Nexdata Releases 11 Multilingual Speech Synthesis Datasets

At Interspeech 2026, recent achievements in speech AI were demonstrated. Among other topics, multilingual speech recognition, modeling natural conversation and expressive speech synthesis attracted particular attention. In accordance with the evolution of speech technology to more languages, more natural interactions and emotional expressiveness, the need for high-quality and diverse training data increases significantly.

Nexdata has released 11 multilingual natural conversation average-tone speech synthesis corpora. The datasets cover a wide range of languages, accents and emotions, offering data for training multilingual TTS, emotional speech synthesis and models of natural speech interaction.

The 11 new corpora represent multiple languages and accents, such as Vietnamese, Spanish, Arabic, Japanese, Italian, Brazilian Portuguese, German and Indian English.

The datasets include:

2 People - Vietnamese‌ Natural Conversation Average Tone Speech Synthesis Corpus

2 People -Argentine Spanish Natural Conversation Average Tone Speech Synthesis Corpus

2 People - Mexican Spanish Natural Conversation Average Tone Speech Synthesis Corpus

2 People - Peninsular Spanish Natural Conversation Average Tone Speech Synthesis Corpus

2 People - Saudi Arabian Arabic Natural Conversation Average Tone Speech Synthesis Corpus

2 People - Modern Standard Arabic Natural Conversation Average Tone Speech Synthesis Corpus

2 People - Japanese Natural Conversation Average Tone Speech Synthesis Corpus

2 People - Italian Natural Conversation Average Tone Speech Synthesis Corpus

2 People - Brazilian Portuguese Natural Conversation Average Tone Speech Synthesis Corpus

2 People - German Natural Conversation Average Tone Speech Synthesis Corpus

2 People - Indian English Natural Conversation Average Tone Speech Synthesis Corpus

Data scale: Each corpus was recorded by two professional voice actors (one male and one female).

Data specifications: 48 kHz/24-bit, mono WAV format, recorded in professional studios using professional recording equipment and software.

Data content: Spontaneous speech on assigned topics, including natural renditions of single-level and multi-level emotions, paralinguistic expressions and various interjections. The recordings represent a wide range of emotional changes and natural speech.

Annotations: Text, emotion and paralinguistic annotations were added to help train multilingual TTS, emotional speech synthesis and models of natural speech interaction, helping improve the naturalness and expressiveness of synthesized speech.

Nexdata constantly expands the number of high-quality speech training data resources for multilingual speech recognition, natural conversation and emotional speech synthesis. With increasing volumes, languages covered and accent diversity, the resources become very helpful for training large speech models and developing related algorithms. To obtain sample data or additional information, please contact your Nexdata account manager.