Joseph (Yossi) Keshet (Hebrew: יוסף (יוסי) קשת; born: 28 February 1973) is an Israeli professor in the Electrical and Computer Engineering Faculty of the Technion, where he is the director of the Speech, Language, and Deep Learning Lab. His research focuses on human speech processing and machine learning. == Early life and education == Keshet was born in Tel-Aviv. He graduated from the Amal School and began his academic studies at the Department of Electrical Engineering-Systems at Tel-Aviv University in 1991 and received his B.Sc. (Cum Laude) in 1994. Keshet served in the IDF Unit 8200 from 1995 to 2002 as the head of the speech processing research section in the R&D Center. During his service, he received a national award from the Administration for the Development of Weapons and Technological Infrastructure (Maf’at). Keshet was award his M.Sc. from the same department after he completed his Israel Defense Force service in 2002. His Dissertation was titled: Stop consonant spotting in continuous speech and was supervised by Dan Chazan from IBM Research Labs, Haifa. He continued his Ph.D. studies at the Hebrew University of Jerusalem until 2008. Prof. Yoram Singer supervised his thesis on Large Margin Algorithms for Discriminative Continuous Speech. == Career == Keshet was a Research Associate (postdoc) at IDIAP Research Institute, Martigny, Switzerland in 2007, and joined the TTI-Chicago and Department of Computer Science, University of Chicago, Chicago, IL in 2009 as Research Assistant Professor. In 2013, he returned to Israel and joined the Computer Science department at Bar-Ilan University as a senior lecturer and head of the Speech, Language, and Deep Learning Lab. In 2020, Keshet became a Founding Venture Partner at the Disruptive AI Venture Capital. In the same year, he also joined Amazon in Tel-Aviv as an Amazon Scholar. In 2022, Keshet joined the Faculty of Electrical and Computer Engineering at the Technion. == Research == Keshet's research work focuses on both machine learning and computational study of human speech and language. His work on speech and language concentrates on speech processing, speech recognition, acoustic phonetics, and pathological speech. In machine learning, Keshet is focused on deep learning and structured tasks. According to Google Scholar (September 2020), Keshet is one of the 15 most cited researchers in the field of spoken language processing. The algorithms that were developed in the Speech, Language, and Deep Learning Lab can analyze different pathological conditions in the throat and vocal cords based on the subject's voice. Other algorithms showed that the voice can be used to estimate physical and emotional state of the speaker. Another research led by Keshet suggested that it is possible to fool structured AI systems (like Google Voice). == Membership in professional societies == Keshet is the founder and chair of the Machine Learning for Speech and Language Processing Special Interest Group (SIGML) of the International Speech Communication Association (ISCA), from 2011. He is a senior member of the IEEE Signal Processing Society since 2018 and a member of ISCA since 2002. == Publications == Prof. Keshet has authored more than 70 scientific publications and edited one book. === Book === Joseph Keshet and Samy Bengio, Eds., Automatic Speech and Speaker Recognition: Large Margin and Kernel Methods, John Wiley & Sons, March 2009. === Selected articles === Jacob T. Cohen, Alma Cohen, Limor Benyamini, Yossi Adi, Joseph Keshet, Predicting glottal closure insufficiency using fundamental frequency contour analysis, Head & Neck, Journal of the Sciences and Specialities of the Head and Neck, Volume 41, Issue 7, pp. 2324–2331, July 2019. Yehoshua Dissen, Jacob Goldberger, and Joseph Keshet, Formant Estimation and Tracking: A Deep Learning Approach, Journal of the Acoustical Society of America, 145 (2), February 2019. Joseph Keshet, Automatic speech recognition: A primer for speech-language pathology researchers, International Journal of Speech-Language Pathology, Vol. 20 No. 6, pp. 599–609, 2018. Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, Joseph Keshet, Turning Your Weakness Into a Strength: Watermarking Deep Neural Networks by Backdooring, Usenix, 2018. Tzeviya Fuchs, Joseph Keshet, Spoken Term Detection Automatically Adjusted for a Given Threshold, IEEE Journal of Selected Topics in Signal Processing, Dec 2017, Volume 11, Issue 8, pp. 1–8. Moustapha Cisse, Yossi Adi, Natalia Neverova, Joseph Keshet, Houdini: Fooling Deep Structured Visual and Speech Recognition Models with Adversarial Examples, Neural Information and Processing Systems (NIPS), 2017. Joseph Keshet, Subhransu Maji, Tamir Hazan, and Tommi Jaakkola, Perturbation Models and PAC-Bayesian Generalization Bounds, in Perturbations, Optimization, and Statistics, Tamir Hazan, George Papandreou, and Daniel Tarlow, Eds., The MIT Press, 2016. Matthew Goldrick, Joseph