Cryfish: On deep audio analysis with Large Language Models

Anton Mitrofanov; Sergei Novoselov; Tatiana Prisyach; Vladislav Marchevskiy; Arseniy Karelin; Nikita Khmelev; Dmitry Dutov; Stepan Malykh; Igor Agafonov; Aleksandr Nikitin; Oleg Petrov

doi:10.21437/interspeech.2025-2109

arXiv · 2508.12666

Cryfish: On deep audio analysis with Large Language Models

Anton Mitrofanov·Sergei Novoselov·Tatiana Prisyach·Vladislav Marchevskiy·Arseniy Karelin·Nikita Khmelev·Dmitry Dutov·Stepan Malykh·Igor Agafonov·Aleksandr Nikitin·Oleg Petrov

View original source ↗Read source document ↗Export metadata ↓

Abstract

The recent revolutionary progress in text-based large language models (LLMs) has contributed to the growth of interest in extending capabilities of such models to multimodal perception and understanding tasks. Hearing is an essential capability that is highly desired to be integrated into LLMs. However, effective integrating listening capabilities into LLMs is a significant challenge lying in generalizing complex auditory tasks across speech and sounds. To address these issues, we introduce Cryfish, our version of auditory-capable LLM. The model integrates WavLM audio-encoder features into Qwen2 model using a transformer-based connector. Cryfish is adapted to various auditory tasks through a specialized training strategy. We evaluate the model on the new Dynamic SUPERB Phase-2 comprehensive multitask benchmark specifically designed for auditory-capable models. The paper presents an in-depth analysis and detailed comparison of Cryfish with the publicly available models.

Explore related subjects

eess.AS

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Anton Mitrofanov, Sergei Novoselov, Tatiana Prisyach, Vladislav Marchevskiy, Arseniy Karelin, Nikita Khmelev, Dmitry Dutov, Stepan Malykh, Igor Agafonov, Aleksandr Nikitin, Oleg Petrov. 2025-08-18. Cryfish: On deep audio analysis with Large Language Models. https://doi.org/10.21437/interspeech.2025-2109

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Recovering the Zipfian Distribution in Unsupervised Term Discovery

Unsupervised term discovery involves segmenting unlabelled speech into word- or syllable-like units and clustering these into a lexicon of candidate types. True lexicons follow a Zipfian distribution, yet the dominant centre-based clustering approach -- K-means -- produces a more uniform distribution due to an inductive bias toward spherical clusters. In this paper we revisit graph-based clustering as a bottom-up alternative, where segment embeddings are connected by pairwise similarity and partitioned using the Leiden algorithm. We show that graph clustering substantially outperforms centre-based approaches (K-means, GMM, BIRCH) in both word- and syllable-level lexicon discovery across three languages, producing more Zipf-like distributions. Another bottom-up approach, agglomerative clustering with average linkage, also performs well, although it is computationally less efficient and allows for less control over the resulting distribution. Our work calls into question the dominance of centre-based clustering for term discovery, and promotes graph clustering as an attractive alternative.

eess.AS↗

arXiv2026

Preserving Speech-to-Text LLM Capabilities in Speech-to-Speech Generation

Strong speech-to-text (S2T) LLMs already provide robust speech perception and text reasoning, but adding speech-to-speech (S2S) output is challenging: fine-tuning the backbone can degrade the original S2T performance, while attaching a downstream talker reintroduces a serial text-to-speech bottleneck. We present PRIME-Speech, a frozen-backbone S2S conversion framework that trains only speech-generation modules. PRIME-Speech synchronizes a causal audio post-decoder with intermediate hidden states of the frozen backbone, so codec tokens are generated from the model's evolving reasoning trajectory rather than from completed text chunks. The post-decoder uses mixed hidden-state, text, and audio-history conditioning, and a training-time packing strategy with turn-level audio KV-cache and position reset stabilizes multi-turn spoken interaction without additional multi-turn S2S training data. Multi-token prediction further reduces the effective codec prediction rate and improves first-audio latency without modifying the reasoning path. Across speech translation, spoken QA, speech understanding, and multi-turn dialogue, PRIME-Speech preserves the S2T behavior of the frozen backbone while producing accurate, low-WER spoken responses.

eess.AS↗

arXiv2026

SRF-SVB: Style-Consistent Singing Voice Beautifying via Rectified Flow

Singing voice beautifying (SVB) aims to correct pitch and rhythm of amateur singing while enhancing vocal quality, preserving lyrics and the singer's timbre. Existing methods, however, suffer from limited generation quality and efficiency, and tend to neglect the preservation of the singer's style. We propose SRF-SVB, a style-consistent model for SVB via rectified flow, which achieves high-fidelity and efficient beautification covering pitch and rhythm correction. Furthermore, we design a context-guided masked mel-spectrogram inpainting mechanism that effectively preserves the amateur singer's style, including unique timbre and expressive patterns. Experiments on both English and Chinese test sets show that SRF-SVB outperforms baseline models in most objective and subjective metrics.

eess.AS↗