The paper investigates whether audio‑language models capture paralinguistic cues beyond spoken content. Using the Expresso dataset and four open‑source models, the authors trace how speaking style information is encoded in the late layers of the audio encoder but is degraded before reaching the final output. They find that some models are content‑driven while others are acoustic‑driven, revealing a gap between what is encoded and what is utilized in current audio‑language models.
arXiv:2609.05871v1 Announce Type: cross Abstract: Audio-conditioned language models often underuse acoustic cues such as prosody, emotion, and non-speech sounds, raising the question of whether ASR-s...
arXiv:2610.03125v1 Announce Type: cross Abstract: Speech delivery varies with both the requested paralinguistic attribute and the linguistic content. We introduce ParaGeo, a matched-content decomposi...
AudioTETRIS is a training‑free, two‑stage token elimination framework designed for large audio language models. It leverages the geometry of audio encoder representations as an acoustic prior to compute per‑token importance scores, which guide progressive top‑k token selection across decoder layers. Experiments on Qwen2‑Audio‑7B, Phi‑4‑Multimodal, and Audio Flamingo 3 show that AudioTETRIS retains near‑full performance while using only 25% of audio tokens and outperforms existing training‑free baselines.
The paper introduces an Encoding Probe that reconstructs language model representations using interpretable features, addressing limitations of traditional decoding probes such as incomparable feature contributions and correlation effects. It evaluates this approach on text and speech transformer models, examining features from acoustics, phonetics, syntax, lexicon, and speaker identity. Findings reveal that speaker-related effects vary with training objectives and datasets, while syntactic and lexical features independently contribute to reconstruction, offering a complementary perspective on model interpretation.
arXiv:2610.23745v1 Announce Type: cross Abstract: Recent large audio language models (LALMs) have extended their capabilities beyond lexical speech understanding to tasks requiring paralinguistic and...
arXiv:2607. 03806v1 Announce Type: cross Abstract: Audio foundation models are widely adopted as general-purpose feature extractors, yet the internal structure of their learned representations remains insufficiently understood.
arXiv:2606. 10368v1 Announce Type: cross Abstract: Speech-to-text (S2T) systems for recognition (ASR) and translation (S2TT) typically generate discrete text tokens.
FireRedAudio is a 9‑billion‑parameter audio language model that separates continuous input representations for audio understanding and speech generation, enabling a single autoregressive LLM to perform tasks such as ASR, zero‑shot TTS, Instruct TTS, and semantic/acoustic speech editing. The model uses a dedicated Audio Encoder for recognition and a RedAE‑based pathway for generation, with the LLM directly generating text or conditioning a flow‑matching DiT to produce acoustic latents. Evaluations show competitive or leading performance in multilingual ASR, content‑accurate zero‑shot TTS, strong instruction following, and significant improvements in speech editing over prior work.
arXiv:2609.15313v1 Announce Type: cross Abstract: Autoregressive generation of interleaved text and acoustic tokens is a common approach to spoken-response generation in speech large language models....
arXiv:2609.38887v1 Announce Type: cross Abstract: Real-time voice conversion (VC) systems commonly rely on pretrained speaker embeddings from automatic speaker verification (ASV) models. While effect...
The paper investigates where speaker timbre information is located within speech signals, examining frequency bands, phase/magnitude structure, source/filter decomposition, voiced/unvoiced regions, SSL layers, and codec tokens across Mandarin, English, and Japanese data. It finds that 1–8 kHz magnitude and vocal‑tract spectral‑envelope features dominate perception, while above‑8 kHz bands remain machine‑discriminative; shallow SSL layers reveal compact speaker cues, and token leakage varies with tokenizer and layer choice. These insights guide where generative speech systems should condition on or transform speaker information and where to audit residual leakage.
Large Audio-Language Models (LALMs) have been widely used as judge models for the automatic evaluation of generated speech. However, prior approaches predominantly focus on holistic naturalness, leaving fine-grained paralinguistic distinctions underexplored.