Keshet, Erin Gustafson, Jordana Heller, and Jeremy Needle, Automatic Analysis of Slips of the Tongue: Insights into the Cognitive Architecture of Speech Production, Cognition, 149, 31–39, 2016. Joseph Keshet, Optimizing the Measure of Performance in Structured Prediction, in Advanced Structured Prediction, Sebastian Nowozin, Peter V. Gehler, Jeremy January, and Christoph H. Lampert, Eds., The MIT Press, 2014. Morgan Sonderegger and Joseph Keshet, Automatic Measurement of Voice Onset Time using Discriminative Structured Prediction, Journal of the Acoustical Society of America, Vol. 132, Issue 6, pp. 3965−3979, 2012. David McAllester, Tamir Hazan and Joseph Keshet, Direct Loss Minimization for Structured Prediction, The 24th Annual Conference on Neural Information Processing Systems (NIPS), 2010. Joseph Keshet, David Grangier and Samy Bengio, Discriminative Keyword Spotting, Speech Communication, Volume 51, Issue 4, pp. 317–329, April 2009. == Personal life == Keshet is married to Lital. They have three children.
Wispr
Wispr AI is a software company founded in 2021 by Tanay Kothari and Sahaj Garg that develops voice-based interfaces for computers and other devices. The company’s main product, Wispr Flow, is an AI-powered speech-to-text application available on macOS, Windows and iOS. == History == Wispr was founded in 2021 with the goal of building a non-invasive wearable device that would allow users to control smartphones without touch input. The device was intended to translate neurological signals into actions and to enable silent text entry by mouthing words, drawing on techniques similar to brain–computer interfaces. Early funding was directed toward this hardware-focused effort. After around three years of development, Wispr concluded that contemporary AI systems were not sufficient for the requirements of the wearable device. The company shifted its focus to Flow voice dictation software, the software layer originally built for the wearable, and in 2024 released a macOS application based on this platform. == Wispr Flow == Wispr Flow (often referred to as Flow) is a speech-to-text application for macOS, Windows and iOS. It provides real-time dictation and transcription in more than 100 languages and can operate across applications, including email clients, messaging platforms and chatbots. In June 2025 Wispr released an iOS version that functions as a third-party keyboard, allowing voice input in any app. == Technology == Wispr Flow is based on automatic speech recognition (ASR) and other AI models. The system adapts to individual users over time, learning their vocabulary and preferred style with the aim of reducing manual editing. Flow operates through configurable “Flow Sessions”, defined as time windows during which the app has access to the microphone; users can set session timeouts or disable automatic time limits. == Users and Adoption == Wispr initially targeted users such as venture capitalists, entrepreneurs and executives who process large volumes of text and often work in private or flexible environments. The user base later expanded via platforms such as Product Hunt to students, software developers, writers, lawyers and consultants. Flow has also been adopted by users with conditions such as ADHD, dyslexia, paralysis and carpal tunnel syndrome. About 40% of users are in the United States, 30% in Europe and the remaining 30% in other regions. More than 30% of users come from non-technical backgrounds. Flow supports 104 languages, with approximately 40% of dictations in English and 60% in other languages, including Spanish, French, German, Dutch, Hindi and Mandarin. Wispr has reported monthly user growth above 50%, a six-month active-user retention rate of about 80%, a payment rate around 19%, and revenue of approximately US$3.8 million between July 2024 and July 2025. == Development == Wispr has announced plans for an Android application and maintains waiting lists for Android, Linux and web versions of Flow. The company is developing shared-context features for teams so that the software can recognize common terminology within organizations and has stated that it aims to evolve Flow into a broader AI assistant for tasks such as messaging, note-taking and reminders. Wispr has also reported working with unnamed AI hardware partners on interaction layers for future devices. == Funding == In 2025 Wispr raised US$30 million in a Series A funding round led by Menlo Ventures, with participation from NEA, 8VC and several individual investors, including Evan Sharp and Henry Ward. Earlier investors include Neo, MVP Ventures and AIX Ventures. In November of that same year, the company raised a US$25 million Series A extension led by Notable Capital, with participation from Flight Fund, bringing its total funding to US$81 million. Wispr competes with other AI-based dictation and voice-input tools, including Aqua, Talktastic, Superwhisper and Betterdication.
