Jiangyan Yi

Jiangyan Yi, Chenglong Wang, Jianhua Tao et al.

Many datasets have been designed to further the development of fake audio detection. However, fake utterances in previous datasets are mostly generated by altering timbre, prosody, linguistic content or channel noise of original audio. These datasets leave out a scenario, in which the acoustic scene of an original audio is manipulated with a forged one. It will pose a major threat to our society if some people misuse the manipulated audio with malicious purpose. Therefore, this motivates us to fill in the gap. This paper proposes such a dataset for scene fake audio detection named SceneFake, where a manipulated audio is generated by only tampering with the acoustic scene of an real utterance by using speech enhancement technologies. Some scene fake audio detection benchmark results on the SceneFake dataset are reported in this paper. In addition, an analysis of fake attacks with different speech enhancement technologies and signal-to-noise ratios are presented in this paper. The results indicate that scene fake utterances cannot be reliably detected by baseline models trained on the ASVspoof 2019 dataset. Although these models perform well on the SceneFake training set and seen testing set, their performance is poor on the unseen test set. The dataset (https://zenodo.org/record/7663324#.Y_XKMuPYuUk) and benchmark source codes (https://github.com/ADDchallenge/SceneFake) are publicly available.

Jiafeng Zhong, Bin Li, Jiangyan Yi

The task of partially spoofed audio localization aims to accurately determine audio authenticity at a frame level. Although some works have achieved encouraging results, utilizing boundary information within a single model remains an unexplored research topic. In this work, we propose a novel method called Boundary-aware Attention Mechanism (BAM). Specifically, it consists of two core modules: Boundary Enhancement and Boundary Frame-wise Attention. The former assembles the intra-frame and inter-frame information to extract discriminative boundary features that are subsequently used for boundary position detection and authenticity decision, while the latter leverages boundary prediction results to explicitly control the feature interaction between frames, which achieves effective discrimination between real and fake frames. Experimental results on PartialSpoof database demonstrate our proposed method achieves the best performance. The code is available at https://github.com/media-sec-lab/BAM.

Zheng Lian, Haiyang Sun, Licai Sun et al.

The first Multimodal Emotion Recognition Challenge (MER 2023) was successfully held at ACM Multimedia. The challenge focuses on system robustness and consists of three distinct tracks: (1) MER-MULTI, where participants are required to recognize both discrete and dimensional emotions; (2) MER-NOISE, in which noise is added to test videos for modality robustness evaluation; (3) MER-SEMI, which provides a large amount of unlabeled samples for semi-supervised learning. In this paper, we introduce the motivation behind this challenge, describe the benchmark dataset, and provide some statistics about participants. To continue using this dataset after MER 2023, please sign a new End User License Agreement and send it to our official email address merchallenge.contact@gmail.com. We believe this high-quality dataset can become a new benchmark in multimodal emotion recognition, especially for the Chinese research community.

Xinrui Yan, Jiangyan Yi, Jianhua Tao et al.

Many effective attempts have been made for fake audio detection. However, they can only provide detection results but no countermeasures to curb this harm. For many related practical applications, what model or algorithm generated the fake audio also is needed. Therefore, We propose a new problem for detecting vocoder fingerprints of fake audio. Experiments are conducted on the datasets synthesized by eight state-of-the-art vocoders. We have preliminarily explored the features and model architectures. The t-SNE visualization shows that different vocoders generate distinct vocoder fingerprints.

Jun Xue, Cunhang Fan, Zhao Lv et al.

Recently, pioneer research works have proposed a large number of acoustic features (log power spectrogram, linear frequency cepstral coefficients, constant Q cepstral coefficients, etc.) for audio deepfake detection, obtaining good performance, and showing that different subbands have different contributions to audio deepfake detection. However, this lacks an explanation of the specific information in the subband, and these features also lose information such as phase. Inspired by the mechanism of synthetic speech, the fundamental frequency (F0) information is used to improve the quality of synthetic speech, while the F0 of synthetic speech is still too average, which differs significantly from that of real speech. It is expected that F0 can be used as important information to discriminate between bonafide and fake speech, while this information cannot be used directly due to the irregular distribution of F0. Insteadly, the frequency band containing most of F0 is selected as the input feature. Meanwhile, to make full use of the phase and full-band information, we also propose to use real and imaginary spectrogram features as complementary input features and model the disjoint subbands separately. Finally, the results of F0, real and imaginary spectrogram features are fused. Experimental results on the ASVspoof 2019 LA dataset show that our proposed system is very effective for the audio deepfake detection task, achieving an equivalent error rate (EER) of 0.43%, which surpasses almost all systems.

Chenglong Wang, Jiangyan Yi, Jianhua Tao et al.

The existing fake audio detection systems often rely on expert experience to design the acoustic features or manually design the hyperparameters of the network structure. However, artificial adjustment of the parameters can have a relatively obvious influence on the results. It is almost impossible to manually set the best set of parameters. Therefore this paper proposes a fully automated end-toend fake audio detection method. We first use wav2vec pre-trained model to obtain a high-level representation of the speech. Furthermore, for the network structure, we use a modified version of the differentiable architecture search (DARTS) named light-DARTS. It learns deep speech representations while automatically learning and optimizing complex neural structures consisting of convolutional operations and residual blocks. The experimental results on the ASVspoof 2019 LA dataset show that our proposed system achieves an equal error rate (EER) of 1.08%, which outperforms the state-of-the-art single system.

Xiaohui Zhang, Jiangyan Yi, Jianhua Tao et al.

Current fake audio detection algorithms have achieved promising performances on most datasets. However, their performance may be significantly degraded when dealing with audio of a different dataset. The orthogonal weight modification to overcome catastrophic forgetting does not consider the similarity of genuine audio across different datasets. To overcome this limitation, we propose a continual learning algorithm for fake audio detection to overcome catastrophic forgetting, called Regularized Adaptive Weight Modification (RAWM). When fine-tuning a detection network, our approach adaptively computes the direction of weight modification according to the ratio of genuine utterances and fake utterances. The adaptive modification direction ensures the network can effectively detect fake audio on the new dataset while preserving its knowledge of old model, thus mitigating catastrophic forgetting. In addition, genuine audio collected from quite different acoustic conditions may skew their feature distribution, so we introduce a regularization constraint to force the network to remember the old distribution in this regard. Our method can easily be generalized to related fields, like speech emotion recognition. We also evaluate our approach across multiple datasets and obtain a significant performance improvement on cross-dataset experiments.

Jun Xue, Cunhang Fan, Jiangyan Yi et al.

