Luca A. Lanzendörfer

Frédéric Berdoz, Luca A. Lanzendörfer, Fabian Farestam et al.

Large reasoning models (LRMs) are often evaluated using metrics such as final-answer accuracy or token count. However, identical scores on these metrics can hide fundamentally different reasoning structures. To address this limitation, we introduce a scalable LRM benchmark of logic puzzles and a pipeline that converts unstructured traces into verifiable reasoning graphs of claims and dependencies. This turns reasoning into a structured, measurable object whose topology can be quantitatively analyzed. Building on this, we define a reasoning efficiency metric that quantifies how concentrated the model's logical flow is. Our analysis on open-source reasoning models shows that structural measurements separate behaviors that token count and accuracy conflate, providing a practical tool for diagnosing failure modes and comparing how reasoning scales with puzzle difficulty.

Frédéric Berdoz, Luca A. Lanzendörfer, Kaan Bayraktar et al.

Large Language Models (LLMs) are increasingly deployed across diverse applications, raising critical questions for governance, accountability, and data provenance. Understanding which training data most influenced a model's output remains a fundamental open problem. We address this challenge through training data attribution (TDA) for auto-regressive LLMs by expanding upon the inverse formulation: How would training data be affected if the model had seen the generated output during training? Our method perturbs the base model using bidirectional gradient optimization (gradient ascent and descent) on a generated text sample and measures the resulting change in loss across training samples. Our framework supports attribution at arbitrary data granularity, enabling both factual and stylistic attribution. We evaluate our method against baselines on pretrained models with known datasets, and show that it outperforms previous work on influence metrics, thereby enhancing model interpretability, an essential requirement for accountable AI systems.

Florian Grötschla, Luca A. Lanzendörfer, Marco Calzavara et al.

Image datasets serve as the foundation for machine learning models in computer vision, significantly influencing model capabilities, performance, and biases alongside architectural considerations. Therefore, understanding the composition and distribution of these datasets has become increasingly crucial. To address the need for intuitive exploration of these datasets, we propose AEye, an extensible and scalable visualization tool tailored to image datasets. AEye utilizes a contrastively trained model to embed images into semantically meaningful high-dimensional representations, facilitating data clustering and organization. To visualize the high-dimensional representations, we project them onto a two-dimensional plane and arrange images in layers so users can seamlessly navigate and explore them interactively. AEye facilitates semantic search functionalities for both text and image queries, enabling users to search for content. We open-source the codebase for AEye, and provide a simple configuration to add datasets.

Luca A. Lanzendörfer, Florian Grötschla, Emil Funke et al.

Music datasets play a crucial role in advancing research in machine learning for music. However, existing music datasets suffer from limited size, accessibility, and lack of audio resources. To address these shortcomings, we present DISCO-10M, a novel and extensive music dataset that surpasses the largest previously available music dataset by an order of magnitude. To ensure high-quality data, we implement a multi-stage filtering process. This process incorporates similarities based on textual descriptions and audio embeddings. Moreover, we provide precomputed CLAP embeddings alongside DISCO-10M, facilitating direct application on various downstream tasks. These embeddings enable efficient exploration of machine learning applications on the provided data. With DISCO-10M, we aim to democratize and facilitate new research to help advance the development of novel machine learning models for music.

Luca A. Lanzendörfer, Roger Wattenhofer

Implicit Neural Representations (INRs) have emerged as a promising method for representing diverse data modalities, including 3D shapes, images, and audio. While recent research has demonstrated successful applications of INRs in image and 3D shape compression, their potential for audio compression remains largely unexplored. Motivated by this, we present a preliminary investigation into the use of INRs for audio compression. Our study introduces Siamese SIREN, a novel approach based on the popular SIREN architecture. Our experimental results indicate that Siamese SIREN achieves superior audio reconstruction fidelity while utilizing fewer network parameters compared to previous INR architectures.