Subvocal recognition
Subvocal recognition (SVR) is the process of taking subvocalization and converting the detected results to a digital output, aural or text-based. A silent speech interface is a device that allows speech communication without using the sound made when people vocalize their speech sounds. It works by the computer identifying the phonemes that an individual pronounces from nonauditory sources of information about their speech movements. These are then used to recreate the speech using speech synthesis. == Input methods == Silent speech interface systems have been created using ultrasound and optical camera input of tongue and lip movements. Electromagnetic devices are another technique for tracking tongue and lip movements. The detection of speech movements by electromyography of speech articulator muscles and the larynx is another technique. Another source of information is the vocal tract resonance signals that get transmitted through bone conduction called non-audible murmurs. They have also been created as a brain–computer interface using brain activity in the motor cortex obtained from intracortical microelectrodes. == Uses == Such devices are created as aids to those unable to create the sound phonation needed for audible speech such as after laryngectomies. Another use is for communication when speech is masked by background noise or distorted by self-contained breathing apparatus. A further practical use is where a need exists for silent communication, such as when privacy is required in a public place, or hands-free data silent transmission is needed during a military or security operation. In 2002, the Japanese company NTT DoCoMo announced it had created a silent mobile phone using electromyography and imaging of lip movement. The company stated that "the spur to developing such a phone was ridding public places of noise," adding that, "the technology is also expected to help people who have permanently lost their voice." The feasibility of using silent speech interfaces for practical communication has since then been shown. In 2019, Arnav Kapur, a researcher from the Massachusetts Institute of Technology, conducted a study known as AlterEgo. Its implementation of the silent speech interface enables direct communication between the human brain and external devices through stimulation of the speech muscles. By leveraging neural signals associated with speech and language, the AlterEgo system deciphers the user's intended words and translates them into text or commands without the need for audible speech. == Research and patents == With a grant from the U.S. Army, research into synthetic telepathy using subvocalization is taking place at the University of California, Irvine under lead scientist Mike D'Zmura. NASA's Ames Research Laboratory in Mountain View, California, under the supervision of Charles Jorgensen is conducting subvocalization research. The Brain Computer Interface R&D program at Wadsworth Center under the New York State Department of Health has confirmed the existing ability to decipher consonants and vowels from imagined speech, which allows for brain-based communication using imagined speech, however using EEGs instead of subvocalization techniques. US Patents on silent communication technologies include: US Patent 6587729 "Apparatus for audibly communicating speech using the radio frequency hearing effect", US Patent 5159703 "Silent subliminal presentation system", US Patent 6011991 "Communication system and method including brain wave analysis and/or use of brain activity", US Patent 3951134 "Apparatus and method for remotely monitoring and altering brain waves". Latter two rely on brain wave analysis. == In fiction == The decoding of silent speech using a computer played an important role in Arthur C. Clarke's story and Stanley Kubrick's associated film A Space Odyssey. In this, HAL 9000, a computer controlling spaceship Discovery One, bound for Jupiter, discovers a plot to deactivate it by the mission astronauts Dave Bowman and Frank Poole through lip reading their conversations. In Orson Scott Card's series (including Ender's Game), the artificial intelligence can be spoken to while the protagonist wears a movement sensor in his jaw, enabling him to converse with the AI without making noise. He also wears an ear implant. In Speaker for the Dead and subsequent novels, author Orson Scott Card described an ear implant, called a "jewel", that allows subvocal communication with computer systems. Author Robert J. Sawyer made use of subvocal recognition to allow silent commands to the cybernetic 'companion implants' used by the advanced Neanderthal characters in his Neanderthal Parallax trilogy of science fiction novels. In Earth, David Brin depicts this technology and its uses as a normal gear in the near future. In Down and Out in the Magic Kingdom, Cory Doctorow has cellphone technology become silent through a cochlear implant and miking the throat to pick up subvocalization. William Gibson's Sprawl Trilogy frequently uses sub-vocalization systems in various devices. In Kage Baker's Company novels, the immortal cyborgs communicate subvocally. In the Hugo Award-winning Hyperion Cantos by Dan Simmons, the characters often use subvocalization to communicate. In the Culture novels by Iain M. Banks, more highly advanced species often communicate subvocally through their technology. In Deus Ex: Human Revolution (2011), the protagonist is augmented with a subvocalization implant for sending covert communications (and a corresponding cochlear implant for receiving covert communications). In the tabletop RPG and video game series Shadowrun, player characters can communicate via subvocal microphones in some instances. In Paranoia, all citizens can speak to the computer via their "cerebral cortech" implants. Alistair Reynolds Revelation Space trilogy frequently uses sub-vocalization systems in various devices.