In this paper, we propose a novel self-distillation method for fake speech detection (FSD), which can significantly improve the performance of FSD without increasing the model complexity. For FSD, some fine-grained information is very important, such as spectrogram defects, mute segments, and so on, which are often perceived by shallow networks. However, shallow networks have much noise, which can not capture this very well. To address this problem, we propose using the deepest network instruct shallow network for enhancing shallow networks. Specifically, the networks of FSD are divided into several segments, the deepest network being used as the teacher model, and all shallow networks become multiple student models by adding classifiers. Meanwhile, the distillation path between the deepest network feature and shallow network features is used to reduce the feature difference. A series of experimental results on the ASVspoof 2019 LA and PA datasets show the effectiveness of the proposed method, with significant improvements compared to the baseline.

Chenglong Wang, Jiangyan Yi, Xiaohui Zhang et al.

Self-supervised speech models are a rapidly developing research topic in fake audio detection. Many pre-trained models can serve as feature extractors, learning richer and higher-level speech features. However,when fine-tuning pre-trained models, there is often a challenge of excessively long training times and high memory consumption, and complete fine-tuning is also very expensive. To alleviate this problem, we apply low-rank adaptation(LoRA) to the wav2vec2 model, freezing the pre-trained model weights and injecting a trainable rank-decomposition matrix into each layer of the transformer architecture, greatly reducing the number of trainable parameters for downstream tasks. Compared with fine-tuning with Adam on the wav2vec2 model containing 317M training parameters, LoRA achieved similar performance by reducing the number of trainable parameters by 198 times.

Cunhang Fan, Hongyu Zhang, Wei Huang et al.

Auditory Attention Detection (AAD) aims to detect target speaker from brain signals in a multi-speaker environment. Although EEG-based AAD methods have shown promising results in recent years, current approaches primarily rely on traditional convolutional neural network designed for processing Euclidean data like images. This makes it challenging to handle EEG signals, which possess non-Euclidean characteristics. In order to address this problem, this paper proposes a dynamical graph self-distillation (DGSD) approach for AAD, which does not require speech stimuli as input. Specifically, to effectively represent the non-Euclidean properties of EEG signals, dynamical graph convolutional networks are applied to represent the graph structure of EEG signals, which can also extract crucial features related to auditory spatial attention in EEG signals. In addition, to further improve AAD detection performance, self-distillation, consisting of feature distillation and hierarchical distillation strategies at each layer, is integrated. These strategies leverage features and classification results from the deepest network layers to guide the learning of shallow layers. Our experiments are conducted on two publicly available datasets, KUL and DTU. Under a 1-second time window, we achieve results of 90.0\% and 79.6\% accuracy on KUL and DTU, respectively. We compare our DGSD method with competitive baselines, and the experimental results indicate that the detection performance of our proposed DGSD method is not only superior to the best reproducible baseline but also significantly reduces the number of trainable parameters by approximately 100 times.

Xiaohui Zhang, Jiangyan Yi, Jianhua Tao et al.

The rapid advancement of spoofing algorithms necessitates the development of robust detection methods capable of accurately identifying emerging fake audio. Traditional approaches, such as finetuning on new datasets containing these novel spoofing algorithms, are computationally intensive and pose a risk of impairing the acquired knowledge of known fake audio types. To address these challenges, this paper proposes an innovative approach that mitigates the limitations associated with finetuning. We introduce the concept of training low-rank adaptation matrices tailored specifically to the newly emerging fake audio types. During the inference stage, these adaptation matrices are combined with the existing model to generate the final prediction output. Extensive experimentation is conducted to evaluate the efficacy of the proposed method. The results demonstrate that our approach effectively preserves the prediction accuracy of the existing model for known fake audio types. Furthermore, our approach offers several advantages, including reduced storage memory requirements and lower equal error rates compared to conventional finetuning methods, particularly on specific spoofing algorithms.

Tao Wang, Jiangyan Yi, Ruibo Fu et al.

Text-based speech editing allows users to edit speech by intuitively cutting, copying, and pasting text to speed up the process of editing speech. In the previous work, CampNet (context-aware mask prediction network) is proposed to realize text-based speech editing, significantly improving the quality of edited speech. This paper aims at a new task: adding emotional effect to the editing speech during the text-based speech editing to make the generated speech more expressive. To achieve this task, we propose Emo-CampNet (emotion CampNet), which can provide the option of emotional attributes for the generated speech in text-based speech editing and has the one-shot ability to edit unseen speakers' speech. Firstly, we propose an end-to-end emotion-selectable text-based speech editing model. The key idea of the model is to control the emotion of generated speech by introducing additional emotion attributes based on the context-aware mask prediction network. Secondly, to prevent the emotion of the generated speech from being interfered by the emotional components in the original speech, a neutral content generator is proposed to remove the emotion from the original speech, which is optimized by the generative adversarial framework. Thirdly, two data augmentation methods are proposed to enrich the emotional and pronunciation information in the training set, which can enable the model to edit the unseen speaker's speech. The experimental results that 1) Emo-CampNet can effectively control the emotion of the generated speech in the process of text-based speech editing; And can edit unseen speakers' speech. 2) Detailed ablation experiments further prove the effectiveness of emotional selectivity and data augmentation methods. The demo page is available at https://hairuo55.github.io/Emo-CampNet/

Xinrui Yan, Jiangyan Yi, Jianhua Tao et al.

The rapid progress of deep speech synthesis models has posed significant threats to society such as malicious manipulation of content. This has led to an increase in studies aimed at detecting so-called deepfake audio. However, existing works focus on the binary detection of real audio and fake audio. In real-world scenarios such as model copyright protection and digital evidence forensics, binary classification alone is insufficient. It is essential to identify the source of deepfake audio. Therefore, audio deepfake attribution has emerged as a new challenge. To this end, we designed the first deepfake audio dataset for the attribution of audio generation tools, called Audio Deepfake Attribution (ADA), and conducted a comprehensive investigation on system fingerprints. To address the challenges of attribution of continuously emerging unknown audio generation tools in the real world, we propose the Class-Representation Multi-Center Learning (CRML) method for open-set audio deepfake attribution (OSADA). CRML enhances the global directional variation of representations, ensuring the learning of discriminative representations with strong intra-class similarity and inter-class discrepancy among known classes. Finally, the strong class discrimination capability learned from known classes is extended to both known and unknown classes. Experimental results demonstrate that the CRML method effectively addresses open-set risks in real-world scenarios. The dataset is publicly available at: https://zenodo.org/records/13318702, and https://zenodo.org/records/13340666.

