With the development of deep learning, nonparallel voice conversion (VC) has achieved a significant progress recently. Automatic speech recognition (ASR) and text-to-speech (TTS) for leveraging knowledge are the two mainstream methods in VC research. In this paper, we demonstrate that the two bottleneck features (BNFs) in the above methods are complementary. ASR-BNFs are more robust especially in any-to-many tasks, but suffer from leakage of source speaker’s timbre information; TTS-BNFs are less likely to reveal speaker’s timbre information, but lack robustness. Therefore, a nonparallel any-to-many voice conversion model is proposed by combining ASR-BNFs and TTS-BNFs. The whole modules in the proposed model can be trained jointly without any pre-trained models. Experiments are conducted on a private multi-speaker TTS dataset. It is demonstrated that the proposed model achieves the best balance in speech quality, timbre similarity and robustness compared to baseline models.

This paper describes LeVoice automatic speech recognition systems to track2 of intelligent cockpit speech recognition challenge 2022. Track2 is a speech recognition task without limits on the scope of model size. Our main points include deep learning based speech enhancement, text-to-speech based speech generation, training data augmentation via various techniques and speech recognition model fusion. We compared and fused the hybrid architecture and two kinds of end-toend architecture. For end-to-end modeling, we used models based on connectionist temporal classification/attention-based encoder-decoder architecture and recurrent neural network transducer/attention-based encoder-decoder architecture. The performance of these models is evaluated with an additional language model to reduce word error rates. As a result, our system achieved 10.2% character error rate on the challenge test set data and ranked third place among the submitted systems in the challenge.

Due to the difficulty of obtaining parallel data, there are many works focus on non-parallel voice conversion(VC) recently. Bottleneck features(BNFs) from automatic speech recognition(ASR) and text-to-speech(TTS) models play an important role in feature disentangling for VC. In this work, we propose Mix-Guided VC, a non-parallel any-to-many voice conversion model by combining ASR-BNFs and TTS-BNFs. We demonstrate that ASR-BNFs and TTS-BNFs are complementary. ASR-BNFs are more robust especially in any-to-many tasks, but suffer from leaking source speaker’s timbre information; TTS-BNFs are closely correlated with text, but lack robustness. Experiments show that the proposed model achieves the best balance in speech quality, timbre similarity and robustness compares with baseline models. Furthermore, the whole modules in the proposed model can be trained jointly and no more pre-training data is needed.

This paper presents the "Ethiopian" system for the SLT 2021 Children Speech Recognition Challenge. Various data processing and augmentation techniques are proposed to tackle children's speech recognition problem, especially the lack of the children's speech recognition training data issue. Detailed experiments are designed and conducted to show the effectiveness of each technique, across different speech recognition toolkits and model architectures. Step by step, we explain how we come up with our final system, which provides the state-of-the-art results in the SLT 2021 Children Speech Recognition Challenge, with 21.66% CER on the Track 1 evaluation set (4th place overall), and 16.53% CER on the Track 2 evaluation set (1st place overall). Post-challenge analysis shows that our system actually achieves 18.82% CER on the Track 1 evaluation set, but we submitted the wrong version to the challenge organizer for Track 1.

Current evaluation methods for soccer player performance either relies on rating from soccer experts or structured statistics of the match, such as shots and tackles. The former needs a lot of manpower and the evaluation is inevitably subjective. The latter can only record the quantity of a player's match events, but cannot reflect the quality (e.g., a wonderful shot or a terrible shot is regarded as a shot). To solve the above problems, an improved evaluation method for soccer player performance using affective computing is proposed. On the basis of statistics, our method also takes advantage of the text information of post-match reports, and employ the affective computing technology to quantify the quality of events. In this way, both the quantity and quality of events are considered. All the players in the Chinese Super League 2019 season are selected as evaluation objects, and the results show that the improved method can evaluate player performance more effectively and reasonably.

Sports big data technology has become an important technological means of modern sports competition. This paper realizes a novel soccer video event detection algorithm based on self-attention. It extracts key frames through self-attention mechanism, and then obtains the characteristics of time window level through NetVLAD network. Finally, each video clip is classified into 4 types of events (goals, red/yellow card, substitutions and others). The experimental results show that with the introduction of the self-attention mechanism, the classification accuracy on the SoccerNet data set has improved from 67.2% to 74.3%.

Articulatory features (AFs) provide language-independent attribute by exploiting the speech production knowledge. This paper proposes a cross-lingual automatic speech recognition (ASR) based on AF methods. Various neural network (NN) architectures are explored to extract cross-lingual AFs and their performance is studied. The architectures include muti-layer perception(MLP), convolutional NN (CNN) and long short-term memory recurrent NN (LSTM). In our cross-lingual setup, only the source language (English, representing a well-resourced language) is used to train the AF extractors. AFs are then generated for the target language (Mandarin, representing an under-resourced language) using the trained extractors. The frame-classification accuracy indicates that the LSTM has an ability to perform a knowledge transfer through the robust cross-lingual AFs from well-resourced to under-resourced language. The final ASR system is built using traditional approaches (e.g. hybrid models), combining AFs with conventional MFCCs. The results demonstrate that the cross-lingual AFs improve the performance in under-resourced ASR task even though the source and target languages come from different language family. Overall, the proposed cross-lingual ASR approach provides slight improvement over the monolingual LF-MMI and cross-lingual (acoustic model adaptation-based) ASR systems.