In this paper, alternative dynamic features for speech recognition are proposed. The goal of this work is to improve speech recognition accuracy by deriving the representation of distinctive dynamic characteristics from a speech spectrum. This work was inspired by two temporal dynamics of a speech signal. One is the highly non-stationary nature of speech, and the other is the inter-frame change of a speech spectrum. We adopt the use of a sub-frame spectrum analyzer to capture very rapid spectral changes within a speech analysis frame. In addition, we attempt to measure spectral fluctuations of a more complex manner as opposed to traditional dynamic features such as delta or double-delta. To evaluate the proposed features, speech recognition tests over smartphone environments were conducted. The experimental results show that the feature streams simply combined with the proposed features are effective for an improvement in the recognition accuracy of a hidden Markov model–based speech recognizer.
In this paper, we propose a rank‐weighted reconstruction feature to improve the robustness of a feed‐forward deep neural network (FFDNN)‐based acoustic model. In the FFDNN‐based acoustic model, an input feature is constructed by vectorizing a submatrix that is created by slicing the feature vectors of frames within a context window. In this type of feature construction, the appropriate context window size is important because it determines the amount of trivial or discriminative information, such as redundancy, or temporal context of the input features. However, we ascertained whether a single parameter is sufficiently able to control the quantity of information. Therefore, we investigated the input feature construction from the perspectives of rank and nullity, and proposed a rank‐weighted reconstruction feature herein, that allows for the retention of speech information components and the reduction in trivial components. The proposed method was evaluated in the TIMIT phone recognition and Wall Street Journal (WSJ) domains. The proposed method reduced the phone error rate of the TIMIT domain from 18.4% to 18.0%, and the word error rate of the WSJ domain from 4.70% to 4.43%.
In this paper, we propose a policy gradient-based semi-supervised speech recognition acoustic model training. In practice, self-training and teacher/student learning are one of the widely used semi-supervised training methods due to their scalability and effectiveness. These methods are based on generating pseudo labels for unlabeled samples using a pre-trained model and selecting reliable samples using confidence measure. However, there are some considerations in this approach. The generated pseudo labels can be biased depending on which pre-trained model is used, and the training process can be complicated because the confidence measure is usually carried out in post-processing using external knowledge. Therefore, to address these issues, we propose a policy gradient method-based approach. Policy gradient is a reinforcement learning algorithm to find an optimal behavior strategy for an agent to obtain optimal rewards. The policy gradient-based approach provides a framework for exploring unlabeled data as well as exploiting labeled data, and it also provides a way to incorporate external knowledge in the same training cycle. The proposed approach was evaluated on an in-house non-native Korean recognition domain. The experimental results show that the method is effective in semi-supervised acoustic model training.
In this paper, we propose the use of data-driven probabilistic utterance-level decision logic to improve Weighted Finite State Transducer (WFST)-based endpoint detection. In general, endpoint detection is dealt with using two cascaded decision processes. The first process is frame-level speech/non-speech classification based on statistical hypothesis testing, and the second process is a heuristic-knowledge-based utterance-level speech boundary decision. To handle these two processes within a unified framework, we propose a WFST-based approach. However, a WFST-based approach has the same limitations as conventional approaches in that the utterance-level decision is based on heuristic knowledge and the decision parameters are tuned sequentially. Therefore, to obtain decision knowledge from a speech corpus and optimize the parameters at the same time, we propose the use of data-driven probabilistic utterance-level decision logic. The proposed method reduces the average detection failure rate by about 14% for various noisy-speech corpora collected for an endpoint detection evaluation.
Unsupervised learning-based approaches for training speech vector representations (SVR) have recently been widely applied. While pretrained SVR models excel in relatively clean automatic speech recognition (ASR) tasks, such as those recorded in laboratory environments, they are still insufficient for practical applications with various types of noise, intonation, and dialects. To cope with this problem, we present a novel unsupervised SVR learning method for practical end-to-end ASR models. Our approach involves designing a speech feature masking method to stabilize SVR model learning and improve the performance of the ASR model in a downstream task. By introducing a noise masking strategy into diverse combinations of the time and frequency regions of the spectrogram, the SVR model becomes a robust representation extractor for the ASR model in practical scenarios. In pretraining experiments, we train the SVR model using approximately 18,000 h of Korean speech datasets that included diverse speakers and were recorded in environments with various amounts of noise. The weights of the pretrained SVR extractor are then frozen, and the extracted speech representations are used for ASR model training in a downstream task. The experimental results show that the ASR model using our proposed SVR extractor significantly outperforms conventional methods.
This paper proposes a new Utterance Verification (UV) algorithm based on i-vector. Phone segments are extracted and concatenated from the training data, which are used to train the Universal Background Model (UBM) and the Total Variability (TV) matrix, and then, i-vector is extracted from the enrollment and evaluation data using UBM and TV matrix. We compare two Confidence Measures (CMs), cosine distance scoring and Support Vector Machine (SVM). To compensate the channel effect, we use two channel compensation methods, Linear Discriminant Analysis (LDA) and Within-Class Covariance Normalization (WCCN). The decision is made by the word-level CM by combining the phone-level CMs. Experiments are conducted in the Korean isolated word recognition domain. Experimental results show that SVM is superior to cosine distance scoring. Best performance is achieved when SVM is used without any channel compensation method.