LLaSO - Logical Intelligence's open-source speech model

What is LLaSO?

LLaSO (Large Language and Speech Model) is the world's first fully open-source speech model, launched by Beijing DeepLogic Intelligence Technology Co., Ltd. It addresses long-standing issues in the field of Large Language and Speech Models (LSLMs), such as fragmented architecture, data privatization, limited task coverage, and a single interaction modality. LLaSO comprises three core components: LLaSO-Align (a large-scale speech-text alignment dataset), LLaSO-Instruct (a multi-task instruction fine-tuning dataset), and LLaSO-Eval (a standardized evaluation benchmark). It provides a unified, transparent, and reproducible infrastructure for LSLM research, driving the field's shift from "individual efforts" to "collaborative innovation."

Main functions of LLaSO

• Dataset providedLLaSO-Align provides a large-scale speech-text alignment dataset, while LLaSO-Instruct provides a multi-task instruction fine-tuning dataset, offering rich data resources for model training.
• Model training and validationThe LLaSO-Base model, trained on the LLaSO dataset, provides researchers with a performance benchmark, facilitating the comparison and verification of the performance of different models.
• Standardized assessmentLLaSO-Eval provides a standardized evaluation benchmark, ensuring the fairness and reproducibility of model evaluation.
• Multimodal supportIt supports multiple modalities such as "text command + audio input", "audio command + text input" and pure audio interaction, expanding the application scenarios of the model.

LLaSO's technical principles

• Voice-Text AlignmentAutomatic speech recognition (ASR) technology is used to precisely align speech data with text data, establishing a mapping relationship between speech representation and text semantic space.
• Multitasking instruction fine-tuningThe model is fine-tuned using data from various tasks, covering linguistic, semantic, and paralinguistic tasks, to improve the model's comprehensive understanding and generative capabilities.
• Modal projectionTechniques such as Multilayer Perceptron (MLP) are used to achieve spatial mapping between speech features and text features, enabling the model to handle multimodal inputs.
• Two-stage training strategyFirst, perform speech-text alignment training, then fine-tune the multi-task instructions to gradually improve the model's performance and generalization ability.
• Standardized evaluation benchmarksBy designing evaluation benchmarks that cover multiple tasks, the model is comprehensively and systematically evaluated to ensure the objectivity and comparability of the evaluation results.

LLaSO's project address

• GitHub repositoryhttps://github.com/EIT-NLP/LLaSO
• HuggingFace model libraryhttps://huggingface.co/papers/2508.15418
• arXiv technical paper: https://arxiv.org/pdf/2508.15418v1

Application scenarios of LLaSO

• Intelligent voice assistantIt is used in the development of intelligent voice assistants, such as smart home control, intelligent customer service, and in-vehicle voice assistants, to enable device control and information retrieval through voice commands, thereby improving the user experience.
• Voice content creationGenerate audio content, such as audiobooks, podcasts, and audio advertisements, by generating natural and fluent audio based on text content, thereby improving content creation efficiency.
• Education and LearningIt provides learners with a personalized learning experience and improves learning outcomes by enabling pronunciation practice and oral assessment through voice commands.
• HealthcareIt assists doctors in recording and diagnosing speech, helps patients with speech rehabilitation training, and improves medical efficiency and patient rehabilitation outcomes.
• Intelligent Customer ServiceProvide customer support through voice interaction, understand customer questions and generate accurate answers, thereby improving service efficiency and customer satisfaction.