In this paper, we report our submitted system for the ZeroSpeech 2020 challenge on Track 2019. The main theme in this challenge is to build a speech synthesizer without any textual information or phonetic labels. In order to tackle those challenges, we build a system that must address two major components such as 1) given speech audio, extract subword units in an unsupervised way and 2) re-synthesize the audio from novel speakers. The system also needs to balance the codebook performance between the ABX error rate and the bitrate compression rate. Our main contribution here is we proposed Transformer-based VQ-VAE for unsupervised unit discovery and Transformer-based inverter for the speech synthesis given the extracted codebook. Additionally, we also explored several regularization methods to improve performance even further.
Bridging robot action sequences and their natural language captions is an important task to increase explainability of human assisting robots in their recently evolving field. In this paper, we propose a system for generating natural language captions that describe behaviors of human assisting robots. The system describes robot actions by using robot observations; histories from actuator systems and cameras, toward end-to-end bridging between robot actions and natural language captions. Two reasons make it challenging to apply existing sequence-to-sequence models to this mapping: 1) it is hard to prepare a large-scale dataset for any kind of robots and their environment, and 2) there is a gap between the number of samples obtained from robot action observations and generated word sequences of captions. We introduced unsupervised segmentation based on K-means clustering to unify typical robot observation patterns into a class. This method makes it possible for the network to learn the relationship from a small amount of data. Moreover, we utilized a chunking method based on byte-pair encoding (BPE) to fill in the gap between the number of samples of robot action observations and words in a caption. We also applied an attention mechanism to the segmentation task. Experimental results show that the proposed model based on unsupervised learning can generate better descriptions than other methods. We also show that the attention mechanism did not work well in our low-resource setting.
This work proposes a new human-related video processing task named 3D panoramic multi-person localization and tracking. With a benchmark dataset and a simple yet effective solution, it establishes a new paradigm for multi-person tracking systems and related applications. Unlike existing methods that can only work on a 2D coordinate or a narrow-angle-view 3D coordinate, our proposal can maximally explore the 3D trajectory information of tracking targets. This is approached by applying camera geometry to transform human locations from 2D panoramic image coordinates to a 3D panoramic camera coordinate, and then by applying a tracking algorithm that associates human appearance and 3D trajectory together.
Simultaneous machine translation is a variant of machine translation that starts the translation process before the end of an input. This task faces a trade-off between translation accuracy and latency. We have to determine when we start the translation for observed inputs so far, to achieve good practical performance. In this work, we propose a neural machine translation method to determine this timing in an adaptive manner. The proposed method introduces a special token '<wait>', which is generated when the translation model chooses to read the next input token instead of generating an output token. It also introduces an objective function to handle the ambiguity in wait timings that can be optimized using an algorithm called Connectionist Temporal Classification (CTC). The use of CTC enables the optimization to consider all possible output sequences including '<wait>' that are equivalent to the reference translations and to choose the best one adaptively. We apply the proposed method into simultaneous translation from English to Japanese and investigate its performance and remaining problems.
Deep acoustic models typically receive features in the first layer of the network, and process increasingly abstract representations in the subsequent layers. Here, we propose to feed the input features at multiple depths in the acoustic model. As our motivation is to allow acoustic models to re-examine their input features in light of partial hypotheses we introduce intermediate model heads and loss function. We study this architecture in the context of deep Transformer networks, and we use an attention mechanism over both the previous layer activations and the input features. To train this model's intermediate output hypothesis, we apply the objective function at each layer right before feature re-use. We find that the use of such intermediate losses significantly improves performance by itself, as well as enabling input feature re-use. We present results on both Librispeech, and a large scale video dataset, with relative improvements of 10 - 20% for Librispeech and 3.2 - 13% for videos.
In this paper, we explore a method for training speech-to-speech translation tasks without any transcription or linguistic supervision. Our proposed method consists of two steps: First, we train and generate discrete representation with unsupervised term discovery with a discrete quantized autoencoder. Second, we train a sequence-to-sequence model that directly maps the source language speech to the target language's discrete representation. Our proposed method can directly generate target speech without any auxiliary or pre-training steps with a source or target transcription. To the best of our knowledge, this is the first work that performed pure speech-to-speech translation between untranscribed unknown languages.
Although skeleton-based action recognition has achieved great success in recent years, most of the existing methods may suffer from a large model size and slow execution speed. To alleviate this issue, we analyze skeleton sequence properties to propose a Double-feature Double-motion Network (DD-Net) for skeleton-based action recognition. By using a lightweight network structure (i.e., 0.15 million parameters), DD-Net can reach a super fast speed, as 3,500 FPS on one GPU, or, 2,000 FPS on one CPU. By employing robust features, DD-Net achieves the state-of-the-art performance on our experimental datasets: SHREC (i.e., hand actions) and JHMDB (i.e., body actions). Our code will be released with this paper later.