Skip to content

NLP and Speech Processing: from knowledge graphs to speech synthesis and embedding security

    Advances in Natural Language and Speech Processing

    Recent research within the Romanian AI Hub focuses on improving recommendation systems, the security of language models, and advancing speech synthesis technologies for the Romanian language.

    Knowledge Graph-based Recommender Systems (TRL 3)

    A recipe recommendation system is proposed, using a hybrid architecture to overcome the limitations of Transformer-only models (high computational cost, low explainability).

    • Retrieval-Augmented Generation (RAG + KG): The LLM queries the knowledge graph for relevant facts, injecting retrieved triples or subgraphs directly into the prompt to generate grounded responses.
    • Graph-Enhanced Training: Includes graph embeddings or triples in the fine-tuning stage to enable the model to explicitly learn relationships.
    • Graph-Enhanced Training: Includes graph embeddings or triples in the fine-tuning stage to enable the model to explicitly learn relationships.

    Embedding Inversion (TRL 3)

    This proof-of-concept technique aims to reconstruct the original text from the numerical representations (embeddings) of encoder-type models..

    • Performance: The inversion technique surpasses current state-of-the-art (SOTA) on multiple datasets and encoder models.
    • Efficiency: The reconstruction method is optimized at the token level, reducing generation costs without sacrificing retrieval accuracy.
    • Critical Applications: Security analysis (recovering sentences from vectors), testing vulnerabilities of vector databases, and interpreting linguistic features (semantic vs. syntactic) encoded in embeddings.

    Advances in Speech Signal Processing

    The main focus is the development and security of audio technologies specific to the Romanian linguistic context..

    • Romanian TTS: Development of an advanced Text-to-Speech (TTS) system for the Romanian language, already integrated into two distinct software applications.
    • Deep Fake Detection: Implementation of automatic detection for artificially synthesized audio content.
    • Anonymization and Emotion: Technologies for anonymizing Romanian speakers and automatically recognizing emotions directly from vocal data.