Giulia Argüello, Luca A. Lanzendörfer, Roger Wattenhofer

Cue points indicate possible temporal boundaries in a transition between two pieces of music in DJ mixing and constitute a crucial element in autonomous DJ systems as well as for live mixing. In this work, we present a novel method for automatic cue point estimation, interpreted as a computer vision object detection task. Our proposed system is based on a pre-trained object detection transformer which we fine-tune on our novel cue point dataset. Our provided dataset contains 21k manually annotated cue points from human experts as well as metronome information for nearly 5k individual tracks, making this dataset 35x larger than the previously available cue point dataset. Unlike previous methods, our approach does not require low-level musical information analysis, while demonstrating increased precision in retrieving cue point positions. Moreover, our proposed method demonstrates high adherence to phrasing, a type of high-level music structure commonly emphasized in electronic dance music. The code, model checkpoints, and dataset are made publicly available.

Hubert Siuzdak, Florian Grötschla, Luca A. Lanzendörfer

Neural audio codecs have recently gained popularity because they can represent audio signals with high fidelity at very low bitrates, making it feasible to use language modeling approaches for audio generation and understanding. Residual Vector Quantization (RVQ) has become the standard technique for neural audio compression using a cascade of VQ codebooks. This paper proposes the Multi-Scale Neural Audio Codec, a simple extension of RVQ where the quantizers can operate at different temporal resolutions. By applying a hierarchy of quantizers at variable frame rates, the codec adapts to the audio structure across multiple timescales. This leads to more efficient compression, as demonstrated by extensive objective and subjective evaluations. The code and model weights are open-sourced at https://github.com/hubertsiuzdak/snac.

Florian Grötschla, Luca Strässle, Luca A. Lanzendörfer et al.

Music recommender systems frequently utilize network-based models to capture relationships between music pieces, artists, and users. Although these relationships provide valuable insights for predictions, new music pieces or artists often face the cold-start problem due to insufficient initial information. To address this, one can extract content-based information directly from the music to enhance collaborative-filtering-based methods. While previous approaches have relied on hand-crafted audio features for this purpose, we explore the use of contrastively pretrained neural audio embedding models, which offer a richer and more nuanced representation of music. Our experiments demonstrate that neural embeddings, particularly those generated with the Contrastive Language-Audio Pretraining (CLAP) model, present a promising approach to enhancing music recommendation tasks within graph-based frameworks.

Michael Ungersböck, Florian Grötschla, Luca A. Lanzendörfer et al.

Generative models have made significant progress in synthesizing high-fidelity audio from short textual descriptions. However, editing existing audio using natural language has remained largely underexplored. Current approaches either require the complete description of the edited audio or are constrained to predefined edit instructions that lack flexibility. In this work, we introduce SAO-Instruct, a model based on Stable Audio Open capable of editing audio clips using any free-form natural language instruction. To train our model, we create a dataset of audio editing triplets (input audio, edit instruction, output audio) using Prompt-to-Prompt, DDPM inversion, and a manual editing pipeline. Although partially trained on synthetic data, our model generalizes well to real in-the-wild audio clips and unseen edit instructions. We demonstrate that SAO-Instruct achieves competitive performance on objective metrics and outperforms other audio editing approaches in a subjective listening study. To encourage future research, we release our code and model weights.

Luca A. Lanzendörfer, Frédéric Berdoz, Antonis Asonitis et al.

Recent autoregressive transformer-based speech enhancement (SE) methods have shown promising results by leveraging advanced semantic understanding and contextual modeling of speech. However, these approaches often rely on complex multi-stage pipelines and low sampling rate codecs, limiting them to narrow and task-specific speech enhancement. In this work, we introduce DAC-SE1, a simplified language model-based SE framework leveraging discrete high-resolution audio representations; DAC-SE1 preserves fine-grained acoustic details while maintaining semantic coherence. Our experiments show that DAC-SE1 surpasses state-of-the-art autoregressive SE methods on both objective perceptual metrics and in a MUSHRA human evaluation. We release our codebase and model checkpoints to support further research in scalable, unified, and high-quality speech enhancement.

Florian Grötschla, Longxiang Jiao, Luca A. Lanzendörfer et al.

