New Media and Mass Communication www.iiste.

org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

Text To Speech Synthesis for Afaan Oromoo Language Using

Deep Learning Approach
SORESSA BEYENE
Ambo University, Hachalu Hundessa Campus Institute of Technology Department of Computer Science

Raja.K (PhD)
Ambo University, Hachalu Hundessa Campus Institute of Technology Department of Computer Science

Abstract
Text to speech synthesis (TTS) which generate input texts is generate to the speech from texts. TTS is very
important in aiding impaired people, in teaching and learning process. But, to implemented TTS have a lot of
challenging such as text processing, time to phoneme mapping and acoustic modeling for Afaan Oromoo
language. So, Afaan Oromoo language mostly required to text to speech synthesis for development of this
language. The application of Natural Language Processing is provide that input texts pair speech to generate the
desired result outputs of speech in waveforms from prepared text corpus. The normalized text was used for
linguistic features are extracted by using Festival toolkit for Afaan Oromoo TTS. The labeled texts are done
using Festival toolkit, and generated the utterances of texts from scheme file parameters. The Festival toolkit is
used for texts normalized in linguistic extraction from label phoneme alignment to match with speech corpus in
trains and tests. The forced alignment is done by HTK toolkit for prepared environment, checked data extracting
features within timestamps of state level alignment for acoustic feature extracted. So, this study focus on TTS
approach deep learning model based on BLSTM-RNN for Afaan Oromoo language. The RNN model used from
a given input feature sequence to extracted duration model and acoustic model. The implementation is done in
BLSTM-based on RNN using pytorch library on jupyter notebook, create duration model and generated speech
samples from trained acoustic model. We have prepared 1000 texts corpus their matching text transcription from
Afaan Oromoo speech corpus by a female speaker dependent for training 700 sentences and tests 300 sentences
from dataset domains. In this study, two evaluation techniques used. Frist, the Mean Opinion Score (MOS)
evaluation technique is used for intelligibility and naturalness in TTS. The second is Mel Cepstral Distortion
(MCD) which is highly used for objective evaluation in model approach for TTS. So, the performance of this
model was measured and quality of synthesized speech is assessed in terms of intelligibility and naturalness
which results are 3.77 and 3.76 respectively. The total average processed using objective evaluation technique
the speech corpus on 16 kHz standards is generated by MCD BLSTM-based on RNN is 3.89 and merlin wave
generated is 3.71 correspondingly.
Keywords: Text To Speech Synthesis, Mel Cepstral Distortion (MCD), Mean Opinion Square (MOS),
Bidirectional Long Short Term Memory Recurrent Neural Network (BLSTM-RNN)
DOI: 10.7176/NMMC/101-02
Publication date: April 30th 2022

1. Introduction
Text-to-speech (TTS) means input texts is to generate the audio and used for in communication, the sound hear
to human. Natural language processing (NLP) is a field which employs computational techniques for learning,
understanding and producing human language properties at the intersection of computer science, artificial
intelligence and computational linguistics. It is used for both generating human readable information from
computer system and converting human language into more formal structures that a computer can understand.
The Text-To-Speech (TTS) synthesizer is a computer-based system that able to read any text aloud, whether it is
directly introduced in the computer. Among 83 languages which are registered in the country Afaan Oromo is a
Cushitic language that has the greatest number of speakers Ethiopia. Moreover, Afaan Oromoo has 60 million
speaker as a mother thong and as second language. The speech is formed from phonemes and combined together
to form words in Natural Processing Language. The Natural Language Processing study human language learns
with the natural sound and they speak to communication throughout their life. Humans also learns easy and
efficient mode of communication with machines. So, the Natural Language Processing accepts the input texts
pair speech corpus and able to generated speech output after text analysis method. This text analysis contains
text normalization, sentence segmentation, tokenization and non-standard words like abbreviations into full word
covert (Trilla, 2009).
The method used to develop a text to speech system in concatenative synthesis is based on speech signal
processing of natural speech databases and speech signal processing able to perform speech in waveform. In
such a manner that, appropriate speech units are concatenated to construct the required speech. The segmental
database is built to show the main phoneme extract features of a language from the concatenative synthesis

15
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

recorded audio. The method used to set the phonemes is built diphones units, representing the phoneme to
phoneme junctures. No uniform units are also used diphones, syllables and words in this concatenative methods
for speech synthesis.
The aims to generate a mapping between the textual diphones and their equivalent speech units. Each
diphone is represented by two characters, consequently producing speech unit of that diphone. Syllables are any
words with the exception of abbreviations and acronyms were considered in system written words considering
rule based constrictive depend on language in consonant and vowels.
This syllables is used for small database select unit in TTS systems. The words Systems that simply
concatenate isolated words or parts of sentences, are only applicable when a limited vocabulary is required
typically a few words and the sentences to be pronounced respect a very restricted structure. The synthesizer of
the speech segments, and performs some signal processing to smooth unit transitions and to match predefined
prosodic schemes. The direct pitch synchronous waveform processing is one of the most simple and popular
synthesis algorithms.
The multi-pulse excitation linear predictive Coding system produces synthetic speech that is more natural
sounding than the classical linear predictive coder. In the multi-pulse excitation linear predictive coding system,
the excitation signal is modeled with a few pulses per frame of speech regardless of whether the frame is voiced
or unvoiced. The quality of the synthesized speech improves with the number of pulses used per frame. Pulses
are computed by minimizing the weighted square error between the original speech and the synthetic speech.
The Digital Signal Processing (DSP) turns NLP representation into an output signal (Tilahun, 1993). The
advantage of DSP module are the controlling the duration time and frequency (aperiodicity) of the vocal folds so
that the output signal matches the input requirements show speech signal processing.
Generating (converting) text to speech encompasses both natural language processing and digital signal
processing (Morka.M, 2003). The application of Natural Language Processing (NLP) is to produce a phoneme
translation of the text reading with the required intonation and rhythm (Alula, 2010). Text analysis is the
responsible for analysis of input text into soundable texts. To achieve this, it organizes the input texts into
control the lists of words and proposes all possible part of speech categories for each word taken individually on
the basis of words, and then considers words in their context of the recorded and written. Phonetic analysis is
purpose for the finding of the phonetic translation of the incoming text. This work can be organize in different
ways dictionary based and rule-based strategies (context-based).
The speech is greatly affected by accents occur at stressed syllables and form characteristic words in the
pitch tones. The transition periods between syllables place of produce and found to be dependent on the nature of
articulation of boundary sound units. The component are responsible for generate the acoustic sequence required
to synthesize the input text by finding the pronunciation of individual words in the input text. The style of
pronunciation was influenced by the gender, physical state, and emotional state and focused on the speakers. The
prosody features depend on many aspects like the speaker characteristics gender, emotions and meaning of the
sentence (Samuel, 2007).
Speech is the most efficient and natural way to communicate with each other. Speech is the agreement and
common understood of communication between human being. When human read text as the rule based of the
phonology, with native language (mother language) speech, the person hears the individual words and sounds.
Every speech not converted to standard written words or texts. So, speech can be written using letter to sound
format the words. But, this is not true if the person hearing the speech is not familiar with the language.
The conversation of text to speech are several method .The development of society and economic system
since prehistory time has been paralleled by a growth in man’s dependence upon technology. Speech enabled
interfaces are desirable because they promise hands-free, natural, and ubiquitous access to the interacting device
(Solomon, 2005). However, it is one of the least supported and least researched languages in the world. He
remarkable works some contributed doing on text to speech synthesis for Afaan Oromo languages (Solomon,
2005).
Amharic Text-To-Speech Speech Synthesis System stated phonetic once analysis done, the final block of
NLP to prosody generation, which is responsible for finding correct intonation, stress, and duration from written
text in prosodic features. The prosodic features are features that appear when to input sounds together in
connected speech (Alula, 2010). It is advanced in prosodic features as successful communication depends on
intonation, stress and rhythm as on the correct pronunciation of sounds. Intonation, stress and rhythm are
prosodic features. The rule-based methods use manually produced rules, extracted from utterances structures.
Afaan Oromoo speech synthesis system was developed on a hidden Markov model method (Wosho, 2020).
The HMM model stated for the neighbor rule and able to processes limited datasets. The researcher not stated
and mentioned acoustic feature and linguistic feature in statistical parameter used for text to speech synthesis
based on Hidden Markov Model (HMM).
In NLP several methods have been used in the phone duration model working like linear regression models
are based on the assumption that among the features which affect the segmental duration there is linear

