Academia.eduAcademia.edu

Figure 3 – uploaded by mezgebe araya

See full PDFdownloadDownload figure
Fig 3. Performance of LSTM in different Epoch number

Figure 3 Performance of LSTM in different Epoch number

Related Figures (12)

Fig 1. Proposed work Architecture This study describes the TTS system for the Tigrigna linguistic, using speech analysis architecture. When we use concatenative speech synthesis where the segments of recorded speech are concatenated to produce the desired output.
Fig 1. Proposed work Architecture This study describes the TTS system for the Tigrigna linguistic, using speech analysis architecture. When we use concatenative speech synthesis where the segments of recorded speech are concatenated to produce the desired output.
Fig 2. Long short-term memory (LSTM) structure One kind of RNN.The four important layers in LSTM models are point-wise operation, vector transfer, neural network layer, concatenate and copy. the three gates are computed as follow 3.2.1. LSTM 3.2 Models
Fig 2. Long short-term memory (LSTM) structure One kind of RNN.The four important layers in LSTM models are point-wise operation, vector transfer, neural network layer, concatenate and copy. the three gates are computed as follow 3.2.1. LSTM 3.2 Models
Table 1. Scales used in MOSMOS Then the evaluators provide their ranks based on the MOS scale as shown in [Table 1]. To evaluate the synthesizers’ intelligibility and naturalness ten sentences are prepared as a test data for the synthesizer. All words used in the sentence are found in the compiled lexicon. Then the selected individuals listen to the synthesized waveform from the synthesizer and evaluate naturalness and intelligibility based on the MOS scale. The invited native Tigrigna speakers are given with the questionnaire to evaluate intelligibility and naturalness of the synthesized speech.
Table 1. Scales used in MOSMOS Then the evaluators provide their ranks based on the MOS scale as shown in [Table 1]. To evaluate the synthesizers’ intelligibility and naturalness ten sentences are prepared as a test data for the synthesizer. All words used in the sentence are found in the compiled lexicon. Then the selected individuals listen to the synthesized waveform from the synthesizer and evaluate naturalness and intelligibility based on the MOS scale. The invited native Tigrigna speakers are given with the questionnaire to evaluate intelligibility and naturalness of the synthesized speech.
Table 2. Intelligibility (MOS) Scores of Tigrigna Speech Synthesizer Where P is the persons who are invited to evaluate the intelligibility of text to speech synthesizer for Tigrigna language.
Table 2. Intelligibility (MOS) Scores of Tigrigna Speech Synthesizer Where P is the persons who are invited to evaluate the intelligibility of text to speech synthesizer for Tigrigna language.
Where P is the persons who are invited to evaluate the naturalness of text to speech synthesizer for Tigrigna language.
Where P is the persons who are invited to evaluate the naturalness of text to speech synthesizer for Tigrigna language.
Table 4. Average MOS Scores of Tigrigna Speech Synthesizer
Table 4. Average MOS Scores of Tigrigna Speech Synthesizer
The overall intelligibility of the system from twenty listeners for the ten Tigrigna sentences is found to be 3.27. Which means the synthesizer is ‘good’ as per the scale of the MOS test. The overall naturalness of the synthesizer found to be 3.28 which also approach to ‘good’ MOS scale. These values of intelligibility and naturalness look encouraging to come up with a better system. Though we tried to record the corpus data in a quiet room, there was also some noises from the computer itself listened during the synthesis of the data which degrades the naturalness of the sound for the listeners.
The overall intelligibility of the system from twenty listeners for the ten Tigrigna sentences is found to be 3.27. Which means the synthesizer is ‘good’ as per the scale of the MOS test. The overall naturalness of the synthesizer found to be 3.28 which also approach to ‘good’ MOS scale. These values of intelligibility and naturalness look encouraging to come up with a better system. Though we tried to record the corpus data in a quiet room, there was also some noises from the computer itself listened during the synthesis of the data which degrades the naturalness of the sound for the listeners.
Fig 5. Formants of the word ”arba” Fig 6. The LPC feature extraction of the word ”arba”
Fig 5. Formants of the word ”arba” Fig 6. The LPC feature extraction of the word ”arba”
Fig 7. The original signal of the word ”arba”
Fig 7. The original signal of the word ”arba”
Fig 8. The original signal of the word ”arba” Figure 8. Segmentation of Recorded Speech using PRAAT Fig 9. GUI Text-to-Speech Synthesizer for Tigrigna language
Fig 8. The original signal of the word ”arba” Figure 8. Segmentation of Recorded Speech using PRAAT Fig 9. GUI Text-to-Speech Synthesizer for Tigrigna language
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved