BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/New_York BEGIN:DAYLIGHT DTSTART:20250309T030000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:EDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20251102T010000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:EST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20250830T233640Z UID:A1BFEBEC-94C6-4823-A8A8-65DBD8F56390 DTSTART;TZID=America/New_York:20250827T090000 DTEND;TZID=America/New_York:20250827T103000 DESCRIPTION:https://landing.signalprocessingsociety.org/ieee-sps-webinars-2 7-aug-2025\n\nThe intersection of speech and language models offer unique opportunities and challenges. This talk provides a comprehensive walkthrou gh of speech-language model research from NVIDIA NeMo. We cover several ty pes of models such as attention-encoder-decoder Canary-1B\, and LLM-based architectures such as SALM or BESTOW. In particular\, we highlight the cha llenges in training and inference efficiency of such models and propose ro bust solutions via 2D bucketing and batch size OOMptimizer. Finally\, we h ighlight the difficulty of preserving text-domain capabilities in speech-a ugmented training and present several possible solutions: EMMeTT\, VoiceTe xtBlender\, and Canary-Qwen-2.5B.\n\nAbout the Presenter:\n\nPiotr Żelask o received the B.S. and M.Sc. degrees in acoustic engineering\, and the Ph .D. in electronic engineering from AGH-University Krakow\, Poland in 2013\ , 2014\, and 2019 respectively.\n\nHe is currently a research scientist at NVIDIA NeMo building multitask and multimodal models and efficient traini ng infrastructure. He held a research scientist position at JHU’s CLSP a nd developed speech technology at different companies (Techmo\, Avaya\, Me aning.Team).\n\nDr. Żelasko is a co-author of the next-generation Kaldi t oolkit (k2) and the maintainer of Lhotse.\n\nAgenda: \nhttps://landing.sig nalprocessingsociety.org/ieee-sps-webinars-27-aug-2025\n\nPlease register here too and on Vtools too.\n\nVirtual: https://events.vtools.ieee.org/m/4 95161 LOCATION:Virtual: https://events.vtools.ieee.org/m/495161 ORGANIZER:baris.kazar@oracle.com SEQUENCE:13 SUMMARY:Foundational Speech Models and their Efficient Training with NVIDIA NeMo {AI Talks with Coffee/Tea #30} URL;VALUE=URI:https://events.vtools.ieee.org/m/495161 X-ALT-DESC:Description: <br /><p><a href="https://landing.signalprocessings ociety.org/ieee-sps-webinars-27-aug-2025">https://landing.signalprocessing society.org/ieee-sps-webinars-27-aug-2025</a></p>\n<div class="row-fluid-w rapper row-depth-1 row-number-6 dnd-row">\n<div class="row-fluid ">\n<div class="span12 widget-span widget-type-custom_widget dnd-module" data-widge t-type="custom_widget" data-x="0" data-w="12">\n<div id="hs_cos_wrapper_wi dget_1684529549375" class="hs_cos_wrapper hs_cos_wrapper_widget hs_cos_wra pper_type_module widget-type-rich_text" data-hs-cos-general-type="widget" data-hs-cos-type="module">\n<p>The intersection of speech and language mod els offer unique opportunities and challenges. This talk provides a compre hensive walkthrough of speech-language model research from NVIDIA NeMo. We cover several types of models such as attention-encoder-decoder Canary-1B \, and LLM-based architectures such as SALM or BESTOW. In particular\, we highlight the challenges in training and inference efficiency of such mode ls and propose robust solutions via 2D bucketing and batch size OOMptimize r. Finally\, we highlight the difficulty of preserving text-domain capabil ities in speech-augmented training and present several possible solutions: EMMeTT\, VoiceTextBlender\, and Canary-Qwen-2.5B.</p>\n<span id="hs_cos_w rapper_widget_1684529549375_" class="hs_cos_wrapper hs_cos_wrapper_widget hs_cos_wrapper_type_rich_text" data-hs-cos-general-type="widget" data-hs-c os-type="rich_text"></span></div>\n</div>\n</div>\n</div>\n<div class="row -fluid-wrapper row-depth-1 row-number-7 dnd-row">\n<div class="row-fluid " >\n<div class="span12 widget-span widget-type-custom_widget dnd-module" da ta-widget-type="custom_widget" data-x="0" data-w="12">\n<div id="hs_cos_wr apper_module_16896111543583" class="hs_cos_wrapper hs_cos_wrapper_widget h s_cos_wrapper_type_module widget-type-rich_text" data-hs-cos-general-type= "widget" data-hs-cos-type="module">\n<h2>About the Presenter:</h2>\n<p>Pio tr Żelasko received the B.S. and M.Sc. degrees in acoustic engineering\, and the Ph.D. in electronic engineering from AGH-University Krakow\, Polan d in 2013\, 2014\, and 2019 respectively.<br><br>He is currently a researc h scientist at NVIDIA NeMo building multitask and multimodal models and ef ficient training infrastructure. He held a research scientist position at JHU&rsquo\;s CLSP and developed speech technology at different companies ( Techmo\, Avaya\, Meaning.Team).<br><br>Dr. Żelasko is a co-author of the next-generation Kaldi toolkit (k2) and the maintainer of Lhotse.</p>\n<spa n id="hs_cos_wrapper_module_16896111543583_" class="hs_cos_wrapper hs_cos_ wrapper_widget hs_cos_wrapper_type_rich_text" data-hs-cos-general-type="wi dget" data-hs-cos-type="rich_text"></span></div>\n</div>\n</div>\n</div><b r /><br />Agenda: <br /><p><a href="https://landing.signalprocessingsociet y.org/ieee-sps-webinars-27-aug-2025">https://landing.signalprocessingsocie ty.org/ieee-sps-webinars-27-aug-2025</a></p>\n<p>Please register here too and on Vtools too.</p> END:VEVENT END:VCALENDAR