We introduce Panama, an active learning framework to train parametric guitar amp models end-to-end using a combination of an LSTM model and a WaveNet-like architecture. With \model, one can create a virtual amp by recording samples that are determined through an ensemble-based active learning strategy to minimize the amount of datapoints needed (i.e., amp knob settings). Our strategy uses gradient-based optimization to maximize the disagreement among ensemble models, in order to identify the most informative datapoints. MUSHRA listening tests reveal that, with 75 datapoints, our models are able to match the perceptual quality of NAM, the leading open-source non-parametric amp modeler.

Luca A. Lanzendörfer, Florian Grötschla, Cesare Blaser et al.

Speaker diarization is the task of partitioning audio into segments according to speaker identity, answering the question of "who spoke when" in multi-speaker conversation recordings. While diarization is an essential task for many downstream applications, it remains an unsolved problem. Errors in diarization propagate to downstream systems and cause wide-ranging failures. To this end, we examine exact failure modes by evaluating five state-of-the-art diarization models, across four diarization datasets spanning multiple languages and acoustic conditions. The evaluation datasets consist of 196.6 hours of multilingual audio, including English, Mandarin, German, Japanese, and Spanish. Overall, we find that PyannoteAI achieves the best performance at 11.2% DER, while DiariZen provides a competitive open-source alternative at 13.3% DER. When analyzing failure cases, we find that the primary cause of diarization errors stem from missed speech segments followed by speaker confusion, especially in high-speaker count settings.

Benjamin Estermann, Luca A. Lanzendörfer, Yannick Niedermayr et al.

Algorithmic reasoning is a fundamental cognitive ability that plays a pivotal role in problem-solving and decision-making processes. Reinforcement Learning (RL) has demonstrated remarkable proficiency in tasks such as motor control, handling perceptual input, and managing stochastic environments. These advancements have been enabled in part by the availability of benchmarks. In this work we introduce PUZZLES, a benchmark based on Simon Tatham's Portable Puzzle Collection, aimed at fostering progress in algorithmic and logical reasoning in RL. PUZZLES contains 40 diverse logic puzzles of adjustable sizes and varying levels of complexity; many puzzles also feature a diverse set of additional configuration parameters. The 40 puzzles provide detailed information on the strengths and generalization capabilities of RL agents. Furthermore, we evaluate various RL algorithms on PUZZLES, providing baseline comparisons and demonstrating the potential for future research. All the software, including the environment, is available at https://github.com/ETH-DISCO/rlp.

Luca A. Lanzendörfer, Florian Grötschla, Uzeyir Valizada et al.

We introduce Audio Atlas, an interactive web application for visualizing audio data using text-audio embeddings. Audio Atlas is designed to facilitate the exploration and analysis of audio datasets using a contrastive embedding model and a vector database for efficient data management and semantic search. The system maps audio embeddings into a two-dimensional space and leverages DeepScatter for dynamic visualization. Designed for extensibility, Audio Atlas allows easy integration of new datasets, enabling users to better understand their audio data and identify both patterns and outliers. We open-source the codebase of Audio Atlas, and provide an initial implementation containing various audio and music datasets.

Antonis Asonitis, Luca A. Lanzendörfer, Frédéric Berdoz et al.

Automatic speech recognition (ASR) performs well for high-resource languages with abundant paired audio-transcript data, but its accuracy degrades sharply for most languages due to limited publicly available aligned data. To this end, we introduce WorldSpeech, a 24 kHz multilingual speech corpus comprising 65k hours of aligned audio-transcript data across 76 languages, collected from diverse public sources including parliamentary proceedings, international broadcasts, and public-domain audiobooks. For 37 languages, WorldSpeech provides more than 200 hours of aligned speech, with 28 exceeding 500 hours and 24 surpassing 1k hours. Fine-tuning existing ASR models on WorldSpeech results in an average relative Word-Error-Rate reduction of 63.5% across 11 typologically diverse languages.

Florian Grötschla, Ahmet Solak, Luca A. Lanzendörfer et al.

