Dynamic Acoustic Unit Augmentation with BPE-Dropout for Low-Resource End-to-End Speech Recognition

Adapting Off-the-Shelf Speech Recognition Systems for Novel Words

Current speech recognition systems with fixed vocabularies have difficulties recognizing Out-of-Vocabulary words (OOVs) such as proper nouns and new words. This leads to misunderstandings or even failures in dialog systems. Ensuring effective speech recognition is crucial for the proper functioning of robot assistants. Non-native accents, new vocabulary, and aging voices can cause malfunctions in a speech recognition system. If this task is not executed correctly, the assistant robot will inevitably produce false or random responses. In this paper, we used a statistical approach based on distance algorithms to improve OOV correction. We developed a post-processing algorithm to be combined with a speech recognition model. In this sense, we compared two distance algorithms: Damerau–Levenshtein and Levenshtein distance. We validated the performance of the two distance algorithms in conjunction with five off-the-shelf speech recognition models. Damerau–Levenshtein, as compared to the Levenshtein distance algorithm, succeeded in minimizing the Word Error Rate (WER) when using the MoroccanFrench test set with five speech recognition systems, namely VOSK API, Google API, Wav2vec2.0, SpeechBrain, and Quartznet pre-trained models. Our post-processing method works regardless of the architecture of the speech recognizer, and its results on our MoroccanFrench test set outperformed the five chosen off-the-shelf speech recognizer systems.

Deep Learning Framework for Controlling Work Sequence in Collaborative Human-Robot Assembly Processes

The human-robot collaboration (HRC) solutions presented so far have the disadvantage that the interaction between humans and robots is based on the human's state or on specific gestures purposely performed by the human, thus increasing the time required to perform a task and slowing down the pace of human labor, making such solutions uninteresting. In this study, a different concept of the HRC system is introduced, consisting of an HRC framework for managing assembly processes that are executed simultaneously or individually by humans and robots. This HRC framework based on deep learning models uses only one type of data, RGB camera data, to make predictions about the collaborative workspace and human action, and consequently manage the assembly process. To validate the HRC framework, an industrial HRC demonstrator was built to assemble a mechanical component. Four different HRC frameworks were created based on the convolutional neural network (CNN) model structures: Faster R-CNN ResNet-50 and ResNet-101, YOLOv2 and YOLOv3. The HRC framework with YOLOv3 structure showed the best performance, showing a mean average performance of 72.26% and allowed the HRC industrial demonstrator to successfully complete all assembly tasks within a desired time window. The HRC framework has proven effective for industrial assembly applications.

Audio Augmentation for Non-Native Children's Speech Recognition through Discriminative Learning

Automatic speech recognition (ASR) in children is a rapidly evolving field, as children become more accustomed to interacting with virtual assistants, such as Amazon Echo, Cortana, and other smart speakers, and it has advanced the human-computer interaction in recent generations. Furthermore, non-native children are observed to exhibit a diverse range of reading errors during second language (L2) acquisition, such as lexical disfluency, hesitations, intra-word switching, and word repetitions, which are not yet addressed, resulting in ASR's struggle to recognize non-native children's speech. The main objective of this study is to develop a non-native children's speech recognition system on top of feature-space discriminative models, such as feature-space maximum mutual information (fMMI) and boosted feature-space maximum mutual information (fbMMI). Harnessing the collaborative power of speed perturbation-based data augmentation on the original children's speech corpora yields an effective performance. The corpus focuses on different speaking styles of children, together with read speech and spontaneous speech, in order to investigate the impact of non-native children's L2 speaking proficiency on speech recognition systems. The experiments revealed that feature-space MMI models with steadily increasing speed perturbation factors outperform traditional ASR baseline models.