Chunyu Qiang, Wang Geng, Yi Zhao et al.

Deep learning has brought significant improvements to the field of cross-modal representation learning. For tasks such as text-to-speech (TTS), voice conversion (VC), and automatic speech recognition (ASR), a cross-modal fine-grained (frame-level) sequence representation is desired, emphasizing the semantic content of the text modality while de-emphasizing the paralinguistic information of the speech modality. We propose a method called "Vector Quantized Contrastive Token-Acoustic Pre-training (VQ-CTAP)", which uses the cross-modal aligned sequence transcoder to bring text and speech into a joint multimodal space, learning how to connect text and speech at the frame level. The proposed VQ-CTAP is a paradigm for cross-modal sequence representation learning, offering a promising solution for fine-grained generation and recognition tasks in speech processing. The VQ-CTAP can be directly applied to VC and ASR tasks without fine-tuning or additional structures. We propose a sequence-aware semantic connector, which connects multiple frozen pre-trained modules for the TTS task, exhibiting a plug-and-play capability. We design a stepping optimization strategy to ensure effective model convergence by gradually injecting and adjusting the influence of various loss components. Furthermore, we propose a semantic-transfer-wise paralinguistic consistency loss to enhance representational capabilities, allowing the model to better generalize to unseen data and capture the nuances of paralinguistic information. In addition, VQ-CTAP achieves high-compression speech coding at a rate of 25Hz from 24kHz input waveforms, which is a 960-fold reduction in the sampling rate. The audio demo is available at https://qiangchunyu.github.io/VQCTAP/

Haogeng Liu, Tao Wang, Ruibo Fu et al.

Text-to-speech (TTS) and voice conversion (VC) are two different tasks both aiming at generating high quality speaking voice according to different input modality. Due to their similarity, this paper proposes UnifySpeech, which brings TTS and VC into a unified framework for the first time. The model is based on the assumption that speech can be decoupled into three independent components: content information, speaker information, prosody information. Both TTS and VC can be regarded as mining these three parts of information from the input and completing the reconstruction of speech. For TTS, the speech content information is derived from the text, while in VC it's derived from the source speech, so all the remaining units are shared except for the speech content extraction module in the two tasks. We applied vector quantization and domain constrain to bridge the gap between the content domains of TTS and VC. Objective and subjective evaluation shows that by combining the two task, TTS obtains better speaker modeling ability while VC gets hold of impressive speech content decoupling capability.

Tao Wang, Ruibo Fu, Jiangyan Yi et al.

The traditional vocoders have the advantages of high synthesis efficiency, strong interpretability, and speech editability, while the neural vocoders have the advantage of high synthesis quality. To combine the advantages of two vocoders, inspired by the traditional deterministic plus stochastic model, this paper proposes a novel neural vocoder named NeuralDPS which can retain high speech quality and acquire high synthesis efficiency and noise controllability. Firstly, this framework contains four modules: a deterministic source module, a stochastic source module, a neural V/UV decision module and a neural filter module. The input required by the vocoder is just the spectral parameter, which avoids the error caused by estimating additional parameters, such as F0. Secondly, to solve the problem that different frequency bands may have different proportions of deterministic components and stochastic components, a multiband excitation strategy is used to generate a more accurate excitation signal and reduce the neural filter's burden. Thirdly, a method to control noise components of speech is proposed. In this way, the signal-to-noise ratio (SNR) of speech can be adjusted easily. Objective and subjective experimental results show that our proposed NeuralDPS vocoder can obtain similar performance with the WaveNet and it generates waveforms at least 280 times faster than the WaveNet vocoder. It is also 28% faster than WaveGAN's synthesis efficiency on a single CPU core. We have also verified through experiments that this method can effectively control the noise components in the predicted speech and adjust the SNR of speech. Examples of generated speech can be found at https://hairuo55.github.io/NeuralDPS.

Siding Zeng, Jiangyan Yi, Jianhua Tao et al.

When the task of locating manipulation regions in partially-fake audio (PFA) involves cross-domain datasets, the performance of deep learning models drops significantly due to the shift between the source and target domains. To address this issue, existing approaches often employ data augmentation before training. However, they overlook the characteristics in target domain that are absent in source domain. Inspired by the mixture-of-experts model, we propose an unsupervised method named Samples mining with Diversity and Entropy (SDE). Our method first learns from a collection of diverse experts that achieve great performance from different perspectives in the source domain, but with ambiguity on target samples. We leverage these diverse experts to select the most informative samples by calculating their entropy. Furthermore, we introduced a label generation method tailored for these selected samples that are incorporated in the training process in source domain integrating the target domain information. We applied our method to a cross-domain partially fake audio detection dataset, ADD2023Track2. By introducing 10% of unknown samples from the target domain, we achieved an F1 score of 43.84%, which represents a relative increase of 77.2% compared to the second-best method.

Zheng Lian, Haiyang Sun, Licai Sun et al.

Multimodal emotion recognition is an important research topic in artificial intelligence. Over the past few decades, researchers have made remarkable progress by increasing the dataset size and building more effective algorithms. However, due to problems such as complex environments and inaccurate annotations, current systems are hard to meet the demands of practical applications. Therefore, we organize the MER series of competitions to promote the development of this field. Last year, we launched MER2023, focusing on three interesting topics: multi-label learning, noise robustness, and semi-supervised learning. In this year's MER2024, besides expanding the dataset size, we further introduce a new track around open-vocabulary emotion recognition. The main purpose of this track is that existing datasets usually fix the label space and use majority voting to enhance the annotator consistency. However, this process may lead to inaccurate annotations, such as ignoring non-majority or non-candidate labels. In this track, we encourage participants to generate any number of labels in any category, aiming to describe emotional states as accurately as possible. Our baseline code relies on MERTools and is available at: https://github.com/zeroQiaoba/MERTools/tree/master/MER2024.

Yujie Chen, Jiangyan Yi, Cunhang Fan et al.