Recent advancements have brought generated music closer to human-created compositions, yet evaluating these models remains challenging. While human preference is the gold standard for assessing quality, translating these subjective judgments into objective metrics, particularly for text-audio alignment and music quality, has proven difficult. In this work, we generate 6k songs using 12 state-of-the-art models and conduct a survey of 15k pairwise audio comparisons with 2.5k human participants to evaluate the correlation between human preferences and widely used metrics. To the best of our knowledge, this work is the first to rank current state-of-the-art music generation models and metrics based on human preference. To further the field of subjective metric evaluation, we provide open access to our dataset of generated music and human evaluations.

Hei Yi Mak, Flint Xiaofeng Fan, Luca A. Lanzendörfer et al. · eth-zurich

In this study, we delve into Federated Reinforcement Learning (FedRL) in the context of value-based agents operating across diverse Markov Decision Processes (MDPs). Existing FedRL methods typically aggregate agents' learning by averaging the value functions across them to improve their performance. However, this aggregation strategy is suboptimal in heterogeneous environments where agents converge to diverse optimal value functions. To address this problem, we introduce the Convergence-AwarE SAmpling with scReening (CAESAR) aggregation scheme designed to enhance the learning of individual agents across varied MDPs. CAESAR is an aggregation strategy used by the server that combines convergence-aware sampling with a screening mechanism. By exploiting the fact that agents learning in identical MDPs are converging to the same optimal value function, CAESAR enables the selective assimilation of knowledge from more proficient counterparts, thereby significantly enhancing the overall learning efficiency. We empirically validate our hypothesis and demonstrate the effectiveness of CAESAR in enhancing the learning efficiency of agents, using both a custom-built GridWorld environment and the classical FrozenLake-v1 task, each presenting varying levels of environmental heterogeneity.

Luca A. Lanzendörfer, Florian Grötschla, Michael Ungersböck et al.

While neural vocoders have made significant progress in high-fidelity speech synthesis, their application on polyphonic music has remained underexplored. In this work, we propose DisCoder, a neural vocoder that leverages a generative adversarial encoder-decoder architecture informed by a neural audio codec to reconstruct high-fidelity 44.1 kHz audio from mel spectrograms. Our approach first transforms the mel spectrogram into a lower-dimensional representation aligned with the Descript Audio Codec (DAC) latent space before reconstructing it to an audio signal using a fine-tuned DAC decoder. DisCoder achieves state-of-the-art performance in music synthesis on several objective metrics and in a MUSHRA listening study. Our approach also shows competitive performance in speech synthesis, highlighting its potential as a universal vocoder.

Florian Grötschla, Elia Trachsel, Luca A. Lanzendörfer et al.

Graph-based recommender systems are commonly trained in transductive settings, which limits their applicability to new users, items, or datasets. We propose NBF-Rec, a graph-based recommendation model that supports inductive transfer learning across datasets with disjoint user and item sets. Unlike conventional embedding-based methods that require retraining for each domain, NBF-Rec computes node embeddings dynamically at inference time. We evaluate the method on seven real-world datasets spanning movies, music, e-commerce, and location check-ins. NBF-Rec achieves competitive performance in zero-shot settings, where no target domain data is used for training, and demonstrates further improvements through lightweight fine-tuning. These results show that inductive transfer is feasible in graph-based recommendation and that interaction-level message passing supports generalization across datasets without requiring aligned users or items.

Amir Dellali, Luca A. Lanzendörfer, Florian Grötschla et al.

We propose SALSA-V, a multimodal video-to-audio generation model capable of synthesizing highly synchronized, high-fidelity long-form audio from silent video content. Our approach introduces a masked diffusion objective, enabling audio-conditioned generation and the seamless synthesis of audio sequences of unconstrained length. Additionally, by integrating a shortcut loss into our training process, we achieve rapid generation of high-quality audio samples in as few as eight sampling steps, paving the way for near-real-time applications without requiring dedicated fine-tuning or retraining. We demonstrate that SALSA-V significantly outperforms existing state-of-the-art methods in both audiovisual alignment and synchronization with video content in quantitative evaluation and a human listening study. Furthermore, our use of random masking during training enables our model to match spectral characteristics of reference audio samples, broadening its applicability to professional audio synthesis tasks such as Foley generation and sound design.

