Sovereign Speech Foundation Models for High-Fidelity, Context-Aware and Controllable Catalan Speech

Short description of the innovation study

The SOV-CAT-VOICE study addresses the lack of sovereign, broadcast-quality generative voice models for minority languages like Catalan. Dominant commercial Text-to-Speech platforms focus on major languages and raise data sovereignty and copyright concerns for European media. Led by Fluendo and Idea Sonora, this project leverages 30.000 GPU accelerator hours on the MareNostrum 5 supercomputer to execute a High-Throughput Computing architecture search, benchmarking Flow Matching against Latent Diffusion models across a 3,100-hour speech corpus. Using Parameter-Efficient Fine-Tuning (LoRA), the project builds a Sovereign Style Bank of voice adapters and a production-ready inference engine integrated with C2PA provenance standards. The study validates direct compute cost savings, turnaround reduction, and commercial client acceptance, paving the way for a compliant B2B SaaS platform ('Virtual Cast') that ensures digital linguistic sovereignty.

Organisations involved: 

Main SME: Fluendo S.A.

Technology expert: Idea Sonora S.L.