Rapid advancements in speech synthesis and voice conversion bring convenience but also new security risks, creating an urgent need for effective audio deepfake detection. Although current models perform well, their effectiveness diminishes when confronted with the diverse and evolving nature of real-world deepfakes. To address this issue, we propose a continual learning method named Region-Based Optimization (RegO) for audio deepfake detection. Specifically, we use the Fisher information matrix to measure important neuron regions for real and fake audio detection, dividing them into four regions. First, we directly fine-tune the less important regions to quickly adapt to new tasks. Next, we apply gradient optimization in parallel for regions important only to real audio detection, and in orthogonal directions for regions important only to fake audio detection. For regions that are important to both, we use sample proportion-based adaptive gradient optimization. This region-adaptive optimization ensures an appropriate trade-off between memory stability and learning plasticity. Additionally, to address the increase of redundant neurons from old tasks, we further introduce the Ebbinghaus forgetting mechanism to release them, thereby promoting the capability of the model to learn more generalized discriminative features. Experimental results show our method achieves a 21.3% improvement in EER over the state-of-the-art continual learning approach RWM for audio deepfake detection. Moreover, the effectiveness of RegO extends beyond the audio deepfake detection domain, showing potential significance in other tasks, such as image recognition. The code is available at https://github.com/cyjie429/RegO

Chunyu Qiang, Haoyu Wang, Cheng Gong et al.

Speech codecs serve as a crucial bridge in unifying speech and text language models. Existing codec methods face several challenges in semantic encoding, such as residual paralinguistic information (e.g., timbre, emotion), insufficient semantic completeness, limited reconstruction capability, and lack of support for streaming. To address these challenges, we propose SecoustiCodec, a cross-modal aligned low-bitrate streaming speech codec that disentangles semantic and paralinguistic information in a single-codebook space. To ensure semantic completeness and reconstruction fidelity, paralinguistic encoding is introduced to bridge the information gap between semantic and acoustic encoding. A semantic-only efficient quantization method based on VAE (Variational Autoencoder) and FSQ (Finite Scalar Quantization) is proposed. This approach alleviates the long-tail distribution problem of tokens while maintaining high codebook utilization. A semantic disentanglement method based on contrastive learning is proposed, which aligns text and speech in a joint multimodal frame-level space, effectively removing paralinguistic information from semantic encoding. An acoustic-constrained multi-stage optimization strategy is proposed to ensure robust and stable convergence. Figure~\ref{fig:pesq_kbps_below_2kbps} shows SecoustiCodec achieves SOTA (state-of-the-art) reconstruction quality (PESQ) of 1.77/2.58 at 0.27/1 kbps. The code and model weights for SecoustiCodec will be open-sourced upon the completion of the peer-review process. We've open-sourced SecoustiCodec's demo, code, and model weights.

Kang Chen, Zheng Lian, Haiyang Sun et al.

Deception detection has attracted increasing attention due to its importance in real-world scenarios. Its main goal is to detect deceptive behaviors from multimodal clues such as gestures, facial expressions, prosody, etc. However, these bases are usually subjective and related to personal habits. Therefore, we extend deception detection to deception reasoning, further providing objective evidence to support subjective judgment. Specifically, we provide potential lies and basic facts and then analyze why this sentence may be a lie by combining factual inconsistencies and intent behind them. Compared with deception detection, this task is more applicable to real-world scenarios. For example, in interrogation, the police should judge whether a person is lying based on solid evidence. This paper presents our initial attempts at this task, including constructing a dataset and defining evaluation metrics. Meanwhile, this task can serve as a benchmark for evaluating the complex reasoning capability of large language models. Our code and data are provided in the supplementary material.

Cunhang Fan, Jiangyan Yi, Jianhua Tao et al.

The joint training framework for speech enhancement and recognition methods have obtained quite good performances for robust end-to-end automatic speech recognition (ASR). However, these methods only utilize the enhanced feature as the input of the speech recognition component, which are affected by the speech distortion problem. In order to address this problem, this paper proposes a gated recurrent fusion (GRF) method with joint training framework for robust end-to-end ASR. The GRF algorithm is used to dynamically combine the noisy and enhanced features. Therefore, the GRF can not only remove the noise signals from the enhanced features, but also learn the raw fine structures from the noisy features so that it can alleviate the speech distortion. The proposed method consists of speech enhancement, GRF and speech recognition. Firstly, the mask based speech enhancement network is applied to enhance the input speech. Secondly, the GRF is applied to address the speech distortion problem. Thirdly, to improve the performance of ASR, the state-of-the-art speech transformer algorithm is used as the speech recognition component. Finally, the joint training framework is utilized to optimize these three components, simultaneously. Our experiments are conducted on an open-source Mandarin speech corpus called AISHELL-1. Experimental results show that the proposed method achieves the relative character error rate (CER) reduction of 10.04\% over the conventional joint enhancement and transformer method only using the enhanced features. Especially for the low signal-to-noise ratio (0 dB), our proposed method can achieves better performances with 12.67\% CER reduction, which suggests the potential of our proposed method.

Xiaohui Zhang, Jiangyan Yi, Chenglong Wang et al.

The rapid evolution of speech synthesis and voice conversion has raised substantial concerns due to the potential misuse of such technology, prompting a pressing need for effective audio deepfake detection mechanisms. Existing detection models have shown remarkable success in discriminating known deepfake audio, but struggle when encountering new attack types. To address this challenge, one of the emergent effective approaches is continual learning. In this paper, we propose a continual learning approach called Radian Weight Modification (RWM) for audio deepfake detection. The fundamental concept underlying RWM involves categorizing all classes into two groups: those with compact feature distributions across tasks, such as genuine audio, and those with more spread-out distributions, like various types of fake audio. These distinctions are quantified by means of the in-class cosine distance, which subsequently serves as the basis for RWM to introduce a trainable gradient modification direction for distinct data types. Experimental evaluations against mainstream continual learning methods reveal the superiority of RWM in terms of knowledge acquisition and mitigating forgetting in audio deepfake detection. Furthermore, RWM's applicability extends beyond audio deepfake detection, demonstrating its potential significance in diverse machine learning domains such as image recognition.

Xinrui Yan, Jiangyan Yi, Jianhua Tao et al.

Open environment oriented open set model attribution of deepfake audio is an emerging research topic, aiming to identify the generation models of deepfake audio. Most previous work requires manually setting a rejection threshold for unknown classes to compare with predicted probabilities. However, models often overfit training instances and generate overly confident predictions. Moreover, thresholds that effectively distinguish unknown categories in the current dataset may not be suitable for identifying known and unknown categories in another data distribution. To address the issues, we propose a novel framework for open set model attribution of deepfake audio with rejection threshold adaptation (ReTA). Specifically, the reconstruction error learning module trains by combining the representation of system fingerprints with labels corresponding to either the target class or a randomly chosen other class label. This process generates matching and non-matching reconstructed samples, establishing the reconstruction error distributions for each class and laying the foundation for the reject threshold calculation module. The reject threshold calculation module utilizes gaussian probability estimation to fit the distributions of matching and non-matching reconstruction errors. It then computes adaptive reject thresholds for all classes through probability minimization criteria. The experimental results demonstrate the effectiveness of ReTA in improving the open set model attributes of deepfake audio.