Luca A. Lanzendörfer, Kyle Fearne, Florian Grötschla et al.

We present Timbru, a post-hoc audio watermarking model that achieves state-of-the-art robustness and imperceptibility trade-offs without training an embedder-detector model. Given any 44.1 kHz stereo music snippet, our method performs per-audio gradient optimization to add imperceptible perturbations in the latent space of a pretrained audio VAE, guided by a combined message and perceptual loss. The watermark can then be extracted using a pretrained CLAP model. We evaluate 16-bit watermarking on MUSDB18-HQ against AudioSeal, WavMark, and SilentCipher across common filtering, noise, compression, resampling, cropping, and regeneration attacks. Our approach attains the best average bit error rates, while preserving perceptual quality, demonstrating an efficient, dataset-free path to imperceptible audio watermarking.

Ahmet Solak, Florian Grötschla, Luca A. Lanzendörfer et al.

While recent years have seen remarkable progress in music generation models, research on their biases across countries, languages, cultures, and musical genres remains underexplored. This gap is compounded by the lack of datasets and benchmarks that capture the global diversity of music. To address these challenges, we introduce GlobalDISCO, a large-scale dataset consisting of 73k music tracks generated by state-of-the-art commercial generative music models, along with paired links to 93k reference tracks in LAION-DISCO-12M. The dataset spans 147 languages and includes musical style prompts extracted from MusicBrainz and Wikipedia. The dataset is globally balanced, representing musical styles from artists across 79 countries and five continents. Our evaluation reveals large disparities in music quality and alignment with reference music between high-resource and low-resource regions. Furthermore, we find marked differences in model performance between mainstream and geographically niche genres, including cases where models generate music for regional genres that more closely align with the distribution of mainstream styles.

Yannick Hauri, Luca A. Lanzendörfer, Till Aczel

Fashion image generation has so far focused on narrow tasks such as virtual try-on, where garments appear in clean studio environments. In contrast, editorial fashion presents garments through dynamic poses, diverse locations, and carefully crafted visual narratives. We introduce the task of virtual fashion photo-shoot, which seeks to capture this richness by transforming standardized garment images into contextually grounded editorial imagery. To enable this new direction, we construct the first large-scale dataset of garment-lookbook pairs, bridging the gap between e-commerce and fashion media. Because such pairs are not readily available, we design an automated retrieval pipeline that aligns garments across domains, combining visual-language reasoning with object-level localization. We construct a dataset with three garment-lookbook pair accuracy levels: high quality (10,000 pairs), medium quality (50,000 pairs), and low quality (300,000 pairs). This dataset offers a foundation for models that move beyond catalog-style generation and toward fashion imagery that reflects creativity, atmosphere, and storytelling.

Samuel Pfisterer, Florian Grötschla, Luca A. Lanzendörfer et al.

Recent progress in speech processing has highlighted that high-quality performance across languages requires substantial training data for each individual language. While existing multilingual datasets cover many languages, they often contain insufficient data for most languages. Thus, trained models perform poorly on the majority of the supported languages. Our work addresses this challenge by introducing a scalable pipeline for constructing speech datasets from parliamentary recordings. The proposed pipeline includes robust components for media retrieval and a two-stage alignment algorithm designed to handle non-verbatim transcripts and long-form audio. Applying this pipeline to recordings from 22 European parliaments, we extract over 61k hours of aligned speech segments, achieving substantial per-language coverage with 19 languages exceeding 1k hours and 22 languages exceeding 500 hours of high-quality speech data. We obtain an average 41.8\% reduction in word error rates over baselines when finetuning an existing ASR model on our dataset, demonstrating the usefulness of our approach.