16
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

independency. The models used to predictions linear regression with small amount of training data not model the
dependency among the features extracted. On the other hand, decision tree models and in particular classification
and regression tree models, which are based on binary splitting of the feature space, can represent the
dependencies among the features. The phone duration model, where the segment duration prediction was based
on a sum of factors and their product terms that affect the noise duration (Yang, 2014).It can effectively
extracted the hidden internal structures of data and use more powerful modeling capabilities to characterize the
data.
Deep learning is a part of machine learning which trains the model with large datasets using multiple
layers and Feedforward neural networks for single layer. Deep learning that is capable of process in short
time duration large dataset of training become important method for text to speech system. The HMM
based for speech synthesis method maps linguistic features into probability densities of speech parameters
with various decision trees. The deep learning based method directly perform mapping from linguistic
features to acoustic features with BLSTM-RNN which have proven extra ordinarily efficient at learning
inherent features of data. It is important for readers better understand the development process of these
methods used deep learning approach. Deep learning based models approach significant progresses like
handwriting recognition machine translation (Sutskever, Vinyals, & Le, 2014). The speech recognition (Graves
A. Mohamed, 2013) and speech synthesis (Zen & Alan, 2009).
Recurrent Neural Networks (RNNs) is the also the family of deep learning that are well-suited for pattern
classification tasks whose inputs and outputs are sequences, for example tasks such as speech recognition,
speech synthesis, named-entity recognition, language modelling, and machine translation (M. S. Al-Radhi, 2017).
The Recurrent Neural Network (RNN) method to model speech-like sequential data that represents associations
among bordering frames training duration model and acoustic model. It can also practice all the accessible input
features to forecast output features at each frame. Particularly, the RNN model is different from the DNN since it
operates not only on inputs but also on network internal states that are updated as a function of the entire input
history. Training RNN incorporates backpropagation.
The RNN connections are able to mapping the utterance and understanding input datasets for train the
acoustic sequence, which is purpose waveforms to show speech in signal processing to generated prediction
outputs desired (M. S. Al-Radhi, 2017).
Long short-term memory networks (LSTM) are a class of recurrent networks composed of units with a
certain structure to manage better with the vanishing gradient problems during training of recurrent neural
network and maintain potential long-distance dependencies (M. S. Al-Radhi, 2017). This focused on linguistics
adapted with technology. The Text to Speech is soundable communicate information to the user, where digital
audio recordings, for developing a user of speech synthesizes in Natural Language Processes. The performance
of evaluation used intelligibility and naturalness encourage to investigate the text to speech synthesizer in Afaan
Oromo language. So, this research training datasets are extract the linguistic features for Afaan Oromoo
language.

2. Related Work
In this chapter, from presented the review of number of speech synthesizer developed focusing on different
approaches. The deep learning approach is one of advanced in Natural Language Process for text to speech
synthesis. Deep learning approach is extract the hidden internal structures of data and use more powerful
modeling capabilities to characterize the data. Therefore, concerning to Afaan Oromoo language from the
previous work done (literature reviewed), text to speech synthesis for Afaan Oromoo language has not still
methodically exploring using deep learning approach. The researcher used to the deep learning approach with
the BLSTM-based on RNN, text to speech synthesis for Afaan Oromoo language and consideration the model
used to synthesize the desire fully context labels (speech, texts) pairs which are phone mapping,
duration(linguistic) modeling, acoustic modeling, generated speech for Afaan Oromoo language Table 1 showed
below.

17
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

Table 1: Table 1:Summary of Related Work on Text to Speech

Author Title method Number of Result Limitation
datasets
(Morka.M, Text-To- Rhyme Test, for test 15 For person one Isolated word
2003). Speech approximately words (Type 1- people) I utterances are very
System for 43.33% and person slow as compared a
Afaan Oromo 2 83.33 % of the continuous speech
Language test data
(Samson, Text-to- Concatenative unspecified Diphone of 75% Prosodic method
2011). Speech within Residual and triphone 54% not include. model
System for Excited Linear respectively was unidentified
Afaan Predictive Coding
Oromoo (RELPC)
(Tewodros, A TTS Festival tools for 1369 1369 Performance of the The Speaker
2009) synthesizer generated nonsense and system 78.3%. specific intonation
for Wolaytta nonsense the recorded and speaker
language Praat for record words 841 specific duration
audio not considered
Letter-to-Sound
Rule
(Alula, Amharic TTS Hidden Markov 80 standard Standard Lack of prosody
2010) system Model toolkit words and 30 evaluation 76.7% analysis building
non-standard and non-standard part of speech is
words (NSWs) 70%, respectively crucial
(Kedir, Text to Statistical 400 sentences In MOS evaluated Audio conversation
2020). Speech parameter based are used for intelligibility is 4.3 mechanism, spoken
Synthesis for on Hidden training and 10 and naturalness 4.1 language non-
Afaan Markov Model sentences for of the speech standard words like
Oromoo testing synthesized time, acronym are
respectively. not considered
(Alem F., Bangla text- Deep learning 1,35000 words Training 94%,
2007) to-speech based on deep testing 3% and
system neural networks validation 3%.
(Zen H. S., Text-to- Bidirectional 13,100 speech Not specified Not effected
2013) Speech LSTM based on generate Acoustic
Synthesis for Recurrent Neural model
English Networks

3. Research Methodology
This section deals with the Afaan Oromoo text to speech synthesis system methodology within architecture. It
explains the whole of design the representation and description of components. The design architecture easily
understand the method of a text to speech synthesis for Afaan Oromoo approach to deep learning. The training
phase a text corpus is used passes through the text analysis process (tokenization, normalization and linguistic
features extraction).The extracted features are used as an input for the duration model, from the speech corpus
acoustic features are extracted. The input is used for the acoustic model with the linguistic features and duration
information generated by the duration model. Output of the acoustic model is used as input for the Vocoder to
generate the final speech. The generation of speech duration model and acoustic features are extract. The
extracted features are then used as an input for duration model and acoustic model training. Finally, the speech
synthesis is evaluated. So, the architecture of text to speech using BLSTM-based on RNN for the Afaan Oromoo
language is illustrated in Figure 1.