Muhammad Sharif, Zeeshan Abbas, Jiangyan Yi et al.

Automatic Speech Recognition (ASR) technology has witnessed significant advancements in recent years, revolutionizing human-computer interactions. While major languages have benefited from these developments, lesser-resourced languages like Urdu face unique challenges. This paper provides an extensive exploration of the dynamic landscape of ASR research, focusing particularly on the resource-constrained Urdu language, which is widely spoken across South Asian nations. It outlines current research trends, technological advancements, and potential directions for future studies in Urdu ASR, aiming to pave the way for forthcoming researchers interested in this domain. By leveraging contemporary technologies, analyzing existing datasets, and evaluating effective algorithms and tools, the paper seeks to shed light on the unique challenges and opportunities associated with Urdu language processing and its integration into the broader field of speech research.

Muhammad Sharif, Jiangyan Yi, Muhammad Shoaib

The language called Balti belongs to the Sino-Tibetan, specifically the Tibeto-Burman language family. It is understood with variations, across populations in India, China, Pakistan, Nepal, Tibet, Burma, and Bhutan, influenced by local cultures and producing various dialects. Considering the diverse cultural, socio-political, religious, and geographical impacts, it is important to step forward unifying the dialects, the basis of common root, lexica, and phonological perspectives, is vital. In the era of globalization and the increasingly frequent developments in AI technology, understanding the diversity and the efforts of dialect unification is important to understanding commonalities and shortening the gaps impacted by unavoidable circumstances. This article analyzes and examines how artificial intelligence AI in the essence of Large Language Models LLMs, can assist in analyzing, documenting, and standardizing the endangered Balti Language, based on the efforts made in different dialects so far.

Tao Wang, Jiangyan Yi, Ruibo Fu et al.

The text-based speech editor allows the editing of speech through intuitive cutting, copying, and pasting operations to speed up the process of editing speech. However, the major drawback of current systems is that edited speech often sounds unnatural due to cut-copy-paste operation. In addition, it is not obvious how to synthesize records according to a new word not appearing in the transcript. This paper proposes a novel end-to-end text-based speech editing method called context-aware mask prediction network (CampNet). The model can simulate the text-based speech editing process by randomly masking part of speech and then predicting the masked region by sensing the speech context. It can solve unnatural prosody in the edited region and synthesize the speech corresponding to the unseen words in the transcript. Secondly, for the possible operation of text-based speech editing, we design three text-based operations based on CampNet: deletion, insertion, and replacement. These operations can cover various situations of speech editing. Thirdly, to synthesize the speech corresponding to long text in insertion and replacement operations, a word-level autoregressive generation method is proposed. Fourthly, we propose a speaker adaptation method using only one sentence for CampNet and explore the ability of few-shot learning based on CampNet, which provides a new idea for speech forgery tasks. The subjective and objective experiments on VCTK and LibriTTS datasets show that the speech editing results based on CampNet are better than TTS technology, manual editing, and VoCo method. We also conduct detailed ablation experiments to explore the effect of the CampNet structure on its performance. Finally, the experiment shows that speaker adaptation with only one sentence can further improve the naturalness of speech. Examples of generated speech can be found at https://hairuo55.github.io/CampNet.

Jiangyan Yi, Ruibo Fu, Jianhua Tao et al.

Audio deepfake detection is an emerging topic, which was included in the ASVspoof 2021. However, the recent shared tasks have not covered many real-life and challenging scenarios. The first Audio Deep synthesis Detection challenge (ADD) was motivated to fill in the gap. The ADD 2022 includes three tracks: low-quality fake audio detection (LF), partially fake audio detection (PF) and audio fake game (FG). The LF track focuses on dealing with bona fide and fully fake utterances with various real-world noises etc. The PF track aims to distinguish the partially fake audio from the real. The FG track is a rivalry game, which includes two tasks: an audio generation task and an audio fake detection task. In this paper, we describe the datasets, evaluation metrics, and protocols. We also report major findings that reflect the recent advances in audio deepfake detection tasks.

Tao Wang, Ruibo Fu, Jiangyan Yi et al.

End-to-end singing voice synthesis (SVS) is attractive due to the avoidance of pre-aligned data. However, the auto learned alignment of singing voice with lyrics is difficult to match the duration information in musical score, which will lead to the model instability or even failure to synthesize voice. To learn accurate alignment information automatically, this paper proposes an end-to-end SVS framework, named Singing-Tacotron. The main difference between the proposed framework and Tacotron is that the speech can be controlled significantly by the musical score's duration information. Firstly, we propose a global duration control attention mechanism for the SVS model. The attention mechanism can control each phoneme's duration. Secondly, a duration encoder is proposed to learn a set of global transition tokens from the musical score. These transition tokens can help the attention mechanism decide whether moving to the next phoneme or staying at each decoding step. Thirdly, to further improve the model's stability, a dynamic filter is designed to help the model overcome noise interference and pay more attention to local context information. Subjective and objective evaluation verify the effectiveness of the method. Furthermore, the role of global transition tokens and the effect of duration control are explored. Examples of experiments can be found at https://hairuo55.github.io/SingingTacotron.

Shuai Zhang, Jiangyan Yi, Zhengkun Tian et al.

Code-switching deals with alternative languages in communication process. Training end-to-end (E2E) automatic speech recognition (ASR) systems for code-switching is especially challenging as code-switching training data are always insufficient to combat the increased multilingual context confusion due to the presence of more than one language. We propose a language-related attention mechanism to reduce multilingual context confusion for the E2E code-switching ASR model based on the Equivalence Constraint (EC) Theory. The linguistic theory requires that any monolingual fragment that occurs in the code-switching sentence must occur in one of the monolingual sentences. The theory establishes a bridge between monolingual data and code-switching data. We leverage this linguistics theory to design the code-switching E2E ASR model. The proposed model efficiently transfers language knowledge from rich monolingual data to improve the performance of the code-switching ASR model. We evaluate our model on ASRU 2019 Mandarin-English code-switching challenge dataset. Compared to the baseline model, our proposed model achieves a 17.12% relative error reduction.

Haoxin Ma, Jiangyan Yi, Jianhua Tao et al.