Luca A. Lanzendörfer, Constantin Pinkl, Florian Grötschla

In this work, we study the task of multi-singer separation in a cappella music, where the number of active singers varies across mixtures. To address this, we use a power set-based data augmentation strategy that expands limited multi-singer datasets into exponentially more training samples. To separate singers, we introduce SepACap, an adaptation of SepReformer, a state-of-the-art speaker separation model architecture. We adapt the model with periodic activations and a composite loss function that remains effective when stems are silent, enabling robust detection and separation. Experiments on the JaCappella dataset demonstrate that our approach achieves state-of-the-art performance in both full-ensemble and subset singer separation scenarios, outperforming spectrogram-based baselines while generalizing to realistic mixtures with varying numbers of singers.

Frédéric Berdoz, Luca A. Lanzendörfer, Nick Tuninga et al.

Prompt-driven scene synthesis allows users to generate complete 3D environments from textual descriptions. Current text-to-scene methods often struggle with complex geometries and object transformations, and tend to show weak adherence to complex instructions. We address these limitations by introducing Reason-3D, a text-to-scene model powered by large reasoning models (LRMs). Reason-3D integrates object retrieval using captions covering physical, functional, and contextual attributes. Reason-3D then places the selected objects based on implicit and explicit layout constraints, and refines their positions with collision-aware spatial reasoning. Evaluated on instructions ranging from simple to complex indoor configurations, Reason-3D significantly outperforms previous methods in human-rated visual fidelity, adherence to constraints, and asset retrieval quality. Beyond its contribution to the field of text-to-scene generation, our work showcases the advanced spatial reasoning abilities of modern LRMs. Additionally, we release the codebase to further the research in object retrieval and placement with LRMs.

Frédéric Berdoz, Luca A. Lanzendörfer, René Caky et al.

Alignment of large language models remains a central challenge in natural language processing. Preference optimization has emerged as a popular and effective method for improving alignment, typically through training-time or prompt-based interventions. In this paper, we introduce alignment-aware decoding (AAD), a method to enhance model alignment directly at inference. Theoretically, AAD can be interpreted as implicit reward optimization, yet it requires no specialized training beyond the standard DPO setup. Empirically, AAD consistently outperforms strong baselines across diverse alignment benchmarks and model scales. Moreover, in data-constrained settings, AAD can produce high-quality synthetic data to improve alignment under standard decoding, providing a practical solution when labeled data is limited.

Florian Grötschla, Luca A. Lanzendörfer, Longxiang Jiao et al.

We introduce PANAMA, an active learning framework for the training of end-to-end parametric guitar amp models using a WaveNet-like architecture. With \model, one can create a virtual amp by recording samples that are determined by an active learning strategy to use a minimum amount of datapoints (i.e., amp knob settings). We show that gradient-based optimization algorithms can be used to determine the optimal datapoints to sample, and that the approach helps under a constrained number of samples.

Nathan Corecco, Giorgio Piatti, Luca A. Lanzendörfer et al.

Reinforcement learning (RL) has gained popularity in the realm of recommender systems due to its ability to optimize long-term rewards and guide users in discovering relevant content. However, the successful implementation of RL in recommender systems is challenging because of several factors, including the limited availability of online data for training on-policy methods. This scarcity requires expensive human interaction for online model training. Furthermore, the development of effective evaluation frameworks that accurately reflect the quality of models remains a fundamental challenge in recommender systems. To address these challenges, we propose a comprehensive framework for synthetic environments that simulate human behavior by harnessing the capabilities of large language models (LLMs). We complement our framework with in-depth ablation studies and demonstrate its effectiveness with experiments on movie and book recommendations. Using LLMs as synthetic users, this work introduces a modular and novel framework to train RL-based recommender systems. The software, including the RL environment, is publicly available on GitHub.

Peter Belcak, Luca A. Lanzendörfer, Roger Wattenhofer

We conduct a preliminary inquiry into the ability of generative transformer models to deductively reason from premises provided. We observe notable differences in the performance of models coming from different training setups and find that the deductive reasoning ability increases with scale. Further, we discover that the performance generally does not decrease with the length of the deductive chain needed to reach the conclusion, with the exception of OpenAI GPT-3 and GPT-3.5 models. Our study considers a wide variety of transformer-decoder models, ranging from 117 million to 175 billion parameters in size.