Catalan Voice Synthesis: The Aina Models and the Challenge of Licence Fragmentation
Major commercial systems exhibit excellent voice synthesis across nearly all languages, but not in Catalan. The Aina project's open models offer an alternative, yet a mix of licences complicates their usage.
When discussing voice synthesis, a frequently overlooked distinction must be made. Gemini, for instance, generally provides outstanding voice synthesis across languages, often being among the best in each language. However, this is not the case for Catalan. This discrepancy underscores the importance of having indigenous and open infrastructure: if the best tools regard Catalan as a secondary language, the quality of Catalan voices, assistants, and screen readers is contingent upon this.
The Aina models, developed under the public Catalan project, present an alternative. The practical issue lies in the licences. Aina models come under Apache, CC-BY-NC-4.0, or commercial licences, depending on the case, with additional variations in practice: for example, the tts-api repository operates under the Mozilla Public License. This mixture is not a trivial detail.
Licence fragmentation and incompatibility pose tangible challenges for product developers. A licence like CC-BY-NC prohibits commercial use; an Apache licence permits it; an MPL contains unique conditions regarding modified files. When a project incorporates components with differing licences, a case-by-case analysis of permissible actions is required, slowing adoption, particularly for small companies that lack a legal department.
Moreover, there exists an expectation gap. Even if Aina surpasses Google’s Catalan synthesis, its creators understandably aspire to have a competitive product. Having a proprietary language model is a boon for linguistic sovereignty, yet it must be genuinely usable: sufficiently good, well-documented, with clear and consistent usage conditions that do not force every developer to play lawyer before writing code.
Until resolved, each project aiming to incorporate Catalan voice into an application must invest time in deciphering permissible actions for each component. This time investment often dictates whether or not Catalan becomes represented.