Plotting algorithms for the Mandelbrot set
There are many programs and algorithms used to plot the Mandelbrot set and other fractals, some of which are described in fractal-generating software. These programs use a variety of algorithms to determine the color of individual pixels efficiently. == Escape time algorithm == The simplest algorithm for generating a representation of the Mandelbrot set is known as the "escape time" algorithm. A repeating calculation is performed for each x, y point in the plot area and based on the behavior of that calculation, a color is chosen for that pixel. === Unoptimized naïve escape time algorithm === In both the unoptimized and optimized escape time algorithms, the x and y locations of each point are used as starting values in a repeating, or iterating calculation (described in detail below). The result of each iteration is used as the starting values for the next. The values are checked during each iteration to see whether they have reached a critical "escape" condition, or "bailout". If that condition is reached, the calculation is stopped, the pixel is drawn, and the next x, y point is examined. For some starting values, escape occurs quickly, after only a small number of iterations. For starting values very close to but not in the set, it may take hundreds or thousands of iterations to escape. For values within the Mandelbrot set, escape will never occur. The programmer or user must choose how many iterations–or how much "depth"–they wish to examine. The higher the maximal number of iterations, the more detail and subtlety emerge in the final image, but the longer time it will take to calculate the fractal image. Escape conditions can be simple or complex. Because no complex number with a real or imaginary part greater than 2 can be part of the set, a common bailout is to escape when either coefficient exceeds 2. A more computationally complex method that detects escapes sooner, is to compute distance from the origin using the Pythagorean theorem, i.e., to determine the absolute value, or modulus, of the complex number. If this value exceeds 2, or equivalently, when the sum of the squares of the real and imaginary parts exceed 4, the point has reached escape. More computationally intensive rendering variations include the Buddhabrot method, which finds escaping points and plots their iterated coordinates. The color of each point represents how quickly the values reached the escape point. Often black is used to show values that fail to escape before the iteration limit, and gradually brighter colors are used for points that escape. This gives a visual representation of how many cycles were required before reaching the escape condition. To render such an image, the region of the complex plane we are considering is subdivided into a certain number of pixels. To color any such pixel, let c {\displaystyle c} be the midpoint of that pixel. We now iterate the critical point 0 under P c {\displaystyle P_{c}} , checking at each step whether the orbit point has modulus larger than 2. When this is the case, we know that c {\displaystyle c} does not belong to the Mandelbrot set, and we color our pixel according to the number of iterations used to find out. Otherwise, we keep iterating up to a fixed number of steps, after which we decide that our parameter is "probably" in the Mandelbrot set, or at least very close to it, and color the pixel black. In pseudocode, this algorithm would look as follows. The algorithm does not use complex numbers and manually simulates complex-number operations using two real numbers, for those who do not have a complex data type. The program may be simplified if the programming language includes complex-data-type operations. for each pixel (Px, Py) on the screen do x0 := scaled x coordinate of pixel (scaled to lie in the Mandelbrot X scale (-2.00, 0.47)) y0 := scaled y coordinate of pixel (scaled to lie in the Mandelbrot Y scale (-1.12, 1.12)) x := 0.0 y := 0.0 iteration := 0 max_iteration := 1000 while (xx + yy ≤ 22 AND iteration < max_iteration) do xtemp := xx - yy + x0 y := 2xy + y0 x := xtemp iteration := iteration + 1 color := palette[iteration] plot(Px, Py, color) Here, relating the pseudocode to c {\displaystyle c} , z {\displaystyle z} and P c {\displaystyle P_{c}} : z = x + i y {\displaystyle z=x+iy\ } z 2 = x 2 + 2 i x y {\displaystyle z^{2}=x^{2}+2ixy} - y 2 {\displaystyle y^{2}\ } c = x 0 + i y 0 {\displaystyle c=x_{0}+iy_{0}\ } and so, as can be seen in the pseudocode in the computation of x and y: x = R e ( z 2 + c ) = x 2 − y 2 + x 0 {\displaystyle x=\mathop {\mathrm {Re} } (z^{2}+c)=x^{2}-y^{2}+x_{0}} and y = I m ( z 2 + c ) = 2 x y + y 0 . {\displaystyle y=\mathop {\mathrm {Im} } (z^{2}+c)=2xy+y_{0}.