18
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

Figure 1: proposed model for Afaan Oromoo Text to Speech

3.1 Data Collection and Preparation

To collect datasets, for achieving the general objective. In is very challenge in text to speech for Afaan Oromoo
not archived and no standard datasets. Therefore, some text accesses the raw data source from Afaan Oromoo
Wikipedia book, serlugaa Afaan Oromoo book and Oromia broadcast network (OBN) of Afaan Oromoo
language from their websites. Thus, the data collection process represents one of the significant stages of this
study. The acoustic data has its transcription which allows the training of models and duration data separately
from total dataset text corpus and speech corpus total 1000 are label to phoneme alignment. Text corpus consists
of total 1000 sentences, 12359 total words and 2043 unique words. The process of text and speech corpus
mapping is performed by applying the forced alignment tool used htk toolkit for linguistic and an acoustic
model. So, this study focused on thesis scopes to collect datasets (texts and speech corpus) for Afaan Oromoo
language. The speech sample is recorded at 16000 Hz mono sound standards and the files are stored in waveform
format (.wav). The Festvox system and full style label use to htk toolkit to make it easier to create raw text pair
speech signal to suitable machine learning and easily understanding format. The Praat is free open source
software that used to record the each text into speech in wave form. For reduce noise using microphone and a
normal office computer made up the hardware equipment record the texts for preparation speech corpus in wave
format.
The corresponding text files are prepare by manual ways from data sources due to benchmark results
dataset is not available for Afaan Oromoo regard to text to speech. The database and experiments contains an
audio record by a female speaker dependent automatically parameter extracted was used. These dataset
preparation is focused on linguistic and acoustic feature train for text to speech conversion. The linguistic and
acoustic features are focused the on original texts translation and recorded audio respectively.
The next steps is to record the speech corpus, the text parallel recording speech is done by a native and
professional in linguistics female 25 year old speaker dependent in a quiet room to reduce disturbance. For high-
quality recording microphone with noise isolation facility used the personal computer. The recording process is
carried out used praat free software which provides tools for sound file recording, end-pointing, rate
manipulation, and noise reduction.

3.2 Text Analysis

The text analysis module is used to do tokenization and text normalization and then the normalized text is used

19
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

to extract linguistic features and the extracted features are used as an input for the duration model. The durations
for each phone are first is predicted using the duration model using the pre-trained model. Then after, the
duration model is used to predict the timestamps of each phoneme mapping duration of each phone to be
synthesized. The abbreviations and acronyms are not pronounced as they are input or written. The first work for
the raw input abbreviation to replaces the abbreviations and acronyms in their expanded as rule in Afaan
Oromoo language. The python scripting language used for performs this normalization with sample Afaan
Oromoo abbreviations is common known when expanded. Text corpus preparation for linguistic feature
extraction for each of wave (.wav) files, it requires text (.txt) files with exactly the same name that contains
exactly the text that was recorded in the speech corpus. Text to speech (TTS) synthesis is conversion of text into
speech.TTS system consists in text analysis, where the input text is transcribed into linguistic representation. In
this TTS part, the input sentence is segmented into tokenize.
The Utterance structures are central when synthesizing speech used the Festival speech synthesis system for
text transcribed into phonetics. The relevant properties of the utterance to be synthesized in a way they are
specifications of desired utterances. Instead of created an utterance structure by synthesizing some text, we can
create an utterance structure from a speaker's natural utterance, taking all properties from the speaker's utterance
instead of predicting them from texts. We have an utterance that looks like a synthesized utterance, but which is
guaranteed to be a valid natural utterance, with correct phonetic properties.
An utterance structure consists of many items like single words, syllables, phones, phrases and items are
connected through several relations. Exactly which relations are present depends on the synthesis method,
among other things. But some of them are always present, such as the word relation connecting word items, the
syllable relation connecting syllables, the segment relation connecting phones, and the syllables structure relation
connecting items.
The requirements for generated utterances structures by the Festival system and provided a script to create
utterance structures from text corpus, linking the hierarchical relations by their time information ( the script
looks at the end times of items in lower relations to decide which items in a higher relation must be their parent).
However, these label files need and for punctuation marks. The punctuation is represented as a feature at token
level by Festival into utterances. The aligner did not differentiate between words and tokens, instead of word
label files created by the aligner are actually tokens in Festival. So in order to have punctuation at the token level
in Festival system, we have that information in the word label files. We have able to make use of the features and
the scripts for converting to utterances created one utterance per word label file. Unfortunately, the librivox
recordings are usually quite long, and it is not possible to cut them into reasonably small pieces for Festival
system before running the aligner, because then you would have to sit down and segment the text(.txt) files
accordingly. Speech corpus files move all short wave files into the wav directory within
/home/soro/merlin/Text2Speech/AO_Speech_Syntesis/soro and phonemic files and the words files to this
directory, so everything is at the same place. The raw audio preparation for training is used to htk toolkit audio
make feature acoustic extraction Mel generation coefficient (mgc), fundamental log frequency (lf0), and band
periodicities (bap).
The made the file was the extracted acoustic feature from composed acoustic features. These made train
was generated in scripts. The full context train labeled extracted from text corpus for utterance build. The made
file was generated used HTK tools for full context label like monophone mapped and full labeled texts pair
speech. The list of made file is generated under folder master label files and model list files. The acoustic
features accesses is extracted audio waveform from make features which includes log fundamental frequency
(fl0) represented the pitch and Mel generalized cepstral features which displayed spectral parameter of the
speech.
3.2.1 Text pre-processing
Text pre-processing is task must be expanded into full words digits and numerals. Another task is to find correct
pronunciation for different contexts in the text. The synthesis speech in text analysis of the raw input text into
pronounceable words. From the texts contains string punctuation marks to clean like!”#$ %&()*+,-./:;?@[\]^_{|}
~\t\n providing many functionalities in deep learning.
The construct datasets from pertaining linguistic duration and acoustic features because computing features
from the label files on-demand are performance, particularly for duration model extract features used to python
script and using bash in form (./filename.sh) as following steps:-
Step1. Prepare corpus similar text within speech corpus in this same file name in text and audio filename
separately in folders. The purpose for preprocessing text pair speech easily undestanded for machine learning in
linguistic (text) analysis.
Step2. From saved file text pair speech automatically created label phoneme alignment and state alignment using
method htk tool and festival fronted tools
Step3. Create the duration model and acoustic model
step4. Create training duration model and training acoustic model

20
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

step5. Speech synthesis from feature prepared

step6. Test speech synthesized duration and acoustic using the name of voice created.
3.2.2 Normalization
The first iteration in the loop again gets the basename. The second eliminates pauses and the header from the
aligner word label file. The third extracts the lines of pause labels from the word label file and stores them in a
file with extension pauses. These lines are important because they contain the label times for the pauses, which
we don't want to lose. The words with punctuation file by writing a hash sign for the end of the header into it
(overwriting a possibly existing file of that name). The speech preprocess was normalization speech record in
wave format. The displayed the graph used to matplotlib and librosa display the wave plot attribute for sample
rate when reducing noises.

Figure 2: Noise reduced for speech corpus

The Figure 11 displayed linguistic features output the text translation, which typical desired additional re-
sources of pronunciation of the language.
The TTS that translate raw text converted to linguistic used festival tool fronted and WORLD tools the
phone alimented for all datasets after label the files text pair speech corpus. So, the linguistic representation text
that is generated from festival tool outputs. The HTK scripts transform with the contextual labeled information
and structured Festival represent the list of phoneme mapped in utterance information full context labeled style.
To create the directory name of folder to copy into template desired the datasets speech pair texts.

3.3 Linguistic Feature Extraction

The forced alignment in state align is created the training labels used to htk tools and phone align created the
training labels with Festvox tools. To check how your sentence was synthesized. We have synthesized an
utterance and stored it in a variable utterance used to Festival system (set! utt (Say Text “iji waaqayyoo iddoo
hundumaa jira, isa hamaa fi isa gaarii ni arga. ")) checked which relations are present by (utt.relation names utt)
generated the following relations token , word , phrase ,syllable , segment, syllable structure , intonation , target
unit source segments and wave.
To train a speech corpus in wave files converted into phone align labels for mapped phone and timestamps
for duration of the texts in corpus when recorded. So, choice one convenience and set the option accordingly in
global settings configuration file.
The training process for both synthesis systems requires that each sentence of the is obtained from phone
align created training labels transcription was match with texts pair speech corpus. The global settings
configuration file helped for feature extract in duration mode depends on source kind from waveform, source
format wave files, and target rate at 50000 for timestamps and target kind for Mel frequency cepstral coefficients.
The linguistic features alongside timestamps are then used because the input for the BLSTM acoustic model
to get corresponding compressed acoustic features, which include parameters Mel generated coefficient (mgc),
fundamental logarithm frequency (log F0) and band (BAP) are generated form linguistic source files from label
phone aligns text files label (.lab).
Features required to train duration model assign the variable X as linguistic and Y as duration. The variable
X as duration source find the linguistic source, add frame features used phone alignment from question paths file
dataset sources the declared variables X duration source and Y duration. We saved the features for duration
mode duration linguistic feature dimension in variable X duration and duration feature dimension variable Y
duration. For the variable x, y to enumerate in zipped code, name split text basename the collected files in
variable X duration join with automatically make directory x and y path converted into binary formatted(.bin)
this processed used for normalization duration model produce linguistic features.

21
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

Figure 3: Prepare Features for Linguistic Feature Training

3.3.1 Utterance Length by Histogram Visualization
Then, the duration features are taken as an input for the Vocoder to generate speech in waveform.

Figure 4: Utterance Lengths Histogram Train and Text Visualization

The Figure 4 is show the final step to synthesized speech waveform from the recover spectrum and can
listen to the sound and pysptk speech representation within waveform using on jupyter notebook using librosa
audio visualization the sentences show the plotted different duration records. The utterance lengths histogram
as total number of utterances 144 and total number of frames 9153 sample.
The feature normalization in this linguistic feature extracted is to show the utterance training feature from
constructed duration and acoustic datasets in precomputed the input variable within the focus (target) in acoustic
datasets. The features load on demand to epoch the utterance length from total number of duration and acoustic
training. The linguistic feature is variant in every scale of at everyone dimension. This linguistic clear to
visualization when the normalization the feature of linguistic and variance while applying the normalization
duration and acoustic features is computed. The linguistic features become to normalization at any one indexed
the feature domain between 0.01 and 0.99 values.

3.4 Acoustic Feature Extraction

The process for training the acoustic to extracted feature model was no more difference that for train duration
model. Since, configuration files which used to initialize and BLSTM based on RNN model. The utterance
lengths histogram as total number of utterances 144 and frame is 128033 sample utterance acoustic train show as
Figure 15 below.

Figure 5: Utterance Lengths Histogram Acoustic Train

The Figure 5 shows, the output total number of utterances lengths acoustic and train. The output
visualization histogram in utterances acoustic within binary at 64 computing the total number of frames used to
summation in attributes of numpy acoustic training.

22
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

3.4.1 Spectrogram
The first spectrograms to generation of speech synthesis to analysis is used python speech parameter toolkit
(pysptk) and important used the librosa for representation features. The pysptk was contains have sequential
include windowing, mel-generalized cestrum analysis, visualize spectral envelope estimates and F0 estimation.
The spectral parameter estimation and visualize its spectral envelop estimate. The first option for an audio
features representation is spectrogram. The spectrogram two dimensional tensor is displayed vertical dimension
indexes times and horizontal dimension frequency.

Figure 6: Shape of Spectrogram Feature

Figure 6 shows acoustic features that are the acoustic properties of speech signals for speech analysis. When
comparing a spectrogram to a normalized spectrogram, the resolution between the frequencies that matter for
speech is higher and there is less redundant information. It emphasizes details in lower frequencies that are
critical for speech modeling and de-emphasizes the high frequencies.
3.4.2 Mel-Frequency Cepstral Coefficients (MFCC)
The second of the audio feature representation is MFCCS. The aim of the behind MFCCs features same as
spectrogram feature at time window. The MFCCs feature representation feature vectors that characteristics
sound within the window. The remember that MFCCs feature is more lower dimension the spectrogram features,
which help as acoustic model to avoid overfitting to the training dataset.
The visualization audio feature representation entire in the tensor assume values close to zeros used to
MFCC shape. However, deeply get the details of how MFCCs are calculated. This spectrogram displayed using
the python speech features python package. The generated spectrogram features from the MFCCs are normalized
in was representation. This focused on MFCC features is the same as spectrogram features at each time window
and the MFCC feature yields a feature vector that characterizes the sound within the window. So, that the MFCC
feature, which helped an acoustic model to avoid overfitting to the training dataset shape use torch library sizes.

