BEGIN:VCALENDAR VERSION:2.0 PRODID:-//128.220.36.25//NONSGML kigkonsult.se iCalcreator 2.26.9// CALSCALE:GREGORIAN METHOD:PUBLISH X-FROM-URL:https://www.clsp.jhu.edu X-WR-TIMEZONE:America/New_York BEGIN:VTIMEZONE TZID:America/New_York X-LIC-LOCATION:America/New_York BEGIN:STANDARD DTSTART:20231105T020000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 RDATE:20241103T020000 TZNAME:EST END:STANDARD BEGIN:DAYLIGHT DTSTART:20240310T020000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 RDATE:20250309T020000 TZNAME:EDT END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT UID:ai1ec-20117@www.clsp.jhu.edu DTSTAMP:20240329T140022Z CATEGORIES;LANGUAGE=en-US:Seminars CONTACT: DESCRIPTION:Abstract\nNeural sequence generation systems oftentimes generat e sequences by searching for the most likely sequence under the learnt pro bability distribution. This assumes that the most likely sequence\, i.e. t he mode\, under such a model must also be the best sequence it has to offe r (often in a given context\, e.g. conditioned on a source sentence in tra nslation). Recent findings in neural machine translation (NMT) show that t he true most likely sequence oftentimes is empty under many state-of-the-a rt NMT models. This follows a large list of other pathologies and biases o bserved in NMT and other sequence generation models: a length bias\, large r beams degrading performance\, exposure bias\, and many more. Many of the se works blame the probabilistic formulation of NMT or maximum likelihood estimation. We provide a different view on this: it is mode-seeking search \, e.g. beam search\, that introduces many of these pathologies and biases \, and such a decision rule is not suitable for the type of distributions learnt by NMT systems. We show that NMT models spread probability mass ove r many translations\, and that the most likely translation oftentimes is a rare event. We further show that translation distributions do capture imp ortant aspects of translation well in expectation. Therefore\, we advocate for decision rules that take into account the entire probability distribu tion and not just its mode. We provide one example of such a decision rule \, and show that this is a fruitful research direction.\nBiography\nI am a n assistant professor (UD) in natural language processing at the Institute for Logic\, Language and Computation where I lead the Probabilistic Langu age Learning group.\nMy work concerns the design of models and algorithms that learn to represent\, understand\, and generate language data. Example s of specific problems I am interested in include language modelling\, mac hine translation\, syntactic parsing\, textual entailment\, text classific ation\, and question answering.\nI also develop techniques to approach gen eral machine learning problems such as probabilistic inference\, gradient and density estimation.\nMy interests sit at the intersection of disciplin es such as statistics\, machine learning\, approximate inference\, global optimization\, formal languages\, and computational linguistics.\n \n DTSTART;TZID=America/New_York:20210419T120000 DTEND;TZID=America/New_York:20210419T131500 LOCATION:via Zoom SEQUENCE:0 SUMMARY:Wilker Aziz (University of Amsterdam) “The Inadequacy of the Mode in Neural Machine Translation” URL:https://www.clsp.jhu.edu/events/wilker-aziz-university-of-amsterdam/ X-COST-TYPE:free X-ALT-DESC;FMTTYPE=text/html:\\n\\n
\\nAbstr act
\nNeural sequence generation systems oftentimes generat e sequences by searching for the most likely sequence under the learnt pro bability distribution. This assumes that the most likely sequence\, i.e. t he mode\, under such a model must also be the best sequence it has to offe r (often in a given context\, e.g. conditioned on a source sentence in tra nslation). Recent findings in neural machine translation (NMT) show that t he true most likely sequence oftentimes is empty under many state-of-the-a rt NMT models. This follows a large list of other pathologies and biases o bserved in NMT and other sequence generation models: a length bias\, large r beams degrading performance\, exposure bias\, and many more. Many of the se works blame the probabilistic formulation of NMT or maximum likelihood estimation. We provide a different view on this: it is mode-seeking search \, e.g. beam search\, that introduces many of these pathologies and biases \, and such a decision rule is not suitable for the type of distributions learnt by NMT systems. We show that NMT models spread probability mass ove r many translations\, and that the most likely translation oftentimes is a rare event. We further show that translation distributions do capture imp ortant aspects of translation well in expectation. Therefore\, we advocate for decision rules that take into account the entire probability distribu tion and not just its mode. We provide one example of such a decision rule \, and show that this is a fruitful research direction.
