Compared with automatic speech recognition (ASR), the human auditory system is more adept at handling noise-adverse situations, including environmental noise and channel distortion. To mimic this adeptness, auditory models have been widely incorporated in ASR systems to improve their robustness. This paper proposes a novel auditory model which incorporates psychoacoustics and otoacoustic emissions (OAEs) into ASR. In particular, we successfully implement the frequency-dependent property of psychoacoustic models and effectively improve resulting system performance. We also present a novel double-transform spectrum-analysis technique, which can qualitatively predict ASR performance for different noise types. Detailed theoretical analysis is provided to show the effectiveness of the proposed algorithm. Experiments are carried out on the AURORA2 database and show that the word recognition rate using our proposed feature extraction method is significantly increased over the baseline. Given models trained with clean speech, our proposed method achieves up to 85.39% word recognition accuracy on noisy data.
The multi-stream paradigm of audio processing, in which several sources are simultaneously considered, has been an active research area for information fusion. Our previous study offered a promising direction within end-to-end automatic speech recognition, where parallel encoders aim to capture diverse information followed by a stream-level fusion based on attention mechanisms to combine the different views. However, with an increasing number of streams resulting in an increasing number of encoders, the previous approach could require substantial memory and massive amounts of parallel data for joint training. In this work, we propose a practical two-stage training scheme. Stage-1 is to train a Universal Feature Extractor (UFE), where encoder outputs are produced from a single-stream model trained with all data. Stage-2 formulates a multi-stream scheme intending to solely train the attention fusion module using the UFE features and pretrained components from Stage-1. Experiments have been conducted on two datasets, DIRHA and AMI, as a multi-stream scenario. Compared with our previous method, this strategy achieves relative word error rate reductions of 8.2--32.4%, while consistently outperforming several conventional combination methods.
We develop an algorithm which can learn from partially labeled and unsegmented sequential data. Most sequential loss functions, such as Connectionist Temporal Classification (CTC), break down when many labels are missing. We address this problem with Star Temporal Classification (STC) which uses a special star token to allow alignments which include all possible tokens whenever a token could be missing. We express STC as the composition of weighted finite-state transducers (WFSTs) and use GTN (a framework for automatic differentiation with WFSTs) to compute gradients. We perform extensive experiments on automatic speech recognition. These experiments show that STC can recover most of the performance of supervised baseline when up to 70% of the labels are missing. We also perform experiments in handwriting recognition to show that our method easily applies to other sequence classification tasks.
End-to-end (E2E) systems are fast replacing the conventional systems in the domain of automatic speech recognition. As the target labels are learned directly from speech data, the E2E systems need a bigger corpus for effective training. In the context of code-switching task, the E2E systems face two challenges: (i) the expansion of the target set due to multiple languages involved, and (ii) the lack of availability of sufficiently large domain-specific corpus. Towards addressing those challenges, we propose an approach for reducing the number of target labels for reliable training of the E2E systems on limited data. The efficacy of the proposed approach has been demonstrated on two prominent architectures, namely CTC-based and attention-based E2E networks. The experimental validations are performed on a recently created Hindi-English code-switching corpus. For contrast purpose, the results for the full target set based E2E system and a hybrid DNN-HMM system are also reported.
Deep neural networks (DNNs) have been successfully applied to a wide variety of acoustic modeling tasks in recent years. These include the applications of DNNs either in a discriminative feature extraction or in a hybrid acoustic modeling scenario. Despite the rapid progress in this area, a number of challenges remain in training DNNs. This paper presents an effective way of training DNNs using a manifold learning based regularization framework. In this framework, the parameters of the network are optimized to preserve underlying manifold based relationships between speech feature vectors while minimizing a measure of loss between network outputs and targets. This is achieved by incorporating manifold based locality constraints in the objective criterion of DNNs. Empirical evidence is provided to demonstrate that training a network with manifold constraints preserves structural compactness in the hidden layers of the network. Manifold regularization is applied to train bottleneck DNNs for feature extraction in hidden Markov model (HMM) based speech recognition. The experiments in this work are conducted on the Aurora-2 spoken digits and the Aurora-4 read news large vocabulary continuous speech recognition tasks. The performance is measured in terms of word error rate (WER) on these tasks. It is shown that the manifold regularized DNNs result in up to 37% reduction in WER relative to standard DNNs.
