The field of speech and language processing is witnessing significant advancements, driven by innovations in large language models (LLMs) and their applications in various tasks. A notable trend is the development of efficient methods for processing long-form audio and text, which has been a longstanding challenge. Researchers are proposing novel approaches to adapt pre-trained models for tasks such as music structure analysis, speech summarization, and dialogue systems. These advancements have the potential to improve the performance and efficiency of speech and language processing systems, enabling them to handle complex and dynamic inputs. Noteworthy papers in this area include LoopServe, which introduces an adaptive dual-phase inference acceleration framework for LLMs in multi-turn dialogues, and LaCache, which proposes a ladder-shaped KV caching paradigm for efficient long-context modeling. Additionally, papers like FastLongSpeech and TalkLess demonstrate innovative approaches to speech processing and editing, highlighting the potential for significant improvements in this field.
Advancements in Speech and Language Processing
Sources
Temporal Adaptation of Pre-trained Foundation Models for Music Structure Analysis
LoopServe: An Adaptive Dual-phase LLM Inference Acceleration System for Multi-Turn Dialogues
Political Leaning and Politicalness Classification of Texts
LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models
Identifying Algorithmic and Domain-Specific Bias in Parliamentary Debate Summarisation
FastLongSpeech: Enhancing Large Speech-Language Models for Efficient Long-Speech Processing
TalkLess: Blending Extractive and Abstractive Speech Summarization for Editing Speech to Preserve Content and Style
On the Inevitability of Left-Leaning Political Bias in Aligned Language Models
Leveraging Context for Multimodal Fallacy Classification in Political Debates
Left Leaning Models: AI Assumptions on Economic Policy
Nonlinear Framework for Speech Bandwidth Extension
Who Attacks, and Why? Using LLMs to Identify Negative Campaigning in 18M Tweets across 19 Countries