https://www2.econ.iastate.edu/tesfatsi/DeepLearningInNeuralNetworksOverview.JSchmidhuber2015.pdf

Deep learning in neural networks

We describe an open-source toolkit for statistical machine translation whose novel contributions are (a) support for linguistically motivated factors, (b) confusion network decoding, and (c) efficient data formats for translation models and language models. In addition to the SMT decoder, the toolkit also includes a wide variety of tools for training, tuning and applying the system to many translation tasks.

/pdf/moses-open-source-toolkit-for-statistical-machine-3pmfuwk8uc.pdf

Moses: Open Source Toolkit for Statistical Machine Translation

Several variants of the long short-term memory (LSTM) architecture for recurrent neural networks have been proposed since its inception in 1995. In recent years, these networks have become the state-of-the-art models for a variety of machine learning problems. This has led to a renewed interest in understanding the role and utility of various computational components of typical LSTM variants. In this paper, we present the first large-scale analysis of eight LSTM variants on three representative tasks: speech recognition, handwriting recognition, and polyphonic music modeling. The hyperparameters of all LSTM variants for each task were optimized separately using random search, and their importance was assessed using the powerful functional ANalysis Of VAriance framework. In total, we summarize the results of 5400 experimental runs ( $\approx 15$ years of CPU time), which makes our study the largest of its kind on LSTM networks. Our results show that none of the variants can improve upon the standard LSTM architecture significantly, and demonstrate the forget gate and the output activation function to be its most critical components. We further observe that the studied hyperparameters are virtually independent and derive guidelines for their efficient adjustment.

LSTM: A Search Space Odyssey

This paper introduces WaveNet, a deep neural network for generating raw audio waveforms. The model is fully probabilistic and autoregressive, with the predictive distribution for each audio sample conditioned on all previous ones; nonetheless we show that it can be efficiently trained on data with tens of thousands of samples per second of audio. When applied to text-to-speech, it yields state-of-the-art performance, with human listeners rating it as significantly more natural sounding than the best parametric and concatenative systems for both English and Mandarin. A single WaveNet can capture the characteristics of many different speakers with equal fidelity, and can switch between them by conditioning on the speaker identity. When trained to model music, we find that it generates novel and often highly realistic musical fragments. We also show that it can be employed as a discriminative model, returning promising results for phoneme recognition.

WaveNet: A Generative Model for Raw Audio

This paper introduces WaveNet, a deep neural network for generating raw audio
waveforms. The model is fully probabilistic and autoregressive, with the
predictive distribution for each audio sample conditioned on all previous ones;
nonetheless we show that it can be efficiently trained on data with tens of
thousands of samples per second of audio. When applied to text-to-speech, it
yields state-of-the-art performance, with human listeners rating it as
significantly more natural sounding than the best parametric and concatenative
systems for both English and Mandarin. A single WaveNet can capture the
characteristics of many different speakers with equal fidelity, and can switch
between them by conditioning on the speaker identity. When trained to model
music, we find that it generates novel and often highly realistic musical
fragments. We also show that it can be employed as a discriminative model,
returning promising results for phoneme recognition.

Feed-forward, Deep neural networks (DNN)-based text-tospeech (TTS) systems have been recently shown to outperform decision-tree clustered context-dependent HMM TTS systems [1, 4]. However, the long time span contextual effect in a speech utterance is still not easy to accommodate, due to the intrinsic, feed-forward nature in DNN-based modeling. Also, to synthesize a smooth speech trajectory, the dynamic features are commonly used to constrain speech parameter trajectory generation in HMM-based TTS [2]. In this paper, Recurrent Neural Networks (RNNs) with Bidirectional Long Short Term Memory (BLSTM) cells are adopted to capture the correlation or co-occurrence information between any two instants in a speech utterance for parametric TTS synthesis. Experimental results show that a hybrid system of DNN and BLSTM-RNN, i.e., lower hidden layers with a feed-forward structure which is cascaded with upper hidden layers with a bidirectional RNN structure of LSTM, can outperform either the conventional, decision tree-based HMM, or a DNN TTS system, both objectively and subjectively. The speech trajectory generated by the BLSTM-RNN TTS is fairly smooth and no dynamic constraints are needed.

TTS Synthesis with Bidirectional LSTM based Recurrent Neural Networks

Research in the field of automatic speech and speaker recognition has made a number of significant advances in the last two decades, influenced by advances in signal processing, algorithms, architectures, and hardware. These advances include: the adoption of a statistical pattern recognition paradigm; the use of the hidden Markov modeling framework to characterize both the spectral and the temporal variations in the speech signal; the use of a large set of speech utterance examples from a large population of speakers to train the hidden Markov models of some fundamental speech units; the organization of speech and language knowledge sources into a structural finite state network; and the use of dynamic, programming based heuristic search methods to find the best word sequence in the lexical network corresponding to the spoken utterance. Automatic Speech and Speaker Recognition: Advanced Topics groups together in a single volume a number of important topics on speech and speaker recognition, topics which are of fundamental importance, but not yet covered in detail in existing textbooks. Although no explicit partition is given, the book is divided into five parts: Chapters 1-2 are devoted to technology overviews; Chapters 3-12 discuss acoustic modeling of fundamental speech units and lexical modeling of words and pronunciations; Chapters 13-15 address the issues related to flexibility and robustness; Chapter 16-18 concern the theoretical and practical issues of search; Chapters 19-20 give two examples of algorithm and implementational aspects for recognition system realization. Audience: A reference book for speech researchers and graduate students interested in pursuing potential research on the topic. May also be used as a text for advanced courses on the subject.

Automatic Speech and Speaker Recognition: Advanced Topics

Deep Neural Network (DNN), which can model a long-span, intricate transform compactly with a deep-layered structure, has recently been investigated for parametric TTS synthesis with a fairly large corpus (33,000 utterances) [6]. In this paper, we examine DNN TTS synthesis with a moderate size corpus of 5 hours, which is more commonly used for parametric TTS training. DNN is used to map input text features into output acoustic features (LSP, F0 and V/U). Experimental results show that DNN can outperform the conventional HMM, which is trained in ML first and then refined by MGE. Both objective and subjective measures indicate that DNN can synthesize speech better than HMM-based baseline. The improvement is mainly on the prosody, i.e., the RMSE of natural and generated F0 trajectories by DNN is improved by 2 Hz. This benefit is likely from the key characteristics of DNN, which can exploit feature correlations, e.g., between F0 and spectrum, without using a more restricted, e.g. diagonal Gaussian probability family. Our experimental results also show: the layer-wise BP pre-training can drive weights to a better starting point than random initialization and result in a more effective DNN; state boundary info is important for training DNN to yield better synthesized speech; and a hyperbolic tangent activation function in DNN hidden layers yields faster convergence than a sigmoidal one.

On the training aspects of Deep Neural Network (DNN) for parametric TTS synthesis

Described is a voice persona service by which users convert text into speech waveforms, based on user-provided parameters and voice data from a service data store. The service may be remotely accessed, such as via the Internet. The user may provide text tagged with parameters, with the text sent to a text-to-speech engine along with base or custom voice data, and the resulting waveform morphed based on the tags. The user may also provide speech. Once created, a voice persona corresponding to the speech waveform may be persisted, exchanged, made public, shared and so forth. In one example, the voice persona service receives user input and parameters, and retrieves a base or custom voice that may be edited by the user via a morphing algorithm. The service outputs a waveform, such as a .wav file for embedding in a software program, and persists the voice persona corresponding to that waveform.

Frank K. Soong

Papers

TTS Synthesis with Bidirectional LSTM based Recurrent Neural Networks

Automatic Speech and Speaker Recognition: Advanced Topics

On the training aspects of Deep Neural Network (DNN) for parametric TTS synthesis

Voice persona service for embedding text-to-speech features into software programs

Improved mispronunciation detection with deep neural network trained acoustic models and transfer learning based logistic regression classifiers