\ } To get colorful images of the set, the assignment of a color to each value of the number of executed iterations can be made using one of a variety of functions (linear, exponential, etc.). One practical way, without slowing down calculations, is to use the number of executed iterations as an entry to a palette initialized at startup. If the color table has, for instance, 500 entries, then the color selection is n mod 500, where n is the number of iterations. === Optimized escape time algorithms === The code in the previous section uses an unoptimized inner while loop for clarity. In the unoptimized version, one must perform five multiplications per iteration. To reduce the number of multiplications the following code for the inner while loop may be used instead: x2:= 0 y2:= 0 w:= 0 while (x2 + y2 ≤ 4 and iteration < max_iteration) do x:= x2 - y2 + x0 y:= w - x2 - y2 + y0 x2:= x x y2:= y y w:= (x + y) (x + y) iteration:= iteration + 1 The above code works via some algebraic simplification of the complex multiplication: ( i y + x ) 2 = − y 2 + 2 i y x + x 2 = x 2 − y 2 + 2 i y x {\displaystyle {\begin{aligned}(iy+x)^{2}&=-y^{2}+2iyx+x^{2}\\&=x^{2}-y^{2}+2iyx\end{aligned}}} Using the above identity, the number of multiplications can be reduced to three instead of five. The above inner while loop can be further optimized by expanding w to w = x 2 + 2 x y + y 2 {\displaystyle w=x^{2}+2xy+y^{2}} Substituting w into y = w − x 2 − y 2 + y 0 {\displaystyle y=w-x^{2}-y^{2}+y_{0}} yields y = 2 x y + y 0 {\displaystyle y=2xy+y_{0}} and hence calculating w is no longer needed. The further optimized pseudocode for the above is: x:= 0 y:= 0 x2:= 0 y2:= 0 while (x2 + y2 ≤ 4 and iteration < max_iteration) do x2:= x x y2:= y y y:= 2 x y + y0 x:= x2 - y2 + x0 iteration:= iteration + 1 Note that in the above pseudocode, 2 x y {\displaystyle 2xy} seems to increase the number of multiplications by 1, but since 2 is the multiplier the code can be optimized via ( x + x ) y {\displaystyle (x+x)y} . == Coloring algorithms == In addition to plotting the set, a variety of algorithms have been developed to efficiently color the set in an aesthetically pleasing way show structures of the data (scientific visualisation) === Histogram coloring === A more complex coloring method involves using a histogram which pairs each pixel with said pixel's maximum iteration count before escape/bailout. This method will equally distribute colors to the same overall area, and, importantly, is independent of the maximum number of iterations chosen. This algorithm has four passes. The first pass involves calculating the iteration counts associated with each pixel (but without any pixels being plotted). These are stored in an array IterationCounts[x][y], where x and y are the x and y coordinates of said pixel on the screen respectively. The first step of the second pass is to create an array NumIterationsPerPixel[n], where the array size n is the maximum iteration count. Next, one must iterate over the array of pixel-iteration count pairs IterationCounts[x][y], and retrieve each pixel's saved iteration count, i, via e.g. i = IterationCounts[x][y]. After each pixel's iteration count i is retrieved, it is necessary to index the NumIterationsPerPixel array at i and increment the indexed value (which is initially zero) -- e.g. NumIterationsPerPixel[i] = NumIterationsPerPixel[i] + 1. for (x = 0; x < width; x++) do for (y = 0; y < height; y++) do i:= IterationCounts[x][y] NumIterationsPerPixel[i]++ The third pass iterates through the NumIterationsPerPixel array and adds up all the stored values, saving them in total. The array index represents the number of pixels that reached that iteration count before bailout. total: = 0 for (i = 0; i < max_iterations; i++) do total += NumIterationsPerPixel[i] After this, the fourth pass begins and all the values in the IterationCounts array are indexed, and, for each iteration count i, associated with each pixel, the count is added to a global sum of all the iteration counts from 1 to i in the NumIterationsPerPixel array . This value is then normalized by dividing the sum by the total value computed earlier. hue[][]:= 0.0 for (x = 0; x < width; x++) do for (y = 0; y < height; y++) do iteration:= Iteration
