Speech Files Research Articles

Spoken Language Identification System Using Convolutional Recurrent Neural Network

Following recent advancements in deep learning and artificial intelligence, spoken language identification applications are playing an increasingly significant role in our day-to-day lives, especially in the domain of multi-lingual speech recognition. In this article, we propose a spoken language identification system that depends on the sequence of feature vectors. The proposed system uses a hybrid Convolutional Recurrent Neural Network (CRNN), which combines a Convolutional Neural Network (CNN) with a Recurrent Neural Network (RNN) network, for spoken language identification on seven languages, including Arabic, chosen from subsets of the Mozilla Common Voice (MCV) corpus. The proposed system exploits the advantages of both CNN and RNN architectures to construct the CRNN architecture. At the feature extraction stage, it compares the Gammatone Cepstral Coefficient (GTCC) feature and Mel Frequency Cepstral Coefficient (MFCC) feature, as well as a combination of both. Finally, the speech signals were represented as frames and used as the input for the CRNN architecture. After conducting experiments, the results of the proposed system indicate higher performance with combined GTCC and MFCC features compared to GTCC or MFCC features used individually. The average accuracy of the proposed system was 92.81% in the best experiment for spoken language identification. Furthermore, the system can learn language-specific patterns in various filter size representations of speech files.

법과학적 활용을 위한 삼성 스마트폰 음성 녹음 파일의 메타데이터 구조 및 속성 비교 분석 연구*

Due to the popularization of smartphones, most of the recorded speech files submitted as evidence of recent crimes are produced by smartphones, and the integrity (forgery) of the submitted speech files based on smartphones is emerging as a major issue in the investigation and trial process. Samsung smartphones with the highest domestic market share are distributed with built-in speech recording applications that can record calls and voice, and can edit recorded speech. Unlike editing through third-party speech (audio) applications, editing by their own builtin speech applications has a high similarity to the original file in metadata structures and attributes, so more precise analysis techniques need to prove integrity. In this study, we constructed a speech file metadata database for speech files (original files) recorded by 34 Samsung smartphones and edited speech files edited by their built-in speech recording applications. We analyzed by comparing the metadata structures and attributes of the original files to their edited ones. As a result, we found significant metadata differences between the original speech files and the edited ones.

Lombard effect in Algerian Dialect Speech: Investigation of boosting and bypass strategies in narrowband noisy environment

The paper presents the first investigation on the Lombard effect (LE) in the Algerian Dialect (AD) speech. It is devoted to the examination of boosting and bypass strategies in narrowband noisy environments (0.49–6.8 kHz). Speakers who adopt the boosting tactic would attempt to improve their audibility in noise by increasing the global intensity (I), the local intensity, and the duration (D) of their speech. The boost of the local intensity consists in amplifying the spectral magnitude of the speech waveform in the frequency region wherein the masking noise is beyond 55 dB. However, speakers who apply the bypass strategy would raise their speech in noise so that their fundamental frequency (F0) and first Formant (F1) will decrease. We built the AD Lombard Speech Corpus (ADLSC) that consists of continuous speech files uttered by eight native speakers in quiet and noisy conditions. The corpus contains phonetically rich words that cover all AD phonemes. To examine these strategies, we assessed the change in I, D, F0 and F1 from quiet to noisy environments. Subsequently, we evaluated the change in the local intensity via the computation of the Power Spectrum Ratio (PSR). Statistical tests were conducted to confirm the significance of the acoustic analysis. We found out that speakers raised their vocal effort in I and D. Furthermore, the measured PSRs confirmed the evidence that LE is frequency-specific in humans. Similarly, we noticed that speakers increased their F0 and F1 in noisy conditions. The results we obtained confirmed the boosting strategy but denied the bypass tactic. We compared our findings with the results reported in the state-of-the-art studies. The comparison revealed that LE in narrowband does not exhibit the same acoustic variations in AD and English.