\nBi ography
\nI am an assistant professor (UD) in natu ral language processing at the Institute for Logic\, Language and Computation where I lead the Probabilistic Language Learning group.
\nMy work concerns the design of models and algorithms that learn to represe nt\, understand\, and generate language data. Examples of specific problem s I am interested in include language modelling\, machine translation\, sy ntactic parsing\, textual entailment\, text classification\, and question answering.
\nI also develop techniques to approach general machine l earning problems such as probabilistic inference\, gradient and density es timation.
\nMy interests sit at the intersection of disciplines such as statistics\, machine learning\, approximate inference\, global optimiz ation\, formal languages\, and computational linguistics.
\n\n< p> \n X-TAGS;LANGUAGE=en-US:2021\,April\,Aziz END:VEVENT BEGIN:VEVENT UID:ai1ec-21031@www.clsp.jhu.edu DTSTAMP:20240329T140022Z CATEGORIES;LANGUAGE=en-US:Seminars CONTACT: DESCRIPTION:Abstract\nMost people take for granted that when they speak\, t hey will be heard and understood. But for the millions who live with speec h impairments caused by physical or neurological conditions\, trying to co mmunicate with others can be difficult and lead to frustration. While ther e have been a great number of recent advances in Automatic Speech Recognit ion (ASR) technologies\, these interfaces can be inaccessible for those wi th speech impairments.\nIn this talk\, we will present Parrotron\, an end- to-end-trained speech-to-speech conversion model that maps an input spectr ogram directly to another spectrogram\, without utilizing any intermediate discrete representation. The system is also trained to emit words in addi tion to a spectrogram\, in parallel. We demonstrate that this model can be trained to normalize speech from any speaker regardless of accent\, pro sody\, and background noise\, into the voice of a single canonical target speaker with a fixed accent and consistent articulation and prosody. We fu rther show that this normalization model can be adapted to normalize highl y atypical speech from speakers with a variety of speech impairments (due to\, ALS\, Cerebral-Palsy\, Deafness\, Stroke\, Brain Injury\, etc.) \, r esulting in significant improvements in intelligibility and naturalness\, measured via a speech recognizer and listening tests. Finally\, demonstrat ing the utility of this model on other speech tasks\, we show that the sam e model architecture can be trained to perform a speech separation task.\n Dimitri will give a brief description of some key moments in development o f speech recognition algorithms that he was involved in and their applicat ions to YouTube closed captions\, Live Transcribe and wearable subtitles. \nFadi will then speak about the development of Parrotron.\nBiographies\nD imitri Kanevsky started his career at Google working on speech recognition algorithms. Prior to joining Google\, Dimitri was a Research staff member in the Speech Algorithms Department at IBM. Prior to IBM\, he worked at a number of centers for higher mathematics\, including Max Planck Institu te in Germany and the Institute for Advanced Studies in Princeton. He curr ently holds 295 US patents and was Master Inventor at IBM. MIT Technology Review recognized Dimitri conversational biometrics based security patent as one of five most influential patents for 2003. In 2012 Dimitri was hono red at the White House as a Champion of Change for his efforts to advance access to science\, technology\, engineering\, and math.\nFadi Biadsy is a senior staff research scientist at Google NY for the past ten years. He h as been exploring and leading multiple projects at Google\, including spee ch recognition\, speech conversion\, language modeling\, and semantic unde rstanding. He received his PhD from Columbia University in 2011. At Colum bia\, he researched a variety of speech and language processing projects i ncluding\, dialect and accent recognition\, speech recognition\, charismat ic speech and question answering. He holds a BSc and MSc in mathematics a nd computer science. He worked on handwriting recognition during his maste rs degree and he worked as a senior software developer for five years at D alet digital media systems building multimedia broadcasting systems. DTSTART;TZID=America/New_York:20211105T120000 DTEND;TZID=America/New_York:20211105T131500 LOCATION:Hackerman Hall B17 @ 3400 N. Charles Street\, Baltimore\, MD 21218 SEQUENCE:0 SUMMARY:Fadi Biadsy and Dimitri Kanevsky (Google) “Speech Recognition: From Speaker Dependent to Speaker Independent to Full Personalization” “Parrot ron: A Unified E2E Speech-to Speech Conversion and ASR Model for Atypical Speech” URL:https://www.clsp.jhu.edu/events/fadi-biadsy-and-dimitri-kanevsky-google / X-COST-TYPE:free X-ALT-DESC;FMTTYPE=text/html:\\n\\n\\n
Abstr act
\nMost people take for granted that when they speak\, they will be heard and understood. But for the millions who live with speech impairments caused by physical or neurological condi tions\, trying to communicate with others can be difficult and lead to fru stration. While there have been a great number of recent advances in Autom atic Speech Recognition (ASR) technologies\, these interfaces can be inacc essible for those with speech impairments.