Fake audio attack becomes a major threat to the speaker verification system. Although current detection approaches have achieved promising results on dataset-specific scenarios, they encounter difficulties on unseen spoofing data. Fine-tuning and retraining from scratch have been applied to incorporate new data. However, fine-tuning leads to performance degradation on previous data. Retraining takes a lot of time and computation resources. Besides, previous data are unavailable due to privacy in some situations. To solve the above problems, this paper proposes detecting fake without forgetting, a continual-learning-based method, to make the model learn new spoofing attacks incrementally. A knowledge distillation loss is introduced to loss function to preserve the memory of original model. Supposing the distribution of genuine voice is consistent among different scenarios, an extra embedding similarity loss is used as another constraint to further do a positive sample alignment. Experiments are conducted on the ASVspoof2019 dataset. The results show that our proposed method outperforms fine-tuning by the relative reduction of average equal error rate up to 81.62%.

Jiangyan Yi, Ye Bai, Jianhua Tao et al.

Diverse promising datasets have been designed to hold back the development of fake audio detection, such as ASVspoof databases. However, previous datasets ignore an attacking situation, in which the hacker hides some small fake clips in real speech audio. This poses a serious threat since that it is difficult to distinguish the small fake clip from the whole speech utterance. Therefore, this paper develops such a dataset for half-truth audio detection (HAD). Partially fake audio in the HAD dataset involves only changing a few words in an utterance.The audio of the words is generated with the very latest state-of-the-art speech synthesis technology. We can not only detect fake uttrances but also localize manipulated regions in a speech using this dataset. Some benchmark results are presented on this dataset. The results show that partially fake audio presents much more challenging than fully fake audio for fake audio detection. The HAD dataset is publicly available: https://zenodo.org/records/10377492.

Zhengkun Tian, Jiangyan Yi, Ye Bai et al.

Transducer-based models, such as RNN-Transducer and transformer-transducer, have achieved great success in speech recognition. A typical transducer model decodes the output sequence conditioned on the current acoustic state and previously predicted tokens step by step. Statistically, The number of blank tokens in the prediction results accounts for nearly 90\% of all tokens. It takes a lot of computation and time to predict the blank tokens, but only the non-blank tokens will appear in the final output sequence. Therefore, we propose a method named fast-skip regularization, which tries to align the blank position predicted by a transducer with that predicted by a CTC model. During the inference, the transducer model can predict the blank tokens in advance by a simple CTC project layer without many complicated forward calculations of the transducer decoder and then skip them, which will reduce the computation and improve the inference speed greatly. All experiments are conducted on a public Chinese mandarin dataset AISHELL-1. The results show that the fast-skip regularization can indeed help the transducer model learn the blank position alignments. Besides, the inference with fast-skip can be speeded up nearly 4 times with only a little performance degradation.

Zhengkun Tian, Jiangyan Yi, Jianhua Tao et al.

The autoregressive (AR) models, such as attention-based encoder-decoder models and RNN-Transducer, have achieved great success in speech recognition. They predict the output sequence conditioned on the previous tokens and acoustic encoded states, which is inefficient on GPUs. The non-autoregressive (NAR) models can get rid of the temporal dependency between the output tokens and predict the entire output tokens in at least one step. However, the NAR model still faces two major problems. On the one hand, there is still a great gap in performance between the NAR models and the advanced AR models. On the other hand, it's difficult for most of the NAR models to train and converge. To address these two problems, we propose a new model named the two-step non-autoregressive transformer(TSNAT), which improves the performance and accelerating the convergence of the NAR model by learning prior knowledge from a parameters-sharing AR model. Furthermore, we introduce the two-stage method into the inference process, which improves the model performance greatly. All the experiments are conducted on a public Chinese mandarin dataset ASIEHLL-1. The results show that the TSNAT can achieve a competitive performance with the AR model and outperform many complicated NAR models.

Ye Bai, Jiangyan Yi, Jianhua Tao et al.

Attention-based encoder-decoder (AED) models have achieved promising performance in speech recognition. However, because the decoder predicts text tokens (such as characters or words) in an autoregressive manner, it is difficult for an AED model to predict all tokens in parallel. This makes the inference speed relatively slow. We believe that because the encoder already captures the whole speech utterance, which has the token-level relationship implicitly, we can predict a token without explicitly autoregressive language modeling. When the prediction of a token does not rely on other tokens, the parallel prediction of all tokens in the sequence is realizable. Based on this idea, we propose a non-autoregressive speech recognition model called LASO (Listen Attentively, and Spell Once). The model consists of an encoder, a decoder, and a position dependent summarizer (PDS). The three modules are based on basic attention blocks. The encoder extracts high-level representations from the speech. The PDS uses positional encodings corresponding to tokens to convert the acoustic representations into token-level representations. The decoder further captures token-level relationships with the self-attention mechanism. At last, the probability distribution on the vocabulary is computed for each token position. Therefore, speech recognition is re-formulated as a position-wise classification problem. Further, we propose a cross-modal transfer learning method to refine semantics from a large-scale pre-trained language model BERT for improving the performance.

Cunhang Fan, Bin Liu, Jianhua Tao et al.

Recurrent neural networks (RNNs) have shown significant improvements in recent years for speech enhancement. However, the model complexity and inference time cost of RNNs are much higher than deep feed-forward neural networks (DNNs). Therefore, these limit the applications of speech enhancement. This paper proposes a deep time delay neural network (TDNN) for speech enhancement with full data learning. The TDNN has excellent potential for capturing long range temporal contexts, which utilizes a modular and incremental design. Besides, the TDNN preserves the feed-forward structure so that its inference cost is comparable to standard DNN. To make full use of the training data, we propose a full data learning method for speech enhancement. More specifically, we not only use the noisy-to-clean (input-to-target) to train the enhanced model, but also the clean-to-clean and noise-to-silence data. Therefore, all of the training data can be used to train the enhanced model. Our experiments are conducted on TIMIT dataset. Experimental results show that our proposed method could achieve a better performance than DNN and comparable even better performance than BLSTM. Meanwhile, compared with the BLSTM, the proposed method drastically reduce the inference time.

Shuai Zhang, Jiangyan Yi, Zhengkun Tian et al.

Despite the recent significant advances witnessed in end-to-end (E2E) ASR system for code-switching, hunger for audio-text paired data limits the further improvement of the models' performance. In this paper, we propose a decoupled transformer model to use monolingual paired data and unpaired text data to alleviate the problem of code-switching data shortage. The model is decoupled into two parts: audio-to-phoneme (A2P) network and phoneme-to-text (P2T) network. The A2P network can learn acoustic pattern scenarios using large-scale monolingual paired data. Meanwhile, it generates multiple phoneme sequence candidates for single audio data in real-time during the training process. Then the generated phoneme-text paired data is used to train the P2T network. This network can be pre-trained with large amounts of external unpaired text data. By using monolingual data and unpaired text data, the decoupled transformer model reduces the high dependency on code-switching paired training data of E2E model to a certain extent. Finally, the two networks are optimized jointly through attention fusion. We evaluate the proposed method on the public Mandarin-English code-switching dataset. Compared with our transformer baseline, the proposed method achieves 18.14% relative mix error rate reduction.

