INESCO Dataset: indonesian expressive speech corpus for emotion speech synthesis

Area of Science:

• Natural Language Processing (NLP)
• Speech Synthesis
• Corpus Linguistics

Background:

• High-quality speech corpora are essential for developing effective language models in NLP.
• Indonesian is a low-resourced language, limiting its digital communication capabilities.
• There is a need for expressive speech datasets to improve Indonesian speech synthesis.

Purpose of the Study:

• To develop the Indonesian Expressive Speech Corpus (INESCO) to address the gap in expressive speech data for Indonesian.
• To create a high-quality dataset for advancing text-to-speech (TTS) and other NLP research in Indonesia.
• To support digital communication, creative industries, and public services with improved speech technology.

Main Methods:

• Selected 600 sentences from diverse literary and everyday sources for emotional expression.
• Recorded 2400 audio files using four professional voice artists in a soundproof studio.
• Utilized professional recording equipment and software, with phonetic annotation and emotional validation (MOS ≥ 3.5).

Main Results:

• The final INESCO dataset comprises 2400 audio files with text transcriptions, totaling 1.33 GB and 3 hours of audio.
• The corpus includes three emotional styles: happiness, anger, and sadness, recorded by experienced theater artists.
• The dataset is validated for usability in end-to-end TTS systems.

Conclusions:

• The INESCO dataset provides comprehensive, curated data crucial for reproducible research in speech synthesis and NLP.
• This publicly accessible dataset will significantly benefit Indonesian language technology development.
• INESCO supports advancements in speech recognition, TTS, and broader natural language processing applications.