Figure 7: Audio Feature Representation in MFCCS

The Figure 7 show the results of the data in sampling rate in Mel frequency cepstral coefficients at 13 plot
in figure size (14, 4) is used to librosa display speech show the MFCC within time rates.
Based on their semantic interpretation audio features are classified into physical and perceptual features.
Perceptual features are basic features that are perceived by human listeners such as loudness, pitch, rhythm, and

23
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

timber. In contrast, physical features are properties that present mathematical, statistical, and physical properties
of audio signals. The most commonly used acoustic features in speech synthesis are the Mel Frequency Cepstral
Coefficients. The acoustic features are often used as a low level audio representation to bridge the synthesizer
and the Vocoder in the backend of a TTS system. In this study, we used Mel Frequency Cepstral Coefficients
and linear scale frequency spectrograms as an intermediate acoustic feature representation. To be specific, the
output of the feature predictor and the conditional input of the Vocoder are sequences of Mel and linear
frequency spectrograms. The frame period at five and the trimming zero frames of the spectrogram is small
power to make good visualize using pyworld (python world).

3.5 BLSTM based on RNNs Model

The defined model used to bidirectional long short term memory based on recurrent neural network variable
declare RNN models for duration and acoustic within bidirectional long short term memory activation using
pytorch for simple implementation.

Figure 8: BLSTM based on RNNs Model

3.5.1 Training Duration Model
Specifically, to create the directory name database for the processed dataset and experiment using on Linux
operating system. The experiment directory include the label texts which contains alignments of phones
generated from linguistic (duration) and the states alignment generated from speech corpus. Additionally, the
experiment directory contains the extracted speech record features in folder linguistic (duration) and acoustic
model. For a little soundness check from the two alignments in labels (time to phoneme) files for the text pair
speech have different lines of pronounced in each characters in words. In everyone line of files label represents a
single alignments in time to phonemes. The desired for a given sound files in its state alignment have more lines
than its phoneme label files. In this one phoneme made up of multiple states in the machinery use for generate
alignments. To look details the acoustic model used to do forced alignment to generated the labels, but the
difference between numbers of lines in the alignments files are around the expected numbers of the text files
‘cafeen teesse’ which label name A0_00202.lab (A= Afaan Oromoo from listed ID generated) have total
phonemes twenty (20) and two hundred state level in acoustic (audio) file. This gives to know how many
acoustic in label state alignment have phonemes is two hundred divided for twenty (200/200) result ten (10)
states per phoneme. The model duration training was nicely. It kept decreasing training and test loss over time.
But there is one thing the test loss was stopped decreasing and started to increase. This means that the network
had started to overfitted to the training set and regularization techniques such dropout and obtain extract of the
duration model displayed the finished train datasets text pair speech within number of epoch 25.

Figure 9: Train Duration Model

24
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

From this Figure 9, the number of epoch or iteration increase it displaying different training and test loss in
datasets. So, good result displayed.
Table 2: Duration Model using Phone Align and State Align
Labels phone align Labels state align
Learnin Valid in Correctio Test in Learnin Valid in Correctio Test in
g Rate RMSE n RMSE g Rate RMSE n RMSE
0.002 6.777 CORR 7.665 0.002 6.826 0.624 7.840
frames/phone 0.633 frames/phone frames/phone frames/phone
me me me me
Table 2 describe duration model label phone alignment and state level alignment to demonstration speech
pair text train data, test data for train, valid, and test respectively evaluation used to RMSE within learning rate at
0.002 in all data sources 63.3 % in label phone align was one of preferred.
3.5.2 Training Acoustic Model
The train acoustic model is somewhat good to decreasing training and validation loss over time. But there is one
thing. The validation loss to be stopped decreasing and started to increase showed. This means that the network
had started to overfitted to the training set and regularization techniques such dropout and obtain predictions of
the acoustic model displayed finished train datasets texts pair speech at number of epoch 25.

Figure 10: Train Acoustic Model

The challenge from Figure 10 when, we repeat execute programming for training the shape of diagram train
and loss are changed. The processing train and validation are parallel increase to control the overfitting models.
Table 3: Acoustic Model Phone Alignment
Labels phone align valid
Learning Rate Valid MCD BAP F0 in RMSE CORR VUV
0.002 6.762 dB 0.246 dB 19.433 Hz 0.538 11.403%
Labels phone test
Learning Rate Test MCD BAP F0 in RMSE CORR VUV
0.002 6.704 dB 0.262 dB 15.264 Hz 0.700 8.907%
The Table 3 describe acoustic model label phone alignment average voice model single female audio
records from A0_00202 to A0_01000 within 9153 utterances and trained datasets evaluation results. From this
label state align test correct or accuracy 70% at learning rate was 0.002.
Table 4: Acoustic Model State Alignment
Labels state align valid
Learning Rate Valid MCD BAP F0 in RMSE CORR VUV
0.002 6.559 dB 0.242 dB 19.573 Hz 0.529 11.655%
Labels state align test
Learning Rate Test MCD BAP F0 in RMSE CORR VUV
0.002 6.586 dB 0.259 dB 15.309 Hz 0.701 8.821%
Table 4 show average speech model single female speech records from A0_00202 to A0_01000 (A is
stands Afaan) within 9153 utterances and data distribution train, valid and test respectively evaluation results.
From this label state align test correct or accuracy 70.1% at learning rate was 0.002. The displayed label state
level alignment for tested the labeled path by using jupyter notebook within different package install in python
scripts.

3.6 Generated of Speech Samples

The acoustic training model text, audio pair, which is typically needed to train acoustic models and load wave
file to gain text file. The generated speech from the label phoneme alignment for objective test displayed used to

25
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

BLSTM base on RNN.

Figure 11: Sample Generated Speech Use Model

This Figure 11 shows the sample generated speech from text pair speech corpus. From this generated
speech it is important to download and adjusted within playback speed listen for regulate for many application of
TTS like teaching style and another.
(A0_0001 "iji waaqayyoo iddoo hundumaa jira, isa hamaa fi isa gaarii ni arga")
(A0_00208 "fardi dammaqe")
(A0_00209 "waraabeessi boolla keessa dhokate")
(A0_00211 "qe'een isaanii onte")
(A0_00212 "sareen dutte")
3.6.1 Speech in Waveforms
The load the speech was generated into with SciPy input and output wave file packages. Speech synthesis using
the spectrogram, and Mel-Frequency Cepstral Coefficients to extract acoustic feature. So, we have converted the
extracted acoustic feature using jupyter notebook within pysptk and pyworld in displayed to waveform showed
as Figure 12 below.

26
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

Figure 12 : Generation of Speech in Waveform

The Figure 12 is described, generated the speech in waveforms sample from test label to display librosa in
wave plot and pyworld speech synthesis method. This is removed part of silent and silent periodicity in frame of
fundamental frequency (f0). The librosa display in attribute wave plot is generated speech within frequency rates
in waveforms.

4.1 Evaluation of Text to Speech Synthesized for Afaan Oromoo

Different ways to evaluate the text-to-speech systems occurs. The process model evaluation performance to
perform used two ways for TTS Afaan Oromoo language the subjective and objective evaluation.
4.1.2 Subjective Evaluation
The subjective evaluation way to evaluating to listen the tests audio and the listener tackles the speech quality
and naturalness.
A total of 13 native Afaan Oromoo speakers and reader randomly selected to evaluated speech quality, the
generated speech is given for evaluator that focus on the intelligibility and naturalness of the synthesized speech
text pair speech given as Figure was showed. The reason we have chosen from total datasets only thirteen (13)
sentences for subjective evaluation to checking the understanding and check the outputs also difficult as the
whole datasets tested.

