Eloquens, commonly known under the alias Er Finestra, is a text-to-speech software whose first version was released in 1993 by the Italian research center CSELT. It holds the distinction of being the first commercial speech synthesis software capable of speaking Italian, marking a significant milestone in the application of Machine learning and Neural network technologies to voice generation.
The software was built using diphone-technology, an approach designed to achieve high computational efficiency while producing intelligible speech. Although the resulting voice retained a somewhat "robotic" quality, it was notably more natural than the previous generation of voice synthesis implemented in CSELT's earlier MUSA system. This technological advancement positioned Eloquens as a pioneering tool in the field of Generative AI applied to audio.
Development and Technology
Eloquens was the first commercial product developed by CSELT in the field of voice technology, emerging from years of research in speech processing and synthesis. The diphone-based method used concatenated speech units to generate utterances, allowing for a balance between quality and processing requirements. This approach was detailed in a 1993 paper presented at the 3rd European Conference on Speech Communication and Technology, authored by CSELT researchers including Marcello Balestri, Stefano Lazzaretto, Pier Luigi Salza, and Stefano Sandri, titled "The CSELT system for Italian text-to-speech synthesis." The underlying techniques later informed advances in Sequence-to-Sequence (Seq2Seq) models and Cross-Attention mechanisms used in modern Large language model applications.
Applications in Public Services
In its early years, Eloquens was deployed in several practical applications across Italy. It was used to automatically read timetables at Italian train stations, providing passengers with real-time schedule information through synthesized voice announcements. The software also powered telephone services operated by Telecom Italia, including an automated address book service that allowed users to retrieve contact information via voice interaction. These deployments, documented in research such as the 2000 paper "Field trials of the Italian Arise train timetable system" by Paolo Baggia and colleagues, demonstrated the viability of speech synthesis in real-world, high-usage environments.
By 1997, Eloquens had evolved to serve as a building block, specifically a synthesizer module, in more complex dialogue systems. This integration marked an early instance of combining speech output with interactive voice response technology, a precursor to modern Reinforcement Learning from AI Feedback (RLAIF)-trained conversational agents.
Er Finestra and Radio DeeJay
The software gained widespread popular recognition in the summer of 2001 when it was adopted by the Italian radio station Radio DeeJay. The station used Eloquens as the voice of a character named Er Finestra, a comedic persona that became a staple of their programming. The character's popularity led to a modified version of the software being released in 2002 for Windows by a user known as DaNieLz, who distributed it under the name Er Finestra.
This updated version featured a custom user interface theme incorporating the Radio DeeJay logoaine and added new words and symbols to the internal pronunciation files. However, the core voice engine and functionality remained identical to the original Eloquens. The Er Finestra release became the most famous version of the program because it could be installed on Windows operating systems without complications, making it accessible to a broad audience.
Legacy and Modern Usage
Both Eloquens and its Er Finestra variant are now distributed as freeware, preserving their availability to enthusiasts and creators. In contemporary use, the software is typically employed to generate spoken comments in many YouTube videos and to voice the performances of fictional characters. This grassroots adoption has kept the software relevant years after its initial development, highlighting a lasting niche for lightweight Text-to-speech tools.
The development of Eloquens also contributed to the broader evolution of speech technology, influencing subsequent projects and spin-offs such as Loquendo, a company that later commercialized advanced synthesis products. The CSELT experience, as documented in publications like the 1995 paper "Interactive voice technology at work: The CSELT experience" by Roberto Billi and colleagues, provided foundational insights that informed later work in Speech synthesis and interactive systems.
Technical and Research Context
Eloquens's design principles, particularly its focus on Data Augmentation and efficient acoustic modeling, were innovative for its time. The software's ability to operate on limited hardware without sacrificing intelligibility was a key factor in its commercial viability. Researchers at CSELT also explored specialized applications, such as the automatic reverse directory service described in a 1998 IEEE paper by Luciano Nebbia, Silvia Quazza, and Pier Luigi Salza, which used speech synthesis to read telephone numbers and subscriber information aloud.
This pioneering work in Italian speech synthesis predated the widespread adoption of deep learning techniques such as Residual Network (ResNet) architectures and Batch Normalization, instead relying on established phonetic and concatenative methods. Despite being superseded by more advanced systems, Eloquens remains a notable example of early Artificial intelligence deployed in consumer-facing technology, bridging the gap between research laboratories and everyday use.