CapCut
CapCut, known domestically as JianYing (Chinese: 剪映; pinyin: Jiǎnyìng) and formerly internationally as ViaMaker, is a video editor developed by ByteDance, available as a mobile app, desktop app, and web app. == History == The app was first released in China in 2019 and was initially available for iPhone and Android. In 2020, it was rebranded in English from ViaMaker to CapCut and became available globally. It later expanded to include web and desktop versions for Mac and Windows. In 2022, CapCut reached 200 million active users. According to The Wall Street Journal, in March 2023, it was the second-most downloaded app in the U.S., behind that of Chinese discount retailer Temu. In January 2025, CapCut had over 1 billion downloads on the Google Play Store. On February 1, 2021, CapCut Pro for Windows was launched. On November 27, the Pro version for Mac was launched. In July 2025, CapCut Pro for HarmonyOS was available on HarmonyOS NEXT tablets. In July 2024, CapCut was reported by the South China Morning Post to be a generative AI (GenAI) application that led global AI app downloads, with approximately 38.42 million downloads and 323 million monthly active users. == Features == CapCut supports basic video editing functions, including editing, trimming, and adding or splitting clips. Editing projects is limited to single-layer editing, but the app supports overlay options that enable additional effects, including multi-layer editing. The app includes a library of pre-made templates and a tool that generates editable video captions. It also provides photo editing tools, including retouch and product photo features integrated within the editing interface. CapCut's video editor includes AI-based features such as video and script generation. Users can export or save completed projects directly to different social media platforms. CapCut includes a free version and a paid Pro version with cloud storage and advanced features. == Controversies == === Illegal data collection === In July 2023, many users of CapCut accused it of illegally profiting off their personal data. A class-action lawsuit filed in the U.S. District Court for the Northern District of Illinois on July 28, 2023, alleged that CapCut illegally harvests and profits from user data including biometric information and geolocation without consent. In September 2025, a federal court excluded most of the lawsuit, which alleged that TikTok’s parent company improperly scraped private data from CapCut's video editing software, as lacking grounds, with some of the class action continuing to move forward. == Bans and restrictions == === Ban in India === As a response to border clashes with China in May 2020, the Indian government banned around 56 Chinese applications including CapCut and TikTok, which is owned by CapCut's parent company ByteDance. Indian users were unable to use and download the application. As of February 2022, around 273 Chinese applications have been banned by the Indian government under the concern of national security and Indian user privacy. === Ban in the United States === On January 18, 2025, at 10 PM EST, CapCut was banned in the United States along with TikTok and all other ByteDance apps due to the implementation of the Protecting Americans from Foreign Adversary Controlled Applications Act. Hours after the suspension of services took effect, President Donald Trump indicated on Truth Social that he would issue an executive order on the day of his inauguration "to extend the period of time before the law's prohibitions take effect". On January 21, CapCut began restoring service. On February 13, Google and Apple restored CapCut on the App Store and Google Play Store.