Zhengkun Tian, Jiangyan Yi, Jianhua Tao et al.

Non-autoregressive transformer models have achieved extremely fast inference speed and comparable performance with autoregressive sequence-to-sequence models in neural machine translation. Most of the non-autoregressive transformers decode the target sequence from a predefined-length mask sequence. If the predefined length is too long, it will cause a lot of redundant calculations. If the predefined length is shorter than the length of the target sequence, it will hurt the performance of the model. To address this problem and improve the inference speed, we propose a spike-triggered non-autoregressive transformer model for end-to-end speech recognition, which introduces a CTC module to predict the length of the target sequence and accelerate the convergence. All the experiments are conducted on a public Chinese mandarin dataset AISHELL-1. The results show that the proposed model can accurately predict the length of the target sequence and achieve a competitive performance with the advanced transformers. What's more, the model even achieves a real-time factor of 0.0056, which exceeds all mainstream speech recognition models.

Ye Bai, Jiangyan Yi, Jianhua Tao et al.

Although attention based end-to-end models have achieved promising performance in speech recognition, the multi-pass forward computation in beam-search increases inference time cost, which limits their practical applications. To address this issue, we propose a non-autoregressive end-to-end speech recognition system called LASO (listen attentively, and spell once). Because of the non-autoregressive property, LASO predicts a textual token in the sequence without the dependence on other tokens. Without beam-search, the one-pass propagation much reduces inference time cost of LASO. And because the model is based on the attention based feedforward structure, the computation can be implemented in parallel efficiently. We conduct experiments on publicly available Chinese dataset AISHELL-1. LASO achieves a character error rate of 6.4%, which outperforms the state-of-the-art autoregressive transformer model (6.7%). The average inference latency is 21 ms, which is 1/50 of the autoregressive transformer model.

Cunhang Fan, Jianhua Tao, Bin Liu et al.

Monaural speech dereverberation is a very challenging task because no spatial cues can be used. When the additive noises exist, this task becomes more challenging. In this paper, we propose a joint training method for simultaneous speech denoising and dereverberation using deep embedding features, which is based on the deep clustering (DC). DC is a state-of-the-art method for speech separation that includes embedding learning and K-means clustering. As for our proposed method, it contains two stages: denoising and dereverberation. At the denoising stage, the DC network is leveraged to extract noise-free deep embedding features. These embedding features are generated from the anechoic speech and residual reverberation signals. They can represent the inferred spectral masking patterns of the desired signals, which are discriminative features. At the dereverberation stage, instead of using the unsupervised K-means clustering algorithm, another supervised neural network is utilized to estimate the anechoic speech from these deep embedding features. Finally, the denoising stage and dereverberation stage are optimized by the joint training method. Experimental results show that the proposed method outperforms the WPE and BLSTM baselines, especially in the low SNR condition.

Jiangyan Yi, Jianhua Tao, Ye Bai et al.

Previous studies demonstrate that word embeddings and part-of-speech (POS) tags are helpful for punctuation restoration tasks. However, two drawbacks still exist. One is that word embeddings are pre-trained by unidirectional language modeling objectives. Thus the word embeddings only contain left-to-right context information. The other is that POS tags are provided by an external POS tagger. So computation cost will be increased and incorrect predicted tags may affect the performance of restoring punctuation marks during decoding. This paper proposes adversarial transfer learning to address these problems. A pre-trained bidirectional encoder representations from transformers (BERT) model is used to initialize a punctuation model. Thus the transferred model parameters carry both left-to-right and right-to-left representations. Furthermore, adversarial multi-task learning is introduced to learn task invariant knowledge for punctuation prediction. We use an extra POS tagging task to help the training of the punctuation predicting task. Adversarial training is utilized to prevent the shared parameters from containing task specific information. We only use the punctuation predicting task to restore marks during decoding stage. Therefore, it will not need extra computation and not introduce incorrect tags from the POS tagger. Experiments are conducted on IWSLT2011 datasets. The results demonstrate that the punctuation predicting models obtain further performance improvement with task invariant knowledge from the POS tagging task. Our best model outperforms the previous state-of-the-art model trained only with lexical features by up to 9.2% absolute overall F_1-score on test set.

Cunhang Fan, Jianhua Tao, Bin Liu et al.

In this paper, we propose an end-to-end post-filter method with deep attention fusion features for monaural speaker-independent speech separation. At first, a time-frequency domain speech separation method is applied as the pre-separation stage. The aim of pre-separation stage is to separate the mixture preliminarily. Although this stage can separate the mixture, it still contains the residual interference. In order to enhance the pre-separated speech and improve the separation performance further, the end-to-end post-filter (E2EPF) with deep attention fusion features is proposed. The E2EPF can make full use of the prior knowledge of the pre-separated speech, which contributes to speech separation. It is a fully convolutional speech separation network and uses the waveform as the input features. Firstly, the 1-D convolutional layer is utilized to extract the deep representation features for the mixture and pre-separated signals in the time domain. Secondly, to pay more attention to the outputs of the pre-separation stage, an attention module is applied to acquire deep attention fusion features, which are extracted by computing the similarity between the mixture and the pre-separated speech. These deep attention fusion features are conducive to reduce the interference and enhance the pre-separated speech. Finally, these features are sent to the post-filter to estimate each target signals. Experimental results on the WSJ0-2mix dataset show that the proposed method outperforms the state-of-the-art speech separation method. Compared with the pre-separation method, our proposed method can acquire 64.1%, 60.2%, 25.6% and 7.5% relative improvements in scale-invariant source-to-noise ratio (SI-SNR), the signal-to-distortion ratio (SDR), the perceptual evaluation of speech quality (PESQ) and the short-time objective intelligibility (STOI) measures, respectively.

Shuai Zhang, Jiangyan Yi, Zhengkun Tian et al.

Recently, language identity information has been utilized to improve the performance of end-to-end code-switching (CS) speech recognition. However, previous works use an additional language identification (LID) model as an auxiliary module, which causes the system complex. In this work, we propose an improved recurrent neural network transducer (RNN-T) model with language bias to alleviate the problem. We use the language identities to bias the model to predict the CS points. This promotes the model to learn the language identity information directly from transcription, and no additional LID model is needed. We evaluate the approach on a Mandarin-English CS corpus SEAME. Compared to our RNN-T baseline, the proposed method can achieve 16.2% and 12.9% relative error reduction on two test sets, respectively.

