New Open Dataset Brings Large-Scale Speech Data to Finnish and Swedish TTS Systems

Text-to-speech technology has historically favored languages with abundant training data, leaving many languages underserved. A new dataset aims to narrow that gap for Nordic speakers by making high-quality speech corpora freely available.

What is Nord-Parl-TTS?#

Researchers have created Nord-Parl-TTS, an open-source collection of speech recordings designed specifically for training text-to-speech systems in Finnish and Swedish. The dataset draws from audio recordings of Nordic parliamentary sessions, making use of naturally spoken material that was already in the public domain.

The effort yielded substantial quantities of training material: approximately 900 hours of Finnish speech and roughly 5,090 hours of Swedish speech. This large volume of data provides a foundation for developing and refining TTS models tailored to these languages.

How the Dataset was Built#

The researchers adapted the Emilia data processing pipeline to extract and prepare speech suitable for TTS training from the raw parliamentary recordings. Rather than collecting new audio, the team worked with existing public material, which allowed them to create the resource efficiently while avoiding privacy complications.

In addition to the raw speech data, the dataset includes unified evaluation sets. These benchmark materials enable researchers to measure progress and compare different TTS models on standardized tasks—a critical resource for coordinating development efforts across the research community.

Why It Matters#

High-resource languages like English, Mandarin, and Spanish have access to vast amounts of openly available speech data for training advanced TTS systems. This creates a significant disparity: speakers of languages with smaller digital footprints often lack the training resources needed to build competitive voice synthesis technology.

By releasing Nord-Parl-TTS publicly, the researchers help address this imbalance for Finnish and Swedish-speaking communities. The open-access model means any researcher or organization can use the data to advance TTS technology without licensing restrictions—democratizing access to tools that were previously difficult to develop for these languages.