27
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

Figure 13:Text pair Speech for Subjective Test

This Figure 13 show the first question is targeting in measuring the intelligibility of the synthesized speech
and the second is aimed at measuring whether the synthesized speech is human like or not.
The mean opinion score (MOS) test is chosen for this evaluation, which allowed us to score and compare
the global quality of TTS systems with respect to naturalness and intelligibility thirteen (13) sentences for tested
by native Afaan Oromoo speaker and professional have in Afaan Oromoo language for Naturalness from each
evaluators represented as participants p1-p13 and sentence one(s).
Table 5: Speech Quality for Naturalness from Evalutors
Evaluators s1 s2 s3 s4 s5 s6 s7 s8 s9 s10 s11 s12 s13 Average
p1 3 4 4 5 4 4 5 3 3 4 5 5 4 4.076923
p2 2 4 4 5 3 4 5 2 4 3 4 5 3 3.692308
p3 4 3 3 5 4 4 3 3 5 5 2 3 2 3.538462
p4 5 3 4 4 5 3 4 4 4 3 4 2 3 3.692308
p5 4 2 3 4 4 4 4 4 4 2 2 3 5 3.461538
p6 5 4 5 5 5 5 5 4 4 4 5 4 2 4.384615
p7 3 3 3 4 4 4 3 4 5 3 4 5 2 3.615385
p8 5 5 5 4 4 4 2 2 3 4 5 5 5 4.076923
p9 4 4 4 3 3 4 2 2 3 3 2 3 4 3.153846
p10 3 4 4 4 5 5 4 4 2 3 5 5 5 4.076923
p11 4 5 1 2 5 3 4 5 5 5 4 2 3 3.692308
p12 3 3 4 2 5 4 4 5 5 5 5 4 4 4.076923
p13 4 4 3 3 3 3 3 4 5 4 4 4 1 3.461538
Total Average 3.769231
The participants are allowed to listen to the audio recorded samples before they check the developed text to
speech synthesizer. Subsequently, each participant plays the sample record speech to check the quality of the
speech. After listening the synthetic speech from the system, the participants are requested to fill marks on the

28
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

properly given text pair speech for intelligibility.

Table 6: Speech Quality for Intelligibility from Evaluators
Evaluators s1 s2 s3 s4 s5 s6 s7 s8 s9 s10 s11 s12 s13 Average
p1 3 4 4 5 4 4 5 3 3 4 5 5 4 4.076923
p2 2 4 4 5 3 4 5 2 4 3 4 5 3 3.692308
p3 4 3 3 5 4 4 3 3 5 5 2 3 2 3.538462
p4 5 3 4 4 5 3 4 4 4 3 4 2 3 3.692308
p5 4 2 3 4 4 4 4 4 4 2 2 3 5 3.461538
p6 5 4 5 5 5 5 5 4 4 4 5 4 2 4.384615
p7 3 3 3 4 4 4 3 4 5 3 4 5 2 3.615385
p8 5 5 5 4 4 4 2 2 3 4 5 5 5 4.076923
p9 4 4 4 3 3 4 2 2 3 3 2 3 4 3.153846
p10 3 4 4 4 5 5 4 4 2 3 5 5 5 4.076923
p11 4 5 1 2 5 3 4 5 5 5 4 2 3 3.692308
p12 3 3 4 2 5 4 4 5 5 5 5 4 4 4.076923
p13 4 4 3 3 3 3 3 4 5 4 4 4 2 3.538462
Total Average calculated by MOS 3.775148
Finally, according to mean opinion score, the mean results are calculated as per the respondents’ responses.
Table 7: Average MOS Result of Afaan Oromoo Speech Synthesizer
Evaluators Intelligibility Naturalness
average score results 3.77 3.76
The results obtained from Table 7, the subjective evaluation are collected for the 13 test sentences. For all
the test sessions, results is organized with their set of scores rated by each subject. In order to select the
appropriate test to analyses the data, the first step was to apply a normality test. The results revealed that data
was normally distributed. Hence, a parametric test was used. The sentence was ranks given by evaluators for
naturalness and intelligibility by using calculate the MOS results. The intelligibility is good result values 37.7%
was responded from evaluators.
4.1.3 Objective Evaluation
Objective evaluation where measurement of speech quality performance approximated by applying appropriate
speech in waveforms. One method using Mel cepstral distortion for evaluating objective and it helping the
difference between speech syntheses. This difference shows how the reproduced speech is related to the natural
one and not extract like natural speech production.
Table 8: MCD Objective Test
Generate BLSTM-based on RNN Merlin _AO_Speech_Syntesis
average score results 3.89 3.71
Table 8 shows the MCD evaluation and test audio pair texts within BLSTM based on RNN and Merlin
Afaan Oromoo speech synthesis female speech corpus analysis used to speech signal toolkit and pyworld on
jupyter notebook. The total average process of MCD at Mel cepstral coefficients, the recoding audio were 16
kHz to generate the MCD BLSTM-based on RNN is 3.89 and merlin wave generated is around 3.71 respectively.

5. Discussion, Conclusion and Recommendation

5.1 Discussion
The results Discussion from experiment overall and all model as tested on the spectrogram features and MFCC
the model training for features synthesis. The output of the using Bidirectional LSTM based on RNN model
convert text into speech. Although, the results were achieved on an overfitted model, with longer training and
hyperparameters optimization, to achieve desired results on validation data within developed system. At the first,
the generated MFCC is perhaps the most important result obtained from the Recurrent Neural Network part of
the system.
Using the parameters is learnt by the model, generate MFCC is predicted sequences model from speech
source. The converted of the MFCC input a shape of coefficients and sequence length. The Mel generated
coefficients are predicted with a source of speech, the effectiveness given in modeling sequential data. Using
RNN for acoustic model training to take the time sequence of audio features as input. At each time step,
including each of the 33 Phoneset in the Afaan Oromoo language hudha (apostrophe) was removed as
punctuations in preprocessed.
The output of the RNN at each time step is a vector of probabilities the entries five short vowels, twenty for

29
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

consonant and seven double phoneme (Qubee dacha) where the entry total phoneme in Afaan oromoo language
in numbers 33.The characters (Phone) are mapped to indices in the estimate numerical form (binary) and look at
the char mapping in the files.
The predictable of RNN model is that they are only able to make use of previous context. In acoustic model,
where whole utterances are transcribed at once, predict the feature context. The BLSTM-based on RNN, to
complete the bidirectional rnn model function in sample models and specifying the bidirectional covering. The
train the deep learning used BLSTM based RNN as specified in input to softmax. Then, the model was finished
training, saved in the path for training duration model and training acoustic model to visualize binary formats.