Cunhang Fan, Bin Liu, Jianhua Tao et al.

Multi-channel deep clustering (MDC) has acquired a good performance for speech separation. However, MDC only applies the spatial features as the additional information. So it is difficult to learn mutual relationship between spatial and spectral features. Besides, the training objective of MDC is defined at embedding vectors, rather than real separated sources, which may damage the separation performance. In this work, we propose a deep attention fusion method to dynamically control the weights of the spectral and spatial features and combine them deeply. In addition, to solve the training objective problem of MDC, the real separated sources are used as the training objectives. Specifically, we apply the deep clustering network to extract deep embedding features. Instead of using the unsupervised K-means clustering to estimate binary masks, another supervised network is utilized to learn soft masks from these deep embedding features. Our experiments are conducted on a spatialized reverberant version of WSJ0-2mix dataset. Experimental results show that the proposed method outperforms MDC baseline and even better than the oracle ideal binary mask (IBM).

Zhengkun Tian, Jiangyan Yi, Ye Bai et al.

For most of the attention-based sequence-to-sequence models, the decoder predicts the output sequence conditioned on the entire input sequence processed by the encoder. The asynchronous problem between the encoding and decoding makes these models difficult to be applied for online speech recognition. In this paper, we propose a model named synchronous transformer to address this problem, which can predict the output sequence chunk by chunk. Once a fixed-length chunk of the input sequence is processed by the encoder, the decoder begins to predict symbols immediately. During training, a forward-backward algorithm is introduced to optimize all the possible alignment paths. Our model is evaluated on a Mandarin dataset AISHELL-1. The experiments show that the synchronous transformer is able to perform encoding and decoding synchronously, and achieves a character error rate of 8.91% on the test set.

Ye Bai, Jiangyan Yi, Jianhua Tao et al.

Attention-based encoder-decoder (AED) models have achieved promising performance in speech recognition. However, because of the end-to-end training, an AED model is usually trained with speech-text paired data. It is challenging to incorporate external text-only data into AED models. Another issue of the AED model is that it does not use the right context of a text token while predicting the token. To alleviate the above two issues, we propose a unified method called LST (Learn Spelling from Teachers) to integrate knowledge into an AED model from the external text-only data and leverage the whole context in a sentence. The method is divided into two stages. First, in the representation stage, a language model is trained on the text. It can be seen as that the knowledge in the text is compressed into the LM. Then, at the transferring stage, the knowledge is transferred to the AED model via teacher-student learning. To further use the whole context of the text sentence, we propose an LM called causal cloze completer (COR), which estimates the probability of a token, given both the left context and the right context of it. Therefore, with LST training, the AED model can leverage the whole context in the sentence. Different from fusion based methods, which use LM during decoding, the proposed method does not increase any extra complexity at the inference stage. We conduct experiments on two scales of public Chinese datasets AISHELL-1 and AISHELL-2. The experimental results demonstrate the effectiveness of leveraging external text-only data and the whole context in a sentence with our proposed method, compared with baseline hybrid systems and AED model based systems.

Zhengkun Tian, Jiangyan Yi, Jianhua Tao et al.

Recurrent neural network transducers (RNN-T) have been successfully applied in end-to-end speech recognition. However, the recurrent structure makes it difficult for parallelization . In this paper, we propose a self-attention transducer (SA-T) for speech recognition. RNNs are replaced with self-attention blocks, which are powerful to model long-term dependencies inside sequences and able to be efficiently parallelized. Furthermore, a path-aware regularization is proposed to assist SA-T to learn alignments and improve the performance. Additionally, a chunk-flow mechanism is utilized to achieve online decoding. All experiments are conducted on a Mandarin Chinese dataset AISHELL-1. The results demonstrate that our proposed approach achieves a 21.3% relative reduction in character error rate compared with the baseline RNN-T. In addition, the SA-T with chunk-flow mechanism can perform online decoding with only a little degradation of the performance.

Cunhang Fan, Bin Liu, Jianhua Tao et al.

Deep clustering (DC) and utterance-level permutation invariant training (uPIT) have been demonstrated promising for speaker-independent speech separation. DC is usually formulated as two-step processes: embedding learning and embedding clustering, which results in complex separation pipelines and a huge obstacle in directly optimizing the actual separation objectives. As for uPIT, it only minimizes the chosen permutation with the lowest mean square error, doesn't discriminate it with other permutations. In this paper, we propose a discriminative learning method for speaker-independent speech separation using deep embedding features. Firstly, a DC network is trained to extract deep embedding features, which contain each source's information and have an advantage in discriminating each target speakers. Then these features are used as the input for uPIT to directly separate the different sources. Finally, uPIT and DC are jointly trained, which directly optimizes the actual separation objectives. Moreover, in order to maximize the distance of each permutation, the discriminative learning is applied to fine tuning the whole model. Our experiments are conducted on WSJ0-2mix dataset. Experimental results show that the proposed models achieve better performances than DC and uPIT for speaker-independent speech separation.

Ye Bai, Jiangyan Yi, Jianhua Tao et al.

Integrating an external language model into a sequence-to-sequence speech recognition system is non-trivial. Previous works utilize linear interpolation or a fusion network to integrate external language models. However, these approaches introduce external components, and increase decoding computation. In this paper, we instead propose a knowledge distillation based training approach to integrating external language models into a sequence-to-sequence model. A recurrent neural network language model, which is trained on large scale external text, generates soft labels to guide the sequence-to-sequence model training. Thus, the language model plays the role of the teacher. This approach does not add any external component to the sequence-to-sequence model during testing. And this approach is flexible to be combined with shallow fusion technique together for decoding. The experiments are conducted on public Chinese datasets AISHELL-1 and CLMAD. Our approach achieves a character error rate of 9.3%, which is relatively reduced by 18.42% compared with the vanilla sequence-to-sequence model.

Jiangyan Yi, Jianhua Tao, Zhengqi Wen et al.

In order to improve the performance for far-field speech recognition, this paper proposes to distill knowledge from the close-talking model to the far-field model using parallel data. The close-talking model is called the teacher model. The far-field model is called the student model. The student model is trained to imitate the output distributions of the teacher model. This constraint can be realized by minimizing the Kullback-Leibler (KL) divergence between the output distribution of the student model and the teacher model. Experimental results on AMI corpus show that the best student model achieves up to 4.7% absolute word error rate (WER) reduction when compared with the conventionally-trained baseline models.