This paper analyzes speech-text interleaved language models and reveals that they go through an implicit transcription phase in which spoken words become decodable as text in intermediate layers.

CORTIS is a text-only adaptation framework that fine-tunes spoken language models for task-oriented voice agents using text-form task supervision.

The paper proposes LaSR, a context-aware training paradigm that uses latent reasoning to significantly improve speech recognition, especially for specialized terminology, without adding latency.

This paper proposes Efficient Chain-of-Modality Reasoning (ECoM Reasoning), a framework to improve reasoning ability in spoken language models (SLMs) for mathematical question answering tasks by compr…

The paper introduces three new Korean speech benchmarks (KVoiceBench, KOpenAudioBench, and KMMAU) to evaluate SpeechLMs, demonstrating that English-centric evaluation fails to capture performance gaps…

This paper identifies the root cause of performance degradation in full-duplex Spoken Language Models (SLMs) due to modality interference and proposes Lychee-FD, a framework that decouples conflicting…

MEUSLI introduces an open-source multilingual projector family linking a Whisper encoder with LLMs, enabling end-to-end ASR in 28 European languages and supporting multilingual speech translation and…

The paper introduces SPEARBench, a benchmark for evaluating naturalness in speech-to-speech language models using a multidimensional protocol.

PolySpeech-100 introduces a massive, multi-lingual benchmark covering 110 linguistic variants to rigorously test Speech-LLMs, demonstrating that open-source models struggle with low-resource languages…

The paper introduces SURE, a unified framework designed to standardize and improve the comparability and reproducibility of evaluations for advanced speech understanding models.

This paper investigates the relationship between language model perplexity and word error rate in modern automatic speech recognition systems, studying the impact of external language models, encoder…

SALSA is a lightweight adaptation method that learns layer-wise steering vectors to significantly improve the performance of speech-aware LLMs on out-of-domain speech tasks.

The paper introduces Dialogs, a new Russian conversational speech corpus with high-quality recordings, segmented utterances, and expressive prosody labels.

The paper introduces GigaSpeechBench, a comprehensive multilingual and multidimensional ASR & AST benchmark with 680 hours of human-annotated speech, featuring 12 low-resource languages, 6 Chinese dia…

The paper introduces BEA-Dialogue+, an expanded 200-hour corpus for Hungarian conversational ASR, demonstrating that while larger data is challenging, specialized fine-tuning techniques significantly…

MOSS-Audio is a unified audio-language model designed for comprehensive understanding of speech, environmental sounds, and music, achieving strong performance across various audio-grounded tasks.

The paper proposes SpeechKV, a method to compress speech sequences inside large language models using a learned pooling, maintaining performance and delivering decoding speedup.

This paper explores the effect of conversational timing properties on automatic speech recognition (ASR) systems by controlling and optimizing pause and overlap timing distributions.

The authors present ZONOS2 8B, a TTS model with improved naturalness, prosody, and voice cloning fidelity, achieved through scaling, data expansion, and simplification.

The paper argues that large language models (LLMs) are a special case of world models and proposes a continuous spectrum between token prediction and latent-space architectures.