5.2 Conclusion
Text-to-speech synthesis (TTS) which means input texts is generate to the audio from text. In Afaan Oromoo
language mostly required to text to speech synthesis for development of this language. So as to transmit
information. In this work, a first attempt is investigated speech synthesis for Afaan Oromoo language using deep
learning approach based on BLSTM- based on RNN model. The filename is long, but it contains some very
important information in activation functions used in hidden layers type is TANH, hidden layer size was 1024
and number of hidden layers were five in numbers. The dimensionality of hidden layers is 1024 number of input
nodes (the dimensionality of labels phonemes in model) was 416 and number of output nodes (the
dimensionality of acoustic features to predict) is five .Datasets for train file number seven hundred valid files
three and tests file three. The buffer size of each block of data to buffer size is 200000 the model file name feed
forward 6tanh and the learning rate used was 0.002. From this automatically created in python training and the
feature extracted prepared using python scripts.
The purpose of prepared the python feature for extracted duration model and acoustic model preprocessed
converted into binary formatted. The dimension of vector was created from parameter generated for visualization
utterances length for linguistic features and acoustic feature training and test.
From duration and acoustic plotted visual utterances generated total utterance number and frames. From this
utterance plotted we prepared normalization in X max, X min, and Y mean. The Y variable and Y scale for both
duration and acoustic train utterance length(utt_lengeth).We Used to the pytorch dataset for generated acoustic
model in waveforms and using Recurrent Neural Network within bidirectional activation true in LSTM within
vector(416,256) the model for duration and model for acoustic was generated.
At the end from label state alignment and label phone alignment used sample generated speech and user can
download the generated speech from Jupyter notebook on python scripting. It can be used as message readers,
teaching assistants, tools to aid in communication and learning for the handicapped and impaired challenged
people. During developing Afaan Oromoo speech synthesis, the system involved collecting text, preprocessing
the text, preparing phonetically balanced sentences, recording the sentences, preparing annotated speech
database, and design a prototype. In training first the text and speech corpus are manually prepared for
processing. Mel-cepstral coefficients parameters are obtained from speech data sources used Mel-cepstral
analysis like pyworld and pysptk. Then using automatic htk toolkit and front end used to festival toolkit context
labeler in state level alignment and label phone alignment. The text corpus and speech parameters are align to
generate linguistic (utterance) feature. However, every feature of Afaan Oromoo language was not considered
because it needs a lot of time and deep linguistic way of creation of Afaan Oromoo phonemes are considered.
The Mean opinion score evaluation technique was used to subjective test the performance of synthesized speech
from the evaluator of native speech to listen the audio recorded within their test. The thirteen sentences used for
testing used for subjective test, the result is 3.76 and 3.77 out of 5 score in terms of naturalness and intelligibility
respectively.

5.3 Contribution of This Study

Development of audio recording tools that facilitate the audio recording process by providing the text
transcription to be recorded with its aligned audio file name
We have developed a phoneme map for generated the linguistic features for Afaan Oromoo language.
We have prepared own corpus which can be used for text to speech synthesis to create sound with good
naturalness and intelligibility
We have addressed the issues of duration and acoustic modeling using BLSTM based on RNN for
Afaan Oromoo.

5.4 Recommendation and Future Works

In this study, Afaan Oromoo deep learning based on BLSTM- based on RNN speech synthesizer was developed.
The bandmat a periodicities form package used independency bandmat. The bandmat kind of feature extracted
from the audio, and one file for every audio file from the prepared text pair speech corpus.
Label files (time-to-phone alignments) these files are the label files which contain alignments for phoneme

30
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

alignments or state alignments audio files from the datasets.

The log-fundamental frequencies (lf0) is the log-fundamental frequency files feature file extracted from our
audio files in datasets.
Mel-generalized cepstral coefficients (mgc) these files contain the generalized cepstral coefficients for our
audio files in our datasets and move next feature file type. In these files contain the generalized cepstral
coefficients for our audio files in our data set and script file for filenames to create the acoustic model and
duration model.
The test speech synthesis the sound generated (utterance and prompt utterance, train sentences, file
identification list). The deep learning model used to Bidirectional LSTM-based RNNs. Using pytorch, simple to
demonistrated.The BLSTM is undestanded the previous input data or gate data used and RNN predicts output
feature sequence given an input feature sequence. The Training acoustic model and duration model generated the
audio from text converted to label (time to phoneme each sentence) using on python command line or jupyter
notebook within Control Process Unit (CPU) device in torch package configuration.
In the future will be use deep neural network model (DNN) hybrid within HMM techniques statistical
speech synthesis parameters unit selection synthesis speech.
In the future, we will try to investigate deep bidirectional long short-term memory recurrent neural network
(DBLSTM-RNN) with an even details structure with a larger corpus and considering all the abbreviated words,
numbers, to provide a high quality of speech synthesized for language. Another task that needs future work is
effectively acoustic feature extraction used to WaveRNN model within Tacotron speech synthesis considered in
developing within NVIDIA Nsight HUD Launcher 5.4 driver (CUDA toolkit) for fast speed processing large
datasets.

References
A. Balyan, S. S. (2013). Speech Synthesis. A Review," International Journal of Engineering Research &
Technology (IJERT), pp. 57-75.
Alem F., K. N. (2007). “Text To Speech for Bangla Language using Festival” . BRSC University, Bangladesh.
Allen Jonathan, H. M. (1987). From Text to Speech: The MITalk system. Cambridge University Press.
Alula. (2010). A generalized approach for Amharic Text To Speech system. Addis Abeba.
Barnwell, A. V. (1995, Jul ). A mixed excitation LPC vocoder model for low bit rate speech coding. in IEEE
Transactions on Speech and Audio Processing, vol. 3, pp. 242-250.
Bluche, T. N. (2013). Tandem HMM with Convolutional Neural Network for Handwritten Word Recognition. In
Proceedings of the IEEE International Conference on Acoustics Speech and Signal Processing
(ICASSP2013),. Vancouver, BC, Canada.
Cassia Valentin. (2013). Intelligibility Enhancement Of Synthetic Speech In Noise. Ph. D. Dissertation.
University Of Edinburgh, Germany.
Charpentier, M. a. (1990). Pitch-synchronous waveform processing techniques for text-to-speech synthesis using
diphones. Speech Communication.
Christian Kratzenstein. (1779). the Danish scientist working at the Russian Academy of Sciences, built the first
talking machine. Danish, Russian Academy.
Dutoit, T. (1997). A Short Introduction to Text-to-Speech. Dordrecht, Boston, London.: Kluwer Academic
Publisher.
Figueiredo, A. I. (2006). Automatically Estimating the Input Parameters of Formant-Based Speech Synthesizers.
Flanagan, J. (1965). Speech analysis, synthesis, and perception. Springer, Berlin.
Goubanova, O. T. (2000 ). Using Bayesian belief networks for model duration in text-to -speech systems. In
Proc. of ICSLP ICSLP-20 00, Beijing, China.
Graves A. Mohamed, A. G. (2013). Speech recognition with deep recurrent neural networks.In Proceedings of
the IEEE International Conference on Acoustics, Speech and Signal Processing, (Vol. 38th). Vancouver,
BC, Canada,.
Graves, A. (2012.). Supervised Sequence Labelling with Recurrent Neural Networks;. Springer: Berlin,
Germany,.
Graves, A., Jaitly, N., & Mohamed, A. (8–12 December 2013;). Hybrid speech recognition with Deep
Bidirectional LSTM. In Proceedings of the 2013 IEEEWorkshop on Automatic Speech Recognition and
Understanding,. Olomouc, Czech Republic,.
Holmes, J. H. (2003). “Speech Synthesis and Recognition” e,. Taylor and Francies New Fetter Lan, London
ECAP 4EE.
I. H. Witten, E. F. (2012). Practical Machine Learning Tools and Techniques of Data Mining .
Javidan, R. (2010). Concatenative Synthesis ofPersian Language Based on Word, Diphone and Triphone Data-
bases. Persian.
Klatt, D. (1976). Linguistic uses of segmental duration in English: Acoustic and perceptual evidence. Journal of

