Stay ahead with advanced broadcasting solutions

High-performance RRS implementation

Ensuring reliability and scalability through thorough testing and feature delivery

Fluendo efficiently delivered the requested features, positioning the RRS as a potential replacement for the client’s existing solution. Robust testing ensured high reliability and performance.

Empowering Teams with Rust and GStreamer

Building a foundation for safe and high-performance multimedia development

The engineering team gained confidence in integrating Rust into their multimedia stack with GStreamer. They developed a strong foundation in safe, high-performance multimedia development, enabling them to create more robust and efficient multimedia systems internally.

Architecting an ethical and high-fidelity voice cloning platform

Delivering Professional TTS through MLOps, Transfer Learning, and C2PA Standard

To address the project’s complex requirements, the solution was divided into two distinct technological pillars: a highly optimized TTS model and a secure, enterprise-grade web application.

The Core AI Engine: Linguistic Precision and High-Fidelity Synthesis

To ensure professional naturalness, the system used Transfer Learning from existing foundational models (Proyecto AINA), which were heavily refined using over 45 hours of studio-mastered recordings. A critical differentiator was the manual linguistic correction of the phonetizer engine, which drastically reduced pronunciation errors and ensured perfect dialectal accuracy. The architecture deployed a modular TTS engine that combined eSpeak-NG, Matcha-TTS, and Vocos. This rigorous acoustic and linguistic optimization successfully reduced the Phoneme Error Rate (PER) to just 0.95%. The application and model were presented to a focus group of the audiovisual industry, and the experts rated it as the best Catalan TTS available.

The enterprise architecture: ethical governance and scalability

Surrounding the AI engine, a robust application was built to manage the complexities of voice actors’ IP management through contracts, GDPR compliance, and high-concurrency requests. Orchestrated via a backend using Django, Redis, and Celery, the entire pipeline guarantees absolute data sovereignty. Crucially, the platform addressed the legal challenges of Generative AI by embedding C2PA cryptographic metadata into every generated audio file, ensuring the synthetic media’s origins were immutable and protected against unauthorized data mining.

Supported by real-time ASR validation on the frontend, this scalable architecture achieved an internal synthesis latency of just 0.12 seconds, a Real-Time Factor (RTF) of 0.17, and a 100% success rate for backend inference requests. Ultimately, the project delivered a transparent, ethical, and high-performance solution ready for the professional dubbing industry.

Implementing trustworthy media pipelines

Securing video bitstreams through standardized metadata integration

The diagnostic process identified that traditional authentication methods relied on container-level signatures, which were easily stripped or manipulated. To address this, the solution focused on embedding cryptographic signatures directly into the video bitstream at the Network Abstraction Layer (NAL) unit level. This ensured that authentication information traveled with the video data itself, making it significantly harder to tamper with.

The implemented solution involved extending the H.266 multimedia parsers to extract SEI messages carrying DSC metadata and expose them as GstMeta. New GStreamer elements were developed to handle DSC signing and verification, allowing for the generation of signing metadata in compliance with JVET standards and the validation of content using public keys. These elements emitted signals to notify the system of validation success or failure, providing a robust framework for real-time media authentication.

The engagement resulted in a modular implementation that was contributed upstream to the GStreamer community, ensuring long-term maintainability and industry adoption. Further technical details were shared in a related blog post.