Recent breakthroughs in solving reasoning, math and coding problems with Large Language Models (LLMs) have been enabled by investing substantial computation budgets at inference time. Therefore, inference speed is one of the most critical properties of LLM architectures, and there is a growing need for LLMs that are efficient and fast at inference. Recently, LLMs built on the xLSTM architecture have emerged as a powerful alternative to Transformers, offering linear compute scaling with sequence length and constant memory usage, both highly desirable properties for efficient inference. However, such xLSTM-based LLMs have yet to be scaled to larger models and assessed and compared with respect to inference speed and efficiency. In this work, we introduce xLSTM 7B, a 7-billion-parameter LLM that combines xLSTM’s architectural benefits with targeted optimizations for fast and efficient inference. Our experiments demonstrate that xLSTM 7B achieves performance on downstream tasks comparable to other similar-sized LLMs, while providing significantly faster inference speeds and greater efficiency compared to Llama- and Mamba-based LLMs. These results establish xLSTM 7B as the fastest and most efficient 7B LLM, offering a solution for tasks that require large amounts of test-time computation. Our work highlights xLSTM’s potential as a foundational architecture for methods building on heavy use of LLM inference. Our model weights, model code and training code are open-source. Model: https://huggingface.co/NX-AI/xLSTM-7b Code: https://github.com/NX-AI/xlstm and https://github.com/NX-AI/xlstm-jax.

In-context learning, the ability of large language models to perform tasks using only examples provided in the prompt, has recently been adapted for time series forecasting. This paradigm enables zero-shot prediction, where past values serve as context for forecasting future values, making powerful forecasting tools accessible to non-experts and increasing the performance when training data are scarce. Most existing zero-shot forecasting approaches rely on transformer architectures, which, despite their success in language, often fall short of expectations in time series forecasting, where recurrent models like LSTMs frequently have the edge. Conversely, while LSTMs are well-suited for time series modeling due to their state-tracking capabilities, they lack strong in-context learning abilities. We introduce *TiRex* that closes this gap by leveraging xLSTM, an enhanced LSTM with competitive in-context learning skills. Unlike transformers, state-space models, or parallelizable RNNs such as RWKV, TiRex retains state tracking, a critical property for long-horizon forecasting. To further facilitate its state-tracking ability, we propose a training-time masking strategy called CPM. TiRex sets a new state of the art in zero-shot time series forecasting on the Hugging Face benchmarks *GiftEval* and *Chronos-ZS*, outperforming significantly larger models including *TabPFN-TS* (Prior Labs), *Chronos Bolt* (Amazon), *TimesFM* (Google), and *Moirai* (Salesforce) across both short- and long-term forecasts.

We investigate a late-fusion approach to piano transcription, combined with a strong temporal prior in the form of a handcrafted Hidden Markov Model (HMM). The network architecture under consideration is compact in terms of its number of parameters and easy to train with gradient descent. The network outputs are fused over time in the final stage to obtain note segmentations, with an HMM whose transition probabilities are chosen based on a model of attack, decay, sustain, release (ADSR) envelopes, commonly used for sound synthesis. The note segments are then subject to a final binary decision rule to reject too weak note segment hypotheses. We obtain state-of-the-art results on the MAPS dataset, and are able to outperform other approaches by a large margin, when predicting complete note regions from onsets to offsets.

In this paper, we present madmom, an open-source audio processing and music information retrieval (MIR) library written in Python. madmom features a concise, NumPy-compatible, object oriented design with simple calling conventions and sensible default values for all parameters, which facilitates fast prototyping of MIR applications. Prototypes can be seamlessly converted into callable processing pipelines through madmom's concept of Processors, callable objects that run transparently on multiple cores. Processors can also be serialised, saved, and re-run to allow results to be easily reproduced anywhere. Apart from low-level audio processing, madmom puts emphasis on musically meaningful high-level features. Many of these incorporate machine learning techniques and madmom provides a module that implements some methods commonly used in MIR such as hidden Markov models and neural networks. Additionally, madmom comes with several state-of-the-art MIR algorithms for onset detection, beat, downbeat and meter tracking, tempo estimation, and chord recognition. These can easily be incorporated into bigger MIR systems or run as stand-alone programs.

We present a system that we call ‘The Piano Music Companion’ and that is able to follow and understand (at least to some extent) a live piano performance. Within a few seconds this system can identify the piece that is being played, and the position within the piece. It then tracks the progress of the performer over time via a robust score following algorithm. The companion is useful in multiple ways, e.g., it can be used for piece identification, music visualisation, during piano rehearsal and for automatic page turning.

Musical onset detection is one of the most elementary tasks in music analysis, but still only solved imperfectly for polyphonic music signals. Interpreted as a computer vision problem in spectrograms, Convolutional Neural Networks (CNNs) seem to be an ideal fit. On a dataset of about 100 minutes of music with 26k annotated onsets, we show that CNNs outperform the previous state-of-the-art while requiring less manual preprocessing. Investigating their inner workings, we find two key advantages over hand-designed methods: Using separate detectors for percussive and harmonic onsets, and combining results from many minor variations of the same scheme. The results suggest that even for well-understood signal processing tasks, machine learning can be superior to knowledge engineering.

In this paper a new approach for polyphonic piano note onset transcription is presented. It is based on a recurrent neural network to simultaneously detect the onsets and the pitches of the notes from spectral features. Long Short-Term Memory units are used in a bidirectional neural network to model the context of the notes. The use of a single regression output layer instead of the often used one-versus-all classification approach enables the system to significantly lower the number of erroneous note detections. Evaluation is based on common test sets and shows exceptional temporal precision combined with a significant boost in note transcription performance compared to current state-of-the-art approaches. The system is trained jointly with various synthesized piano instruments and real piano recordings and thus generalizes much better than existing systems.