Voice Synthesis Improvement by Machine Learning of Natural Prosody.

Area of Science:

• Computer Science
• Artificial Intelligence
• Speech Synthesis

Background:

• Current human-computer interfaces (HCI) strive for seamless interaction.
• Computer-generated speech often lacks natural prosody (intonation and rhythm), sounding non-human.
• Improving synthesized voice realism is crucial for applications like electronic readers and assistive speaking devices.

Purpose of the Study:

• To enhance computer-generated text-to-speech (TTS) algorithms by incorporating human-like melodic and prosodic elements.
• To increase the realism of synthesized voices.
• To improve electronic reading applications and artificial voices for individuals needing speech assistance.

Main Methods:

• Exploration of a novel approach using machine learning, specifically a Long Short-Term Memory (LSTM) neural network.
• Implementation of an LSTM-based encoder to add paralinguistic elements to recorded or generated voices.
• Deployment of a prototype modular platform for digital speech improvement through laboratory experiments.

Main Results:

• The LSTM-based encoder demonstrated the ability to produce realistic speech.
• Laboratory experiments provided encouraging results for the developed speech improvement algorithms.
• Optimization of algorithm combinations and performance in edge cases was explored.

Conclusions:

• The novel LSTM-based approach shows significant promise for generating more natural and realistic computer-generated speech.
• Further research will focus on algorithm optimization and comparative performance analysis against existing methods.
• The developed technology has potential applications in improving audio codecs, restoring old recordings, and enhancing HCI.