Speech_Recognition_Using_Deep_Neural_Networks_A_Systematic_Review

They are not effective for modeling non-linear or near non-linear functions.

They allow parameter sharing through different layers of the network.

To extract specific features and information from inputs.

Deep learning arose as a new area of machine learning.

Speaker identification and speaker verification.

Learning by interacting with the problem environment, where an agent learns from its own actions.

Regression algorithms and classification algorithms.

To produce a classifier function for discrete outputs or a regression function for continuous outputs.

Applying inclusion/exclusion criteria to ensure only relevant papers are included.

Speech spectrogram features are more advanced than MFCC with deep neural networks compared to traditional GMMs-HMMs.

It requires less human intervention and utilizes cheaper, easier-to-access unlabeled datasets.

To which registered speaker a given utterance corresponds.

Gaussian Mixture Models (GMMs) based on Hidden Markov Models (HMMs).

A type of machine learning that uses labeled data to train the algorithm.

Greedy layerwise unsupervised pre-training.

They allow for discriminative training more efficiently than HMMs.

They are considered hard to train to capture long-term dependencies.

Dictating computers instead of typing.

Iteratively from analyzed data and new input data.

Observations, examples, instructions, and direct experience.

They require fewer parameters to represent non-linear functions.

Reinforcement learning uses direct interactions with the environment to gain knowledge, while supervised learning learns from examples provided by an external supervisor.

Speech recognition.

Supervised learning, unsupervised learning, semi-supervised learning, reinforcement learning, and deep learning.

A field for human-machine interaction that recognizes user emotions from speech.

Deep belief networks (DBN).

To uncover the best function that fits points in the training dataset.

To find common points between inputs in the dataset, often through clustering.

To conduct a comparison with the current review.

Differentiating between closely correlated languages.

A type of learning that uses trial and error to maximize a cumulative reward metric.

Papers that use deep neural networks in areas other than speech, papers related to speech but not using deep neural networks, and papers with no clear publication information.

To learn more about the data by identifying the fundamental structure or distribution patterns within it.

Utilizing deep learning for speech-related applications.

174 papers published between 2006 and 2018.

To admit or discard the claimed speaker identity.

Speech recognition.

Predicting future data sequences using previous data samples.

Feature extraction, language modeling, acoustic models, understanding speech, and dialogue estimation.

Unsupervised (generative) learning, supervised learning, and hybrid deep networks.

Linear regression, multiple linear regression, and polynomial regression.

A combination of supervised and unsupervised learning using both labeled and unlabeled data.

Papers that use deep neural networks or deep learning in the area of speech.

Unsupervised learning uses an input dataset without any labeled outputs, while supervised learning uses labeled outputs.

Types of speech identified, databases used, languages, environment types, features extracted, publication types, and distribution of papers over the years.

Terms related to deep neural networks and speech.

By incorporating speech properties into the architecture.

Ten QARs.

Google Scholar, IEEE Explorer, Science Direct, ResearchGate, and Springer.

To extract needed information to answer the set of research questions.

Clustering, dimensionality reduction, and anomaly detection.

A field of study that provides computers with the ability to learn from input data without being explicitly programmed.

A method that falls between supervised and unsupervised learning, using a large amount of unlabeled data and a small amount of labeled data.

The persistent occurrence of local optima in the non-convex objective function.

The use of deep neural networks in speech recognition.

Back-propagation (BP).

Recent advances in deep learning capabilities and limitations in speech recognition.

A sub-field of machine learning based on algorithms that learn from multiple levels to represent complex relations among data.

Emotion identification and emotion verification.

Learning the transformation of previously learned features at each new layer.

They struggle with continuous speech signals due to inability to model temporal dependencies.

Basics of state-of-the-art solutions from computational and phonological perspectives.

Sparse interactions, parameter sharing, and equivariant representation.

230 papers.

Estimating the speaker’s age using their speech signals.

Increased processing abilities of computer chips, incorporation of large training datasets, and advances in machine learning.

Using the patient's voice to provide information on their health status.

To perform discriminative deep architecture tasks, particularly in computer vision and image recognition.

To uncover the best fit class for the input data by assigning each input to its correct class.

Kitchenham and Charters methodology.

Identifying the research questions.

To sub-sample the output from the convolutional layer and decrease the data rate.

Is the paper well organized?

Discriminatively trained feed-forward networks and their effectiveness prior to HMM decoding.

Scores range from 1 for fully answered to 0 for completely not answered.

Deep learning architectures have multiple layers of non-linear feature transformation, while shallow architectures typically have one or two layers.

Hessian free optimization.

The recognition of a speaker’s regional accent within a predetermined language.

Because high computational power became more accessible.

174 papers.

Automatic speech recognition, emotional speech recognition, speaker identification, and speech enhancement.

The real challenge is in the extraction process itself.

Social information filtering algorithms, like those used by Amazon.com for recommendations.

Reduced word error rate (WER) by 30% compared to GMM-based models.

Acoustic environment distortion knowledge, model domain vs. feature domain processing, specific environment distortion models, uncertainty processing, and acoustic models trained by the same adaptation process.

A type of machine learning that models abstractions in data using a graph with multiple processing layers.

The paper was excluded from the review.

By grouping inputs based on the features extracted from each input object.

No, they do not assign names but can differentiate among clusters.

They outperform GMM-HMM models on various speech recognition benchmarks.

Speech, speaker, emotion, health, language, accent, age, and gender recognition.

174 publications.

230 papers.

Applying quality assessment rules to identify the final list of papers.

The process of recognizing whether the speaker is male or female.

The capability of a machine or computer to recognize the content of words and phrases in an uttered language.

Research patterns, gaps, and future directions in the use of deep neural networks in speech recognition.