\nIn this talk\, we will present Parrotron\, an end-to-end-trained speech-to-sp eech conversion model that maps an input spectrogram directly to another s pectrogram\, without utilizing any intermediate discrete representation. T he system is also trained to emit words in addition to a spectrogram\, in parallel. We demonstrate that this model can be trained to normalize spe ech from any speaker regardless of accent\, prosody\, and background noise \, into the voice of a single canonical target speaker with a fixed accent and consistent articulation and prosody. We further show that this normal ization model can be adapted to normalize highly atypical speech from spea kers with a variety of speech impairments (due to\, ALS\, Cerebral-Palsy\, Deafness\, Stroke\, Brain Injury\, etc.) \, resulting in significant imp rovements in intelligibility and naturalness\, measured via a speech recog nizer and listening tests. Finally\, demonstrating the utility of this mod el on other speech tasks\, we show that the same model architecture can be trained to perform a speech separation task.
\nDimitri will give a brief description of some key moments in development o f speech recognition algorithms that he was involved in and their applicat ions to YouTube closed captions\, Live Transcribe and wearable subtitles.
\nFadi will then speak about the development of Parrotron.
\nBiographies
\nDimitri K anevsky started his career at Google working on speech recognitio n algorithms. Prior to joining Google\, Dimitri was a Research staff membe r in the Speech Algorithms Department at IBM. Prior to IBM\, he worked a t a number of centers for higher mathematics\, including Max Planck Instit ute in Germany and the Institute for Advanced Studies in Princeton. He cur rently holds 295 US patents and was Master Inventor at IBM. MIT Technology Review recognized Dimitri conversational biometrics based security patent as one of five most influential patents for 2003. In 2012 Dimitri was hon ored at the White House as a Champion of Change for his efforts to advance access to science\, technology\, engineering\, and math.
\nFadi Biadsy is a senior staff research scientist at Google NY for the past ten years. He has been exploring and leading multiple projects a t Google\, including speech recognition\, speech conversion\, language mod eling\, and semantic understanding. He received his PhD from Columbia Uni versity in 2011. At Columbia\, he researched a variety of speech and langu age processing projects including\, dialect and accent recognition\, speec h recognition\, charismatic speech and question answering. He holds a BSc and MSc in mathematics and computer science. He worked on handwriting rec ognition during his masters degree and he worked as a senior software deve loper for five years at Dalet digital media systems building multimedia br oadcasting systems.
