[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-118685-en":3,"doc-seo-118685-105":29,"detail-sidebar-cat-0-en-105":90},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":20,"language":21,"language_code":22,"site_id":23,"html_lang":22,"table_of_contents":24,"faqs":25,"seo_title":26,"seo_description":14,"update_tm":27,"read_time":28},118685,7971461741311,"Ophelia","https://ap-avatar.wpscdn.com/avatar/74000253aff267980c6?x-image-process=image/resize,m_fixed,w_180,h_180&k=1779345379180704826",6,"Technology","Korvax - Audio Processing and Machine Learning in JAX","Korvax is an audio processing and machine learning package built for JAX, delivering differentiable, GPU-ready implementations of common audio loss functions and signal-processing utilities. The workshop materials benchmark Korvax against established PyTorch libraries, evaluating CPU and GPU performance across multi-scale spectral loss and spectral optimal transport. Results show consistent speedups, including roughly 2.5× faster gradients for MSS and large GPU gains for SOT. Korvax also supports efficient differentiation for time-varying all-pole filters used in LPC, with comparable GPU runtime and faster gradient computation on CPU.","DMRN+20: DIGITAL MUSIC RESEARCH NETWORK ONE-DAY WORKSHOP 2025  \nKING ’S COLLEGE LONDON TUE 16 DECEMBER 2025  \nKorvax: Audio Processing and Machine Learning in JAX  \nDavid Marttila∗ and Joshua D. Reiss  \nCentre for Digital Music, Queen Mary University of London, United Kingdom, [d.sudholt@qmul.ac.uk](d.sudholt@qmul.ac.uk)  \nAbstract—We present Korvax, a package for audio processing and machine learning in JAX. Korvax provides differentiable, GPU-ready implementations of common audio loss functions and signal processing utilities. We benchmark the performance on CPU and GPU and find that Korvax can achieve noticeable speedups compared to equivalent PyTorch libraries.  \nI. INTRODUCTION  \nPyTorch is the prevalent framework in the audio machine learning community. Many common workflows are implemented in the officially supported Torchaudio library or a mature ecosystem of community-maintained PyTorch packages.  \nJAX [1] is a newer framework for high-performance arrayoriented computations in Python that has gained popularity in recent years, positioning itself as one of the main alternatives to PyTorch. It focuses on providing function transformations for justin-time compilation, vectorization, automatic differentiation, and parallelization. This allows users to compose complex graphs of numerical operations in Python and execute them efficiently, often faster than in other frameworks.  \nDespite these advantages, JAX has seen little adoption in audio signal processing, likely in part due to the lack of domain-specific tooling compared to PyTorch. Korvax aims to address this gap by porting commonly-used features from Torchaudio, librosa [2], and other packages into JAX. The source code of Korvax and documentation of currently implemented features is available online.1 We welcome feedback and contributions to advance the state of audio machine learning research in JAX.  \nII. PERFORMANCE BENCHMARKS  \nWe benchmark selected Korvax features against reference PyTorch packages. The results are shown in Table 1 .  \nLoss Functions: Korvax provides a general interface for calculating frame-wise losses at multiple resolutions. We benchmark its performance by generating two random audio signals x, y with a length of 64000 samples and a batch size of 256 . We measure the time it takes to compute the loss f(x, y) and the partial gradient off w.r.t. x.  \nFor the multi-scale spectral loss (MSS) benchmark, we use 6 resolutions and compare to the auraloss [3] PyTorch package. Korvax computes the gradients consistently about 2.5x faster. We also benchmark spectral optimal transport (SOT) [4] . We use only one resolution and compare to the sot-loss2 PyTorch package. The choice of framework has a significant effect on performance: Kor-  \n∗D. Marttila is supported by UK Research and Innovation [grant number EP/S022694/1] .  \n1 [https://github.com/davidmarttila/korvax](https://github.com/davidmarttila/korvax)  \n[2](2 github.com/bernardo-torres/spectral-optimal-transport/)[ github.com/bernardo-torres/spectral-optimal-transport/](2 github.com/bernardo-torres/spectral-optimal-transport/)  \nTable 1: Performance comparison of Korvax with PyTorch packages. The time taken to compute a given value is measured in seconds and averagedover 100 runs. The CPU is an Apple M2 Pro chip with 12 cores. The GPU is an NVIDIA H100 .  \n\n|  |  | CPU\u003Cbr>f (x, y) ~~ ~~∂f∂(xx,y) | GPU\u003Cbr>f (x, y) ~~ ~~∂f∂(xx,y) |\n| --- | --- | --- | --- |\n| MSS | Korvax\u003Cbr>auraloss | 0.608 0.993 | 0.010 0.018 |\n|  |  | 0.966 2.450 | 0.032 0.045 |\n| SOT | Korvax\u003Cbr>sot-loss | 1.588 1.957 | 0.039 0.043 |\n|  |  | 0.554 0.844 | 1.443 1.760 |\n| LPC | Korvax\u003Cbr>torchlpc | 0.022 0.091 | 0.106 0.198 |\n|  |  | 0.016 0.153 | 0.097 0.197 |\n\nvax is 2–3x slower than sot-loss on the CPU, but dramatically faster on the GPU, by a factor of about 40x.  \nTime-Varying All-Pole Filters: Korvax implements efficient differentiation of linear predictive coding (LPC) with time-varying all-pole filters as proposed in [5], ","cbCaieL03qoSGJXS","https://ap.wps.com/l/cbCaieL03qoSGJXS","pdf",184942,1,"English","en",105,"# Introduction\n## Performance Benchmarks\n### Loss Functions\n### Time-Varying All-Pole Filters\n# References","[{\"question\":\"What is Korvax, and what does it provide for JAX-based audio ML?\",\"answer\":\"Korvax is a package for audio processing and machine learning in JAX. It provides differentiable, GPU-ready implementations of common audio loss functions and signal processing utilities.\"},{\"question\":\"How does Korvax performance compare with PyTorch libraries on CPU and GPU?\",\"answer\":\"Benchmarks compare Korvax features to reference PyTorch implementations. The materials report noticeable speedups, including about 2.5× faster gradient computation for multi-scale spectral loss and much larger GPU speed gains for spectral optimal transport.\"},{\"question\":\"How does Korvax handle differentiation for time-varying LPC with all-pole filters?\",\"answer\":\"Korvax implements efficient differentiation for LPC with time-varying all-pole filters as proposed in prior work. Filtering code is compiled ahead-of-time in C++, and benchmarks show similar GPU speeds with faster gradient computation on CPU.\"}]","Korvax - Audio Processing and Machine Learning in JAX | PDF",1785684883,3,{"code":4,"msg":30,"data":31},"ok",{"site_id":23,"language":22,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":27},"korvax-audio-processing-and-machine-learning-in-jax","",{"@graph":35,"@context":84},[36,52,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,49],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":28},"https://docshare.wps.com/document/technology/",{"item":50,"name":13,"@type":42,"position":51},"https://docshare.wps.com/document/korvax-audio-processing-and-machine-learning-in-jax/118685/",4,{"url":50,"name":13,"@type":53,"author":54,"headline":13,"publisher":56,"fileFormat":59,"inLanguage":22,"description":14,"dateModified":60,"datePublished":61,"encodingFormat":59,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":55},"Person",{"url":40,"name":57,"@type":58},"DocShare","Organization","application/pdf","2026-08-04","2026-08-02",true,{"@type":64,"interactionType":65,"userInteractionCount":20},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What is Korvax, and what does it provide for JAX-based audio ML?","Question",{"text":74,"@type":75},"Korvax is a package for audio processing and machine learning in JAX. It provides differentiable, GPU-ready implementations of common audio loss functions and signal processing utilities.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does Korvax performance compare with PyTorch libraries on CPU and GPU?",{"text":79,"@type":75},"Benchmarks compare Korvax features to reference PyTorch implementations. The materials report noticeable speedups, including about 2.5× faster gradient computation for multi-scale spectral loss and much larger GPU speed gains for spectral optimal transport.",{"name":81,"@type":72,"acceptedAnswer":82},"How does Korvax handle differentiation for time-varying LPC with all-pole filters?",{"text":83,"@type":75},"Korvax implements efficient differentiation for LPC with time-varying all-pole filters as proposed in prior work. Filtering code is compiled ahead-of-time in C++, and benchmarks show similar GPU speeds with faster gradient computation on CPU.","https://schema.org",{"og:url":50,"og:type":86,"og:title":13,"og:site_name":57,"og:description":14},"article",{"robots":88,"canonical":50},"index,follow",{"doc_id":7,"site_id":23},{"code":4,"msg":5,"data":91},[92,96,100,104,109,112,117,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":51,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":110,"slug":111},50,"technology",{"id":113,"doc_module":4,"doc_module_name":45,"category_name":114,"show_sort_weight":115,"slug":116},7,"Healthcare",40,"healthcare",{"id":118,"doc_module":4,"doc_module_name":45,"category_name":119,"show_sort_weight":120,"slug":121},8,"Research & Report",30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":105,"slug":137},19,"General","general"]