Speech recognition is one of the major research regions these days under speech processing. This paper depends on developing a whole process that takes the input as the text file from the user and provides the output in speech form. This paper proposes a text to speech synthesizer for the Hindi language depends on the coefficients of Mel-frequency cepstral (MFCC) features are extracted to the production and linguistic constraints proposed for modeling the parameters such as intonation, duration, and syllable intensities. The features extracted from the MFCC features are phrasing, fundamental frequency, duration, etc. Neural network models are discovered to confine the features as mentioned earlier, employing MFCC. The performance of the proposed ALO-ANN is computed utilizing objective measures such as prediction error (η), standard deviation (σ), and linear correlation coefficient (χ). The accuracy predicted of the proposed ALO-ANN models is high when compared with other models such as DNN and ANN. The prediction accuracy is high for ALO-ANN models when compared with other models.

An efficient adaptive artificial neural network based text to speech synthesizer for Hindi language

We present a Deep Neural Network (DNN) based statistical parametric Text-to-Speech (TTS) system for Bangla (also known as Bengali). A first step in building a DNN-based TTS system is having large speech data. Since good speech dataset for Bangla TTS is not available publicly, we created our own dataset for our system. We prepared a phonetically rich studio-quality speech database containing more than 40 hours of speech. The database consists of 12,500 utterances. We also prepared a pronunciation dictionary (lexicon) of 1,35,000 words for front-end text processing, which, to our knowledge, is the largest lexicon for Bangla. Our system extracts linguistic features from input text. Then it uses deep neural networks for mapping these linguistic features to acoustic features. We developed two TTS voices using our dataset - one male and one female voice. Both objective and subjective evaluation tests show that our system performs significantly better than the traditional Bangla TTS systems and is comparable to the commercially available best Bangla TTS system.

A Bangla Text-to-Speech System using Deep Neural Networks

Speech is the most natural method for people to convey emotions and communicate. Traditional input techniques are used for machine communication. Communication across distinct regional languages is always unavoidable and takes significant technological effort to make meaningful and effective understanding. Methods of Machine Translation (MT) have been developed in an effort to circumvent this obstacle. Neural Machine Translation (NMT) is one of the projects that has lately seen a significant improvement in terms of human judgement compared to traditional methods such as phrase-based machine translation (PMT) and statistical machine translation (SMT). Therefore, developments in NMT have an influence on the perceived difficulty of humans. Numerous internet translation and mobile application solutions, such as Google Translate, etc., were created to solve this issue. However, such simple translation methods are not accurate for the parallel MizoEnglish languages. In this research, we have attempted to improve language translation by using the superior capacity of NMT to boost the Mizo-to-English translation in the BiLingual Evaluation Understudy (BLEU) measure and achieved 42.65 BLEU Score on the 4 grams. Publicly available Mizo-English corpus was developed.

Mizo to English Machine Translation: An Evaluation Benchmark

A Text to Speech (TTS) system empowers illiterate people by speaking, in their native language, the information available in electronic media. Speech data and corresponding transcript is essential for the development of a TTS system. Most tribal languages in India neither have script nor written literature. In this paper, we present our effort to build a TTS system for Lambani, a tribal language spoken by a group of nomadic people living in several regions of India. We generated a text corpus of about 3000 Lambani sentences, written in the script of Kannada language. These sentences were read by a literate Lambani speaker in a recording studio. The Lambani TTS system was developed by adapting the Nvidia’s Tacotron2 model, pretrained with English speech, following the transfer learning approach. We implemented two versions of the Lambani TTS system. The first version uses HiFi-GAN vocoder and the second uses the WaveGlow vocoder. We evaluated the two versions of the TTS system using both objective and subjective measures. Both Perceptual Evaluation of Speech Quality score and Mean Opinion Score were higher for the Lambani TTS system that used WaveGlow vocoder. 

Text to Speech System for Lambani - A Zero Resource, Tribal Language of India

Speech network analysis and anomaly detection based on the FSS model is analyzed in this research Aiming at the problem of parallel detection and processing of massive data in distributed intrusion detection, a data segmentation algorithm based on capability and load is proposed The algorithm evaluates the capabilities and actual load of the nodes, weighs the actual state of each node and the data distribution relationship in the cluster, and allocates more data to be processed to nodes with strong data processing capabilities and light loads High-order statistics are used to describe the intrusion characteristics of persistent attacks in the link layer of the speech sensor networks, and vector quantitative decomposition is used to analyze the fusion characteristics of advanced persistent intrusion symbols in mobile terminals Different terminals are estimated based on the machine learning algorithms, and the FSS model is integrated to achieve the comprehensive analysis of the speech analytic models The experiment compared with the state-of-the-art methods have proven the efficiency of the framework The detection accuracy is higher than the latest methodologies

Speech network analysis and anomaly detection based on FSS model

This paper describes the development of a text-to-speech system for Assamese language, using Deep Neural Network (DNN). The system is trained with speech data, collected by a consortium, that is available free of cost for academic use. The DNN based method eliminates the need for a grapheme to phoneme conversion; rather, it synthesizes speech directly from the UTF-8 based Assamese script. The results of objective and subjective evaluations confirm that the Assamese speech synthesized using DNN approach is better than the ones synthesized using the traditional hidden Markov model based text-to-speech system.

Development of Assamese Text-to-speech System using Deep Neural Network

This work attempts to build an automatic 1D-CNN based tone recognizer of Mizo, an under-studied Tibeto-Burman language of North-East India. Preliminary research findings have confirmed that along with four canonical tones of Mizo (High, Low, Rising and Falling), a phenomenon of Rising tone sandhi (RTS) with distinct phonetic characteristics are also observed. As per the authors’ knowledge, no work has been reported to identify the RTS along with four distinct tones. Moreover, previous tone recognition works have explored hand-crafted features derived from F0 contour which may not provide the explicit representation of a specific tone category. To address these issues, current work attempts to incorporate the RTS along with four lexical tones and learn tone specific features directly from F0 contours using a 1D-CNN model. Experimental results conducted for speaker independent case show that the proposed 1D-CNN model achieves an accuracy of 68.18%.

S. R. M. Prasanna

Papers

Development of Assamese Text-to-speech System using Deep Neural Network

Learning Mizo Tones from F0 Contours Using 1D-CNN