require any human annotations since the model inputs (visual sequences) and outputs (audio signals) are naturally paired in videos. Experimental results show that the LipSound model not only generates quality mel-spectrograms but also outperforms state-of-the-art models on the GRID benchmark dataset in speaker-dependent settings. Moreover, the improved LipSound2 model further verifies the effectiveness on generalizability (speaker-independent) and transferability (Non-Chinese to Chinese) on large vocabulary continuous speech corpora.
Secondly, to exploit the fact that the image of a face contains information about the person's speech sound, we incorporate face embeddings extracted from a pretrained model for face recognition into the target speech separation model, which guide the system for predicting a target speaker mask in the time-frequency domain. The experimental results show that a pre-enrolled face image is able to benefit separating expected speech signals. Additionally, face information is complementary to voice reference. Further improvement can be achieved when combining both face and voice embeddings.
Thirdly, to integrate domain knowledge, i.e. articulatory features (AFs) into end-to-end learning, we present two approaches: (a) fine-tuning networks which reuse hidden layer representations of AF extractors as input for ASR tasks; (b) progressive networks which combine articulatory knowledge by lateral connections from AF extractors. Results show that progressive networks are more effective and accomplish a lower word error rate than fine-tuning networks and other baseline models.
Finally, to enable end-to-end ASR models to acquire Out-of-Vocabulary (OOV) words, instead of just fine-tuning with the audio containing OOV words, we propose to rescale loss at sentence level or word level, which encourages models to pay more attention to unknown words. Experimental results reveal that fine-tuning the baseline ASR model with loss rescaling and L2/EWC (Elastic Weight Consolidation) regularization can significantly improve the recall rate of OOV words and efficiently overcome the model suffering catastrophic forgetting. Furthermore, loss rescaling at the word level is more stable than the sentence level method and results in less ASR performance loss on general non-OOV words and the LibriSpeech dataset.
In sum, this thesis contributes to the robustness of ASR systems by leveraging additional visual sequences, face information and domain knowledge. We achieve significant improvement on speech reconstruction, speech separation, end-to-end modeling and OOV word recognition tasks. |