Hiding Capacity and Audio Steganography Model Based on LSB in Temporal Domain

Background: Researchers for data hiding using the various encrypting techniques to improve the security of transmission of confidential information through an unsecured channel have been carried out. Methods: In this paper, a new higher recognition Least Significant Bit (LSB) audio data concealment is suggested in this framework. This technique is used to embed the hidden audio into cover audio of the same size. Results: The altered speech file appears as the original carrier file after embedding the secret message. This model strengthens hiding capability and audio quality. Conclusion: The strategy outperforms similar studies by enhancing hiding capability up to 30% and preserving stego audio transparency with the SNR value at 72.2 and SDG at 4.8.

Intelligence System for Multi-Language Recognition

Language classification systems are used to classify spoken language from a particular phoneme sample and are usually the first step of many spoken language processing tasks, such as automatic speech recognition (ASR) systems Without automatic language detection, spoken speech cannot be properly analyzed and grammar rules cannot be applied, causing failures Subsequent speech recognition steps. We propose a language classification system that solves the problem in the image field, rather than the sound field. This research identified and implemented several low-level features using Mel Frequency Cepstral Coefficients, which extract traits from speech files of four languages (Arabic, English, French, Kurdish) from the database (M2L_Dataset) as the data source used in this research. A Convolutional Neuron Network is used to operate on spectrogram images of the available audio snippets. In extensive experiments, we showed that our model is applicable to a range of noisy scenarios and can easily be extended to previously unknown languages, while maintaining classification accuracy. We released our own code and extensive training package for language classification systems for the community. CNN algorithm was applied in this research to classify and the result was perfect, as the classification accuracy reached 97% between two languages if the sample length was only one second, but if the sample length was two seconds, the classification accuracy reached 98%. While the classification among three languages, the classification accuracy reached 95% if the sample length was only one second, but if the sample length was two seconds, the classification accuracy reached 96%.

Digital Speech Files Encryption based on Hénon and Gingerbread Chaotic Maps

Speech encryption approaches are used to prevent eavesdropping, tracking, and other security concerns in speech communication. In this paper, a new cryptography algorithm is proposed to encrypt digital speech files. Initially, the digital speech files are rearranged as a cubic model with six sides to scatter speech data. Furthermore, each side is encrypted by random keys that are created by using two chaotic maps (Hénon and Gingerbread chaotic maps). Encryption for each side of the cube is achieved, using the based map vector that is generated randomly by using a simple random function. Map vector that consists of six bits, each bit refers to one of the specific chaotic maps that generate a random key to encrypt each face of the cube. Results show that the pseudo-random keys created by using chaotic maps for cryptographic speech file have an acceptable characteristic concerning randomness tests, which is confirmed in this paper by using five statistical tests. The final evaluation of the speech encryption algorithm is measured by using different quality metrics, and the results show that the algorithm can achieve resist encryption.

Detecting The Speaker Language Using CNN Deep Learning Algorithm

Many languages identification (LID) systems rely on language models that use machine learning (ML) approaches, LID systems utilize rather long recording periods to achieve satisfactory accuracy. This study aims to extract enough information from short recording intervals in order to successfully classify the spoken languages under test. The classification process is based on frames of (2-18) seconds where most of the previous LID systems were based on much longer time frames (from 3 seconds to 2 minutes). This research defined and implemented many low-level features using MFCC (Mel-frequency cepstral coefficients), containing speech files in five languages (English. French, German, Italian, Spanish), from voxforge.org an open-source corpus that consists of user-submitted audio clips in various languages, is the source of data used in this paper. A CNN (convolutional Neural Networks) algorithm applied in this paper for classification and the result was perfect, binary language classification had an accuracy of 100%, and five languages classification with six languages had an accuracy of 99.8%.