This article provides a unifying Bayesian network view on various approaches for acoustic model adaptation, missing feature, and uncertainty decoding that are well-known in the literature of robust automatic speech recognition. The representatives of these classes can often be deduced from a Bayesian network that extends the conventional hidden Markov models used in speech recognition. These extensions, in turn, can in many cases be motivated from an underlying observation model that relates clean and distorted feature vectors. By converting the observation models into a Bayesian network representation, we formulate the corresponding compensation rules leading to a unified view on known derivations as well as to new formulations for certain approaches. The generic Bayesian perspective provided in this contribution thus highlights structural differences and similarities between the analyzed approaches.
The goal of this work is to build flexible video-language models that can generalize to various video-to-text tasks from few examples, such as domain-specific captioning, question answering, and future event prediction. Existing few-shot video-language learners focus exclusively on the encoder, resulting in the absence of a video-to-text decoder to handle generative tasks. Video captioners have been pretrained on large-scale video-language datasets, but they rely heavily on finetuning and lack the ability to generate text for unseen tasks in a few-shot setting. We propose VidIL, a few-shot Video-language Learner via Image and Language models, which demonstrates strong performance on few-shot video-to-text tasks without the necessity of pretraining or finetuning on any video datasets. We use the image-language models to translate the video content into frame captions, object, attribute, and event phrases, and compose them into a temporal structure template. We then instruct a language model, with a prompt containing a few in-context examples, to generate a target output from the composed content. The flexibility of prompting allows the model to capture any form of text input, such as automatic speech recognition (ASR) transcripts. Our experiments demonstrate the power of language models in understanding videos on a wide variety of video-language tasks, including video captioning, video question answering, video caption retrieval, and video future event prediction. Especially, on video future event prediction, our few-shot model significantly outperforms state-of-the-art supervised models trained on large-scale video datasets. Code and resources are publicly available for research purposes at https://github.com/MikeWangWZHL/VidIL .
State-of-the-art encoder-decoder models (e.g. for machine translation (MT) or speech recognition (ASR)) are constructed and trained end-to-end as an atomic unit. No component of the model can be (re-)used without the others. We describe LegoNN, a procedure for building encoder-decoder architectures with decoder modules that can be reused across various MT and ASR tasks, without the need for any fine-tuning. To achieve reusability, the interface between each encoder and decoder modules is grounded to a sequence of marginal distributions over a discrete vocabulary pre-defined by the model designer. We present two approaches for ingesting these marginals; one is differentiable, allowing the flow of gradients across the entire network, and the other is gradient-isolating. To enable portability of decoder modules between MT tasks for different source languages and across other tasks like ASR, we introduce a modality agnostic encoder which consists of a length control mechanism to dynamically adapt encoders' output lengths in order to match the expected input length range of pre-trained decoders. We present several experiments to demonstrate the effectiveness of LegoNN models: a trained language generation LegoNN decoder module from German-English (De-En) MT task can be reused with no fine-tuning for the Europarl English ASR and the Romanian-English (Ro-En) MT tasks to match or beat respective baseline models. When fine-tuned towards the target task for few thousand updates, our LegoNN models improved the Ro-En MT task by 1.5 BLEU points, and achieved 12.5% relative WER reduction for the Europarl ASR task. Furthermore, to show its extensibility, we compose a LegoNN ASR model from three modules -- each has been learned within different end-to-end trained models on three different datasets -- boosting the WER reduction to 19.5%.
In this paper we study the probabilistic properties of the posteriors in a speech recognition system that uses a deep neural network (DNN) for acoustic modeling. We do this by reducing Kaldi's DNN shared pdf-id posteriors to phone likelihoods, and using test set forced alignments to evaluate these using a calibration sensitive metric. Individual frame posteriors are in principle well-calibrated, because the DNN is trained using cross entropy as the objective function, which is a proper scoring rule. When entire phones are assessed, we observe that it is best to average the log likelihoods over the duration of the phone. Further scaling of the average log likelihoods by the logarithm of the duration slightly improves the calibration, and this improvement is retained when tested on independent test data.
It is an effective way that improves the performance of the existing Automatic Speech Recognition (ASR) systems by retraining with more and more new training data in the target domain. Recently, Deep Neural Network (DNN) has become a successful model in the ASR field. In the training process of the DNN based methods, a back propagation of error between the transcription and the corresponding annotated text is used to update and optimize the parameters. Thus, the parameters are more influenced by the training samples with a big propagation error than the samples with a small one. In this paper, we define the samples with significant error as the hard samples and try to improve the performance of the ASR system by adding many of them. Unfortunately, the hard samples are sparse in the training data of the target domain, and manually label them is expensive. Therefore, we propose a hard samples mining method based on an enhanced deep multiple instance learning, which can find the hard samples from unlabeled training data by using a small subset of the dataset with manual labeling in the target domain. We applied our method to an End2End ASR task and obtained the best performance.