Electrical Engineering And Systems Science

Recent advances of end-to-end models have outperformed conventional models through employing a two-pass model. The two-pass model provides better speed-quality trade-offs for on-device speech recognition, where a 1st-pass model generates hypotheses in a streaming fashion, and a 2nd-pass model re-scores the hypotheses with full audio sequence context. The 2nd-pass model plays a key role in the quality improvement of the end-to-end model to surpass the conventional model. One main challenge of the two-pass model is the computation latency introduced by the 2nd-pass model. Specifically, the original design of the two-pass model uses LSTMs for the 2nd-pass model, which are subject to long latency as they are constrained by the recurrent nature and have to run inference sequentially. In this work we explore replacing the LSTM layers in the 2nd-pass rescorer with Transformer layers, which can process the entire hypothesis sequences in parallel and can therefore utilize the on-device computation resources more efficiently. Compared with an LSTM-based baseline, our proposed Transformer rescorer achieves more than 50% latency reduction with quality improvement.

Peking Opera has been the most dominant form of Chinese performing art since around 200 years ago. A Peking Opera singer usually exhibits a very strong personal style via introducing improvisation and expressiveness on stage which leads the actual rhythm and pitch contour to deviate significantly from the original music score. This inconsistency poses a great challenge in Peking Opera singing voice synthesis from a music score. In this work, we propose to deal with this issue and synthesize expressive Peking Opera singing from the music score based on the Duration Informed Attention Network (DurIAN) framework. To tackle the rhythm mismatch, Lagrange multiplier is used to find the optimal output phoneme duration sequence with the constraint of the given note duration from music score. As for the pitch contour mismatch, instead of directly inferring from music score, we adopt a pseudo music score generated from the real singing and feed it as input during training. The experiments demonstrate that with the proposed system we can synthesize Peking Opera singing voice with high-quality timbre, pitch and expressiveness.

We propose PeriodNet, a non-autoregressive (non-AR) waveform generation model with a new model structure for modeling periodic and aperiodic components in speech waveforms. The non-AR waveform generation models can generate speech waveforms parallelly and can be used as a speech vocoder by conditioning an acoustic feature. Since a speech waveform contains periodic and aperiodic components, both components should be appropriately modeled to generate a high-quality speech waveform. However, it is difficult to decompose the components from a natural speech waveform in advance. To address this issue, we propose a parallel model and a series model structure separating periodic and aperiodic components. The features of our proposed models are that explicit periodic and aperiodic signals are taken as input, and external periodic/aperiodic decomposition is not needed in training. Experiments using a singing voice corpus show that our proposed structure improves the naturalness of the generated waveform. We also show that the speech waveforms with a pitch outside of the training data range can be generated with more naturalness.

Existing prescriptive compression strategies used in hearing aid fitting are designed based on gain averages from a group of users which are not necessarily optimal for a specific user. Nearly half of hearing aid users prefer settings that differ from the commonly prescribed settings. This paper presents a human-in-loop deep reinforcement learning approach that personalizes hearing aid compression to achieve improved hearing perception. The developed approach is designed to learn a specific user's hearing preferences in order to optimize compression based on the user's feedbacks. Both simulation and subject testing results are reported which demonstrate the effectiveness of the developed personalized compression.

Recent years have witnessed significant improvement in ASR systems to recognize spoken utterances. However, it is still a challenging task for noisy and out-of-domain data, where substitution and deletion errors are prevalent in the transcribed text. These errors significantly degrade the performance of downstream tasks. In this work, we propose a BERT-style language model, referred to as PhonemeBERT, that learns a joint language model with phoneme sequence and ASR transcript to learn phonetic-aware representations that are robust to ASR errors. We show that PhonemeBERT can be used on downstream tasks using phoneme sequences as additional features, and also in low-resource setup where we only have ASR-transcripts for the downstream tasks with no phoneme information available. We evaluate our approach extensively by generating noisy data for three benchmark datasets - Stanford Sentiment Treebank, TREC and ATIS for sentiment, question and intent classification tasks respectively. The results of the proposed approach beats the state-of-the-art baselines comprehensively on each dataset.