Window method for digital speech files cryptography

Window method for digital speech files cryptography

Comparative Analysis of Methods Used to Extract Speech Signal Features

The stage of extracting the features of the speech file is one of the most important stages of building a system for identifying a person through the use of his voice. Accordingly, the choice of the method of extracting speech features is an important process because of its subsequent negative or positive effects on the speech recognition system. In this paper research we will analyze the most popular methods of speech signal features extraction: LPC, Kmeans clustering, WPT decomposition and MLBP methods. These methods will be implemented and tested using various speech files. The amplitude and sampling frequency will be changed to see the affects of changing on the extracted features. Depending on the results of analysis some recommendations will be given.

Performing forced alignment with Wav2vec 2.0

Forced alignment, the task of aligning segmentation of audio speech files with an orthographic or phonetic transcript, is fundamental to many types of speech research. Yet the available toolkits for forced alignment are mostly based on the classic HMM/GMM systems, which are outperformed by neural network-based speech recognition models, especially the large-scale speech pre-trained models in recent years. We propose a method of forced alignment utilizing the pre-trained transformer-based model, Wav2vec 2.0. This model has been pre-trained on massive audio datasets, and is subsequently fine-tuned on 360 h of speech data guided by connectionist temporal classification (CTC) loss. The model has learned to jointly recognize phonemes and perform segmentation with or without orthographic transcriptions. During inference, the hidden states are converted to frame-level alignments through post-processing. Our preliminary analysis shows that the model performs competitively when segmenting the TIMIT benchmark, even without orthographic transcription provided.

Processing methods for the detection of landmark acoustic cues

Landmarks (Stevens, 2002) are acoustic cues that are correlated with certain changes in speech articulation, and can be used to infer some of the distinctive features useful for speech recognition, largely the manner features. This project identifies and organizes the processing steps involved in extracting eight types of Landmark acoustic cues for a feature-based hierarchical automatic speech recognition system, in which each module detects one landmark cue from samples of continuous speech, based on knowledge about human speech production and acoustic measurements such as formant frequency and energy. It maps out the processing steps necessary to accurately detect speech Landmarks, and describes a standardized system of modules implemented in Matlab that can accurately identify each of the Landmark cues in speech files quickly and efficiently. The processing methods laid out in this project provide a framework that can be later extended to find other acoustic cues related to place and voicing.

Efficient WPT Based Speech Signal Protection

Speech files are now widely circulated using various means of communication, and sometimes the speech file can be secret, which requires protecting it from any unauthorized person to hear and understand it. The methods which use DES standard for data encryption-decryption are considered ineffective because they need a long time to implement the protection of speech files due to their large size and containing double data with fractions. In this paper research a method which uses WPT decomposition will be proposed, tested and implemented. It will be shown how the new introduced method will increase the cryptography process efficiency by minimizing the encryption-decryption times keeping at the same time excellent values for quality parameters.

The Multi-Domain International Search on Speech 2020 ALBAYZIN Evaluation: Overview, Systems, Results, Discussion and Post-Evaluation Analyses

The large amount of information stored in audio and video repositories makes search on speech (SoS) a challenging area that is continuously receiving much interest. Within SoS, spoken term detection (STD) aims to retrieve speech data given a text-based representation of a search query (which can include one or more words). On the other hand, query-by-example spoken term detection (QbE STD) aims to retrieve speech data given an acoustic representation of a search query. This is the first paper that presents an internationally open multi-domain evaluation for SoS in Spanish that includes both STD and QbE STD tasks. The evaluation was carefully designed so that several post-evaluation analyses of the main results could be carried out. The evaluation tasks aim to retrieve the speech files that contain the queries, providing their start and end times and a score that reflects how likely the detection within the given time intervals and speech file is. Three different speech databases in Spanish that comprise different domains were employed in the evaluation: the MAVIR database, which comprises a set of talks from workshops; the RTVE database, which includes broadcast news programs; and the SPARL20 database, which contains Spanish parliament sessions. We present the evaluation itself, the three databases, the evaluation metric, the systems submitted to the evaluation, the evaluation results and some detailed post-evaluation analyses based on specific query properties (in-vocabulary/out-of-vocabulary queries, single-word/multi-word queries and native/foreign queries). The most novel features of the submitted systems are a data augmentation technique for the STD task and an end-to-end system for the QbE STD task. The obtained results suggest that there is clearly room for improvement in the SoS task and that performance is highly sensitive to changes in the data domain.

