Fusion of Self-supervised Learned Models for MOS Prediction

We participated in the mean opinion score (MOS) prediction challenge, 2022. This challenge aims to predict MOS scores of synthetic speech on two tracks, the main track and a more challenging sub-track: out-of-domain (OOD). To improve the accuracy of the predicted scores, we have explored several model fusion-related strategies and proposed a fused framework in which seven pretrained self-supervised learned (SSL) models have been engaged. These pretrained SSL models are derived from three ASR frameworks, including Wav2Vec, Hubert, and WavLM. For the OOD track, we followed the 7 SSL models selected on the main track and adopted a semi-supervised learning method to exploit the unlabeled data. According to the official analysis results, our system has achieved 1st rank in 6 out of 16 metrics and is one of the top 3 systems for 13 out of 16 metrics. Specifically, we have achieved the highest LCC, SRCC, and KTAU scores at the system level on main track, as well as the best performance on the LCC, SRCC, and KTAU evaluation metrics at the utterance level on OOD track. Compared with the basic SSL models, the prediction accuracy of the fused system has been largely improved, especially on OOD sub-track.

MC-UNet Multi-module Concatenation based on U-shape Network for Retinal Blood Vessels Segmentation

Accurate segmentation of the blood vessels of the retina is an important step in clinical diagnosis of ophthalmic diseases. Many deep learning frameworks have come up for retinal blood vessels segmentation tasks. However, the complex vascular structure and uncertain pathological features make the blood vessel segmentation still very challenging. A novel U-shaped network named Multi-module Concatenation which is based on Atrous convolution and multi-kernel pooling is put forward to retinal vessels segmentation in this paper. The proposed network structure retains three layers the essential structure of U-Net, in which the atrous convolution combining the multi-kernel pooling blocks are designed to obtain more contextual information. The spatial attention module is concatenated with dense atrous convolution module and multi-kernel pooling module to form a multi-module concatenation. And different dilation rates are selected by cascading to acquire a larger receptive field in atrous convolution. Adequate comparative experiments are conducted on these public retinal datasets: DRIVE, STARE and CHASE_DB1. The results show that the proposed method is effective, especially for microvessels. The code will be put out at https://github.com/Rebeccala/MC-UNet

MELONS: generating melody with long-term structure using transformers and structure graph

The creation of long melody sequences requires effective expression of coherent musical structure. However, there is no clear representation of musical structure. Recent works on music generation have suggested various approaches to deal with the structural information of music, but generating a full-song melody with clear long-term structure remains a challenge. In this paper, we propose MELONS, a melody generation framework based on a graph representation of music structure which consists of eight types of bar-level relations. MELONS adopts a multi-step generation method with transformer-based networks by factoring melody generation into two sub-problems: structure generation and structure conditional melody generation. Experimental results show that MELONS can produce structured melodies with high quality and rich contents.

AdaptCL: Efficient Collaborative Learning with Dynamic and Adaptive Pruning

In multi-party collaborative learning, the parameter server sends a global model to each data holder for local training and then aggregates committed models globally to achieve privacy protection. However, both the dragger issue of synchronous collaborative learning and the staleness issue of asynchronous collaborative learning make collaborative learning inefficient in real-world heterogeneous environments. We propose a novel and efficient collaborative learning framework named AdaptCL, which generates an adaptive sub-model dynamically from the global base model for each data holder, without any prior information about worker capability. All workers (data holders) achieve approximately identical update time as the fastest worker by equipping them with capability-adapted pruned models. Thus the training process can be dramatically accelerated. Besides, we tailor the efficient pruned rate learning algorithm and pruning approach for AdaptCL. Meanwhile, AdaptCL provides a mechanism for handling the trade-off between accuracy and time overhead and can be combined with other techniques to accelerate training further. Empirical results show that AdaptCL introduces little computing and communication overhead. AdaptCL achieves time savings of more than 41\% on average and improves accuracy in a low heterogeneous environment. In a highly heterogeneous environment, AdaptCL achieves a training speedup of 6.2x with a slight loss of accuracy.

Pretraining Strategies, Waveform Model Choice, and Acoustic Configurations for Multi-Speaker End-to-End Speech Synthesis

We explore pretraining strategies including choice of base corpus with the aim of choosing the best strategy for zero-shot multi-speaker end-to-end synthesis. We also examine choice of neural vocoder for waveform synthesis, as well as acoustic configurations used for mel spectrograms and final audio output. We find that fine-tuning a multi-speaker model from found audiobook data that has passed a simple quality threshold can improve naturalness and similarity to unseen target speakers of synthetic speech. Additionally, we find that listeners can discern between a 16kHz and 24kHz sampling rate, and that WaveRNN produces output waveforms of a comparable quality to WaveNet, with a faster inference time.