31
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

the Acoustic Society of America.

Klatt, D. (1980,). Software for a cascade/parallel formant synthesizer. J. Acoust. Soc. Am., 67, 971–995.
Kubichek, R. (1993). Mel-cepstral distance measure for objective speech quality assesment. In Proceedings of
IEEE Pacific Rim Conference on Communications, vol. volume 1, pp. pages 125–128.
Lazaridis, A. Z. (, 2007). Segmental duration modeling for Greek speech synthesis. In Proc.of IEEE ICTAI
ICTAI-2007, Patras, Greece,.
Lemmetty, S. (1999). Review of speech synthesis technology. Helsinki University of Technology. From
http://www.acoustics.hut.fi/~slemmett/dippa/%5Cnhttp://www.acoustics.hut.fi/public
Lloret, M. (1997). “Oromo Phonology” ,Phonologies of Asia and Africa . Winona Lake, Ind. Eisenbrauns.
M. S. Al-Radhi, T. G. (2017). "Deep Recurrent Neural Networks in Speech Synthesis Using a Continuous
Vocoder. n International Conference on Speech and Computer.
Melba, G. (1988). Introduction of Oromo people and Oromia.
Morise.M, F. Y. (2016). WORLD A vocoder-based high quality speech synthesis system for real-time
applications in IEICE Transactions on Information and Systems.
Morka, H. (20013). Afaan Oromo TTS system. Addis Abeba.
Morka.M. (2003). Text-To-Speech System for Afan Oromo Masters of Thesis. Addis Ababa University.
Moulines, E., & Charpentier, F. (1990,). Pitch-synchronous waveform processing techniques for text-to-speech
synthesis using diphone. Speech Commun.
Naslund, P. (2018). Artificial Neural Networks in Swedish Speech Synthesis Master in Computer Science.
Ofgaa, S. T. (,May, 2011). Concatenative Text-To-Speech System for Afaan Oromo Language .
Rabiner L.R and Juang, B. (1993). Fundamentals of Speech Recognition. Englewood Cliff. New Jersey: Prentice
Hall, Inc.
Rashad, M. E.-B. (2010). An overview of text-tospeech synthesis techniques.
Rodman, R. D. (1999). Computer Speech Technology. A. rtech House, Inc., London.
Samson, T. O. (2011). Concatenative Text-To-Speech System for Afaan Oromo Language. Addis Ababa
Universisty, Ethiopia.
Samuel, T. (2007). “Natural Sounding Text-ToSpeech Synthesis, based on Syllable-Like Units”, Master of
Science . India.
Sangramsing N. Kayte, D. G. ( 2015). The Marathi Text-To-Speech Synthesizer Based On Artificial Neural
Networks. International Research Journal of Engineering and Technology (IRJET), Volume: 02 Issue,.
From www.irjet.net
Solomon, T. (2005). Automatic Speech Recognition for Amharic. Doctoral Dissertation. Hamburg
University,Germany.
Sproat, R. B. (2001). “Normalization of Non-standard Words, Computer Speech and Language” (Vols. Vol. 15,).
Sutskever, I. V. (2014). Sequence to sequence learning with neural networks. In Proceedings of the Annual
Conference on Neural Information Processing Systems, ,. Montreal, QB, Canada.
Sutskever, I., Vinyals, O., & Le, Q. (2014). Sequence-to-Sequence Neural Network Models for Grapheme-to-
Phoneme Conversion. ,Lake Tahoe, NV, USA.
Takeda, K. S. (1989 .). On sentence sentence-level factors governing segmental duration in Japanese. Journal of
the Acoustical Society of America.
Tesfaye. (2004). Diphone Based Text To Speech System for Tigrigna Language. Addis Abeba.
Tewodros. ( 2009). Text To Speech synthesizer for Wolaytta language . Addi Abeba.
Tilahun, G. (1993). Qubee Afaan Oromoo: Reasons for Choosing the Latin Script for Developing an Oromo
Alphabet t.The Journal of OromoStudies.
Van Santen, J. (1992). Contextual effects on vowel durations. Speech Communication.
Wosho, K. M. (2020). Text to Speech Synthesizer for Afaan Oromoo using Statistical Parametric Speech
Synthesis. Addis Ababa, Ethiopia.
Xu, S. (2007). Study on HMM-Based Chinese Speech Synthesis; Beijing University of Posts and
Telecommunications:. Beijing, China.
Y. Fan, Y. Q. (2014). TTS Synthesis with Bidirectional LSTM. Asia, Beijing, China.
Yamagishi et al. ( 2004). speaker-independent and speaker adaptability which give special emphasis on voice
characteristics such as speaker individualities, speaking styles, and emotions.
Yang, J. ( 2014). Deep learning theory and its application in speech recognition.Commun. Countermeas.
Yao, K., & Zweig, G. (2015). Sequence-to-Sequence Neural Network Models for Grapheme-to-Phoneme
Conversion. In Proceedings of the Annual Conference of the International Speech Communication
Association, Dresden,. Germany.
Yoshimura, T. (2002). Simultaneous modeling of phonetic and prosodic parameters, and characteristic
conversion for HMM-based Text-To-Speech systems. Nagoya Institute of Technology: PhD dissertation.
Yoshimura, T. T. (1999). Simultaneous modeling of spectrum, pitch and duration in HMM-based speech

32
New Media and Mass Communication www.iiste.org
ISSN 2224-3267 (Paper) ISSN 2224-3275 (Online)
Vol.101, 2022

synthesis. In Proceedings of the Sixth European Conference on Speech Communication and Technology
(EUROSPEECH’99). Budapest, Hungary,.
Zen, H. S. (2013). Deep learning method speech sysnthesis. Vancouver, BC, Canada.
Zen, H. T., & Alan, W. (2009). Statistical parametric speech synthesis. Speech Commun.

33