A novel NMF-based authentication scheme for encrypted speech in cloud computing

Authentication of encrypted speeches is a technique that can judge the integrity of encrypted speech in cloud computing, even the encrypted speeches have been subjected to common signal processing operations. In this paper, a novel and effective authentication scheme for encrypted speech is proposed. At first, the host speech signal is first scrambled and encrypted by Advanced Encryption Standard (AES). Then, Integer Wavelet Transform (IWT) is performed to obtain the approximation coefficients and the detail coefficients. At last, Non-negative Matrix Factorization (NMF) is employed to generate perceptual hashing, which is embedded into the encrypted speech by differential expansion. In authentication section, the tampered region of encrypted speech is located by comparing the reconstructed perceptual hashing with the extracted perceptual hashing version. Extensive experiments are carried out, which demonstrate that the proposed scheme is not only sensitive to malicious tampering of encrypted speech files, but also robust to tolerate common signal processing operations. The comparison shows that our algorithm performs better than the existing methods on Bit Error Rate (BER) values.

The Effects of ESL Immersion and Proficiency on Learners’ Pronunciation Development

Despite the efforts of existing studies in the domain of L2 phonology to examine ESL learners’ pronunciation development, little research has comprehensively demonstrated ESL learners’ pronunciation improvement in academic immersion contexts. Similarly, few studies have focused on learners’ proficiency levels linked to their developmental success. The current exploratory study investigated the changes of learners’ pronunciation constructs as a result of their ESL program. Seventy-five newly arrived ESL students (25 in each proficiency; beginner, intermediate, and advanced) enrolled in an Intensive English Program in the United States provided their speech responses (to the placement and exit tests from the program). One hundred fifty speaking samples were linguistically analyzed for the following suprasegmental features: fluency (speech rates and pauses) and prosody (prominence and pitch range). Segmental features were analyzed by employing a functional load approach with randomly selected 90 speech files. Findings revealed different developmental patterns among phonological features and proficiency levels; that is, the upper-level learners improved more in fluency and prominence than the lower-level learners. Segmental changes were minimal, suggesting that both high functional and low functional load sounds involve a complex process in learning. Overall findings provide important implications for ESL curriculum planning and development: 1) intonation acquisition can be difficult; 2) skill improvement differs by proficiency level; and 3) level-specific curriculum may be needed.

Secure image hiding in speech signal by steganography-mining and encryption

<span>Information hiding techniques are constantly evolving due to the increased need for security and confidentiality. This paper proposes a working mechanism in three phases. The first phase includes scrambling the values of the gray image depending on a series of keys that are generated using a quantum chaotic map. The second phase generates hybrid keys by mixing a Zaslavsky and a 3D Hanon map that are used to encrypt the gray image values produced after the scramble. Finally, in the third phase, a new algorithm is suggested to hide the encrypted gray image at random locations within a speech file. This algorithm includes the LSB algorithm to determine the hidden bits and the zero-crossing K-means algorithm in selecting locations mining in a scattered manner so that hackers cannot easily retrieve the hidden data of any hacked person. Also used a fractional fourier transform to choose magnitude value as specific data to hide encoded image data. The measures MSE, PSNR, NSCR, and UACI are using to measure the work efficiency in the encryption algorithm, and in measuring the efficiency of the hidden algorithm, use the measures SNR, PSNR, and MSE. The results of the paper are encouraging and efficient compared to other algorithms that performed the same work. Hence our results show the larger the image dimensions used, the better the values.</span>

Speech file compression by eliminating unvoiced/silence components

The main objective of this study is to have the noise component of a speech signal eliminated and compressed by storing the locations and durations of silence regions. The separation between voiced, unvoiced, and silence regions is done by using the Short-Time Energy (STE) and Zero Crossing Rate (ZCR) methodologies. All operations in this study have been performed by using the User Interface (UI) developed on MATLAB®. These operations include voice recording, playing the recording, eliminating the unwanted regions, playing the modified recording, saving original and compressed files, and loading the recording compressed.