Robust Skeletonization for Plant Root Structure Reconstruction from MRI

Structural reconstruction of plant roots from MRI is challenging, because of low resolution and low signal-to-noise ratio of the 3D measurements which may lead to disconnectivities and wrongly connected roots. We propose a two-stage approach for this task. The first stage is based on semantic root vs. soil segmentation and finds lowest-cost paths from any root voxel to the shoot. The second stage takes the largest fully connected component generated in the first stage and uses 3D skeletonization to extract a graph structure. We evaluate our method on 22 MRI scans and compare to human expert reconstructions.

Learning Disentangled Phone and Speaker Representations in a Semi-Supervised VQ-VAE Paradigm

We present a new approach to disentangle speaker voice and phone content by introducing new components to the VQ-VAE architecture for speech synthesis. The original VQ-VAE does not generalize well to unseen speakers or content. To alleviate this problem, we have incorporated a speaker encoder and speaker VQ codebook that learns global speaker characteristics entirely separate from the existing sub-phone codebooks. We also compare two training methods: self-supervised with global conditions and semi-supervised with speaker labels. Adding a speaker VQ component improves objective measures of speech synthesis quality (estimated MOS, speaker similarity, ASR-based intelligibility) and provides learned representations that are meaningful. Our speaker VQ codebook indices can be used in a simple speaker diarization task and perform slightly better than an x-vector baseline. Additionally, phones can be recognized from sub-phone VQ codebook indices in our semi-supervised VQ-VAE better than self-supervised with global conditions.

Cross-regional oil palm tree counting and detection via multi-level attention domain adaptation network

Providing an accurate evaluation of palm tree plantation in a large region can bring meaningful impacts in both economic and ecological aspects. However, the enormous spatial scale and the variety of geological features across regions has made it a grand challenge with limited solutions based on manual human monitoring efforts. Although deep learning based algorithms have demonstrated potential in forming an automated approach in recent years, the labelling efforts needed for covering different features in different regions largely constrain its effectiveness in large-scale problems. In this paper, we propose a novel domain adaptive oil palm tree detection method, i.e., a Multi-level Attention Domain Adaptation Network (MADAN) to reap cross-regional oil palm tree counting and detection. MADAN consists of 4 procedures: First, we adopted a batch-instance normalization network (BIN) based feature extractor for improving the generalization ability of the model, integrating batch normalization and instance normalization. Second, we embedded a multi-level attention mechanism (MLA) into our architecture for enhancing the transferability, including a feature level attention and an entropy level attention. Then we designed a minimum entropy regularization (MER) to increase the confidence of the classifier predictions through assigning the entropy level attention value to the entropy penalty. Finally, we employed a sliding window-based prediction and an IOU based post-processing approach to attain the final detection results. We conducted comprehensive ablation experiments using three different satellite images of large-scale oil palm plantation area with six transfer tasks. MADAN improves the detection accuracy by 14.98% in terms of average F1-score compared with the Baseline method (without DA), and performs 3.55%-14.49% better than existing domain adaptation methods.

PBRnet: Pyramidal Bounding Box Refinement to Improve Object Localization Accuracy

Many recently developed object detectors focused on coarse-to-fine framework which contains several stages that classify and regress proposals from coarse-grain to fine-grain, and obtains more accurate detection gradually. Multi-resolution models such as Feature Pyramid Network(FPN) integrate information of different levels of resolution and effectively improve the performance. Previous researches also have revealed that localization can be further improved by: 1) using fine-grained information which is more translational variant; 2) refining local areas which is more focused on local boundary information. Based on these principles, we designed a novel boundary refinement architecture to improve localization accuracy by combining coarse-to-fine framework with feature pyramid structure, named as Pyramidal Bounding Box Refinement network(PBRnet), which parameterizes gradually focused boundary areas of objects and leverages lower-level feature maps to extract finer local information when refining the predicted bounding boxes. Extensive experiments are performed on the MS-COCO dataset. The PBRnet brings a significant performance gains by roughly 3 point of $mAP$ when added to FPN or Libra R-CNN. Moreover, by treating Cascade R-CNN as a coarse-to-fine detector and replacing its localization branch by the regressor of PBRnet, it leads an extra performance improvement by 1.5 $mAP$, yielding a total performance boosting by as high as 5 point of $mAP$.