Code-switching---the intra-utterance use of multiple languages---is prevalent across the world. Within text-to-speech (TTS), multilingual models have been found to enable code-switching. By modifying the linguistic input to sequence-to-sequence TTS, we show that code-switching is possible for languages unseen during training, even within monolingual models. We use a small set of phonological features derived from the International Phonetic Alphabet (IPA), such as vowel height and frontness, consonant place and manner. This allows the model topology to stay unchanged for different languages, and enables new, previously unseen feature combinations to be interpreted by the model. We show that this allows us to generate intelligible, code-switched speech in a new language at test time, including the approximation of sounds never seen in training.

Neural network applications generally benefit from larger-sized models, but for current speech enhancement models, larger scale networks often suffer from decreased robustness to the variety of real-world use cases beyond what is encountered in training data. We introduce several innovations that lead to better large neural networks for speech enhancement. The novel PoCoNet architecture is a convolutional neural network that, with the use of frequency-positional embeddings, is able to more efficiently build frequency-dependent features in the early layers. A semi-supervised method helps increase the amount of conversational training data by pre-enhancing noisy datasets, improving performance on real recordings. A new loss function biased towards preserving speech quality helps the optimization better match human perceptual opinions on speech quality. Ablation experiments and objective and human opinion metrics show the benefits of the proposed improvements.

This report proposes a polyphonic sound event detection (SED) method for the DCASE 2020 Challenge Task 4. The proposed SED method is based on semi-supervised learning to deal with the different combination of training datasets such as weakly labeled dataset, unlabeled dataset, and strongly labeled synthetic dataset. Especially, the target label of each audio clip from weakly labeled or unlabeled dataset is first predicted by using the mean teacher model that is the DCASE 2020 baseline. The data with predicted labels are used for training the proposed SED model, which consists of CNNs with skip connections and self-attention mechanism, followed by RNNs. In order to compensate for the erroneous prediction of weakly labeled and unlabeled data, a semi-supervised loss function is employed for the proposed SED model. In this work, several versions of the proposed SED model are implemented and evaluated on the validation set according to the different parameter setting for the semi-supervised loss function, and then an ensemble model that combines five-fold validation models is finally selected as our final model.

The Voice Conversion Challenge 2020 is the third edition under its flagship that promotes intra-lingual semiparallel and cross-lingual voice conversion (VC). While the primary evaluation of the challenge submissions was done through crowd-sourced listening tests, we also performed an objective assessment of the submitted systems. The aim of the objective assessment is to provide complementary performance analysis that may be more beneficial than the time-consuming listening tests. In this study, we examined five types of objective assessments using automatic speaker verification (ASV), neural speaker embeddings, spoofing countermeasures, predicted mean opinion scores (MOS), and automatic speech recognition (ASR). Each of these objective measures assesses the VC output along different aspects. We observed that the correlations of these objective assessments with the subjective results were high for ASV, neural speaker embedding, and ASR, which makes them more influential for predicting subjective test results. In addition, we performed spoofing assessments on the submitted systems and identified some of the VC methods showing a potentially high security risk.

Spoken language understanding is typically based on pipeline architectures including speech recognition and natural language understanding steps. These components are optimized independently to allow usage of available data, but the overall system suffers from error propagation. In this paper, we propose a novel training method that enables pretrained contextual embeddings to process acoustic features. In particular, we extend it with an encoder of pretrained speech recognition systems in order to construct end-to-end spoken language understanding systems. Our proposed method is based on the teacher-student framework across speech and text modalities that aligns the acoustic and the semantic latent spaces. Experimental results in three benchmarks show that our system reaches the performance comparable to the pipeline architecture without using any training data and outperforms it after fine-tuning with ten examples per class on two out of three benchmarks.