Footing e TRP - A arquitetura informal do C-ORAL-Brasil I sob uma ótica socio(inter)acional

Este artigo contempla o aspecto dinâmico que orienta e organiza as relações interpessoais em eventos microecológicos que Goffman (1979) chamou de footing e considera, também, o momento exato da alternância de turno que linguistas como Sacks, Schegloff e Jefferson (1974; 2003), Marcuschi (2003) e Couper-Kuhlen e Selting (2018) chamam de TRP (Lugar de Relevância para a Transição). O objetivo é verificar em dois arquivos de fala espontânea do C-ORAL-BRASIL I, “bfamcv02” e “bfamcv05”, se o TRP das TCUs (Unidades de Construção de Turno) constituídas nas falas em interação provoca mudanças de footing. Para alcançar este objetivo, elaborou-se uma interface entre a prosódia, Hirst e Di Cristo (1998), Bolinger (1986), Halliday (1967), Robert D. Ladd (2008 [1996]), Gumperz (1988) e Selting, (1995), e as perspectivas teóricas que melhor explicam a estruturação de uma interação falada: a Análise da Conversa, a Linguística Interacional e a Sociolinguística Interacional, Sacks, Schegloff e Jefferson (1974; 2003), Marcuschi (1988; 2003; 2008), Couper-Kuhlen e Selting (2018), Goffman (1972; 1974; 1798; 1981), Gumperz (1982; 1989) e Tannen e Wallat (1987). As análises foram realizadas, portanto, a partir de aspectos sociointeracionais e suprassegmentais e os resultados apontaram que o TRP verificado nas TCUs elaboradas nas duas situações comunicativas escolhidas para este artigo provoca mudanças de footing.

Assessment of Extreme Communication Environment With Ultralow SNR: A Benchmark

Accurate estimation of subjective assessment plays an essential role in not only speech quality perception, but also communication environment assessment. Traditional speech quality perception is almost always with an environment, in which the content of speech can be heard clearly. Unlike speech quality assessment, in extreme communication environment with ultralow SNR, like short wave channel with active jamming, speech intelligibility is impaired. Under this condition, subjective assessment as absolute category ranking (ACR) whose scores are made by experienced staffs, and many objective measurements cannot handle this situation. In this paper, we propose information damage level (IDL) to replace subjective ACR as subjective assessment. IDL is the average of scores which is marked on the subjective recognition rate. Under extreme communication environment, it can effectively avoid excessive differences from people to people to some extent. We also provide a new dataset collected in an environment with active jamming, whose speech files are recoded under three different environments, named as indoor simulation environment (EN1), outdoor simulation environment (EN2), and outdoor real environment (EN3). We also benchmark a novel open framework on random forest for direct predicting subjective assessment by combining all possible objective measurements. Experiment results prove the effectiveness of our open framework together with our dataset.

Neural Network Modelling of Speech Emotion Detection

In making the Machines Intelligent, and enable them to work as human, Speech recognition is one of the most essential requirement. Human Language conveys various types of information such as the energy, pitch, loudness, rhythm etc., in the sound, the speech and its context such as gender, age and the emotion. Identifying the emotion from a speech pattern is a challenging task and the most useful solution especially in the era of widely developing speech recognition systems with digital assistants. Digital assistants like Bixby, Blackberry assistant are building products that consist of emotion identification and reply the user in step with user point of view. The objective of this work is to improve the accuracy of the speech emotion prediction using deep learning models. Our work experiments with the MLP and CNN classification models on three benchmark datasets with 5700 speech files of 7 emotion categories. The proposed model showed improved accuracy.