\n X-TAGS;LANGUAGE=en-US:2021\,Biadsy and Kanevsky\,November END:VEVENT BEGIN:VEVENT UID:ai1ec-24157@www.clsp.jhu.edu DTSTAMP:20240329T140022Z CATEGORIES;LANGUAGE=en-US:Seminars CONTACT: DESCRIPTION:Abstract\nIn this talk\, I will present a simple extension of i mage-based Masked Autoencoders (MAE) to self-supervised representation lea rning from audio spectrograms. Following the Transformer encoder-decoder d esign in MAE\, our Audio-MAE first encodes audio spectrogram patches with a high masking ratio\, feeding only the non-masked tokens through encoder layers. The decoder then re-orders and decodes the encoded context padded with mask tokens\, in order to reconstruct the input spectrogram. We find it beneficial to incorporate local window attention in the decoder\, as au dio spectrograms are highly correlated in local time and frequency bands. We then fine-tune the encoder with a lower masking ratio on target dataset s. Empirically\, Audio-MAE sets new state-of-the-art performance on six au dio and speech classification tasks\, outperforming other recent models th at use external supervised pre-training.\nBio\nFlorian Metze is a Research Scientist Manager at Meta AI in New York\, supporting a team of researche rs and engineers working on multi-modal (image\, video\, audio\, text) con tent understanding for Meta’s Family of Apps (Instagram\, Threads\, Facebo ok\, WhatsApp). He used to be an Associate Research Professor at Carnegie Mellon University\, in the School of Computer Science’s Language Technolog ies Institute\, where he still is an Adjunct Professor. He is also a co-fo under of Abridge\, a company working on extracting information from doctor patient conversations. His work covers many areas of speech recognition a nd multi-media analysis with a focus on end-to-end deep learning. Currentl y\, he focuses on multi-modal processing of videos\, and using that inform ation to recommend unconnected content. In the past\, he has worked on low resource and multi-lingual speech processing\, speech recognition with ar ticulatory features\, large-scale multi-media retrieval and summarization\ , information extraction from medical interviews\, and recognition of pers onality or similar meta-data from speech.\nFor more information\, please s ee http://www.cs.cmu.edu/directory/fmetze\n DTSTART;TZID=America/New_York:20231110T120000 DTEND;TZID=America/New_York:20231110T131500 LOCATION:Hackerman Hall B17 @ 3400 N. Charles Street\, Baltimore\, MD 21218 SEQUENCE:0 SUMMARY:Florian Metze (CMU) “Masked Autoencoders that Listen” URL:https://www.clsp.jhu.edu/events/florian-metze-cmu/ X-COST-TYPE:free X-ALT-DESC;FMTTYPE=text/html:\\n\\n\\nAbstr act
\nIn this talk\, I will present a simple extension of i mage-based Masked Autoencoders (MAE) to self-supervised representation lea rning from audio spectrograms. Following the Transformer encoder-decoder d esign in MAE\, our Audio-MAE first encodes audio spectrogram patches with a high masking ratio\, feeding only the non-masked tokens through encoder layers. The decoder then re-orders and decodes the encoded context padded with mask tokens\, in order to reconstruct the input spectrogram. We find it beneficial to incorporate local window attention in the decoder\, as au dio spectrograms are highly correlated in local time and frequency bands. We then fine-tune the encoder with a lower masking ratio on target dataset s. Empirically\, Audio-MAE sets new state-of-the-art performance on six au dio and speech classification tasks\, outperforming other recent models th at use external supervised pre-training.
\nBio
\nFlorian Metze is a Research Scientist Manager at Meta AI in New York\ , supporting a team of researchers and engineers working on multi-modal (i mage\, video\, audio\, text) content understanding for Meta’s Family of Ap ps (Instagram\, Threads\, Facebook\, WhatsApp). He used to be an Associate Research Professor at Carnegie Mellon University\, in the School of Compu ter Science’s Language Technologies Institute\, where he still is an Adjun ct Professor. He is also a co-founder of Abridge\, a company working on ex tracting information from doctor patient conversations. His work covers ma ny areas of speech recognition and multi-media analysis with a focus on en d-to-end deep learning. Currently\, he focuses on multi-modal processing o f videos\, and using that information to recommend unconnected content. In the past\, he has worked on low resource and multi-lingual speech process ing\, speech recognition with articulatory features\, large-scale multi-me dia retrieval and summarization\, information extraction from medical inte rviews\, and recognition of personality or similar meta-data from speech.< /p>\n
For more information\, please see http://www.cs.cmu.edu/directory/fmetze
\n\n X-TAGS;LANGUAGE=en-US:2023\,Metze\,November END:VEVENT END:VCALENDAR