PatchMatch
PatchMatch is an algorithm used to quickly find correspondences (or matches) between small square regions (or patches) of an image. It has various applications in image editing, such as reshuffling or removing objects from images or altering their aspect ratios without cropping or noticeably stretching them. PatchMatch was first presented in a 2011 paper by researchers at Princeton University. == Algorithm == The goal of the algorithm is to find the patch correspondence by defining a nearest-neighbor field (NNF) as a function f : R 2 → R 2 {\displaystyle f:\mathbb {R} ^{2}\to \mathbb {R} ^{2}} of offsets, which is over all possible matches of patch (location of patch centers) in image A, for some distance function of two patches D {\displaystyle D} . So, for a given patch coordinate a {\displaystyle a} in image A {\displaystyle A} and its corresponding nearest neighbor b {\displaystyle b} in image B {\displaystyle B} , f ( a ) {\displaystyle f(a)} is simply b − a {\displaystyle b-a} . However, if we search for every point in image B {\displaystyle B} , the work will be too hard to complete. So the following algorithm is done in a randomized approach in order to accelerate the calculation speed. The algorithm has three main components. Initially, the nearest-neighbor field is filled with either random offsets or some prior information. Next, an iterative update process is applied to the NNF, in which good patch offsets are propagated to adjacent pixels, followed by random search in the neighborhood of the best offset found so far. Independent of these three components, the algorithm also uses a coarse-to-fine approach by building an image pyramid to obtain the better result. === Initialization === When initializing with random offsets, we use independent uniform samples across the full range of image B {\displaystyle B} . This algorithm avoids using an initial guess from the previous level of the pyramid because in this way the algorithm can avoid being trapped in local minima. === Iteration === After initialization, the algorithm attempted to perform iterative process of improving the N N F {\displaystyle NNF} . The iterations examine the offsets in scan order (from left to right, top to bottom), and each undergoes propagation followed by random search. === Propagation === We attempt to improve f ( x , y ) {\displaystyle f(x,y)} using the known offsets of f ( x − 1 , y ) {\displaystyle f(x-1,y)} and f ( x , y − 1 ) {\displaystyle f(x,y-1)} , assuming that the patch offsets are likely to be the same. That is, the algorithm will take new value for f ( x , y ) {\displaystyle f(x,y)} to be arg min ( x , y ) D ( f ( x , y ) ) , D ( f ( x − 1 , y ) ) , D ( f ( x , y − 1 ) ) {\displaystyle \arg \min \limits _{(x,y)}{D(f(x,y)),D(f(x-1,y)),D(f(x,y-1))}} . So if f ( x , y ) {\displaystyle f(x,y)} has a correct mapping and is in a coherent region R {\displaystyle R} , then all of R {\displaystyle R} below and to the right of f ( x , y ) {\displaystyle f(x,y)} will be filled with the correct mapping. Alternatively, on even iterations, the algorithm search for different direction, fill the new value to be arg min ( x , y ) { D ( f ( x , y ) ) , D ( f ( x + 1 , y ) ) , D ( f ( x , y + 1 ) ) } {\displaystyle \arg \min \limits _{(x,y)}\{D(f(x,y)),D(f(x+1,y)),D(f(x,y+1))\}} . === Random search === Let v 0 = f ( x , y ) {\displaystyle v_{0}=f(x,y)} , we attempt to improve f ( x , y ) {\displaystyle f(x,y)} by testing a sequence of candidate offsets at an exponentially decreasing distance from v 0 {\displaystyle v_{0}} u i = v 0 + w α i R i {\displaystyle u_{i}=v_{0}+w\alpha ^{i}R_{i}} where R i {\displaystyle R_{i}} is a uniform random in [ − 1 , 1 ] × [ − 1 , 1 ] {\displaystyle [-1,1]\times [-1,1]} , w {\displaystyle w} is a large window search radius which will be set to maximum picture size, and α {\displaystyle \alpha } is a fixed ratio often assigned as 1/2. This part of the algorithm allows the f ( x , y ) {\displaystyle f(x,y)} to jump out of local minimum through random process. === Halting criterion === The often used halting criterion is set the iteration times to be about 4~5. Even with low iteration, the algorithm works well.
Speech recognition
Speech recognition (automatic speech recognition (ASR), computer speech recognition, or speech-to-text (STT)) is a sub-field of computational linguistics concerned with methods and technologies that translate spoken language into text or other interpretable forms. Speech recognition applications include voice user interfaces, where the user speaks to a device, which "listens" and processes the audio. Common voice applications include interpreting commands for calling, call routing, home automation, and aircraft control. These applications are called direct voice input. Productivity applications include searching audio recordings, creating transcripts, and dictation. Speech recognition can be used to analyse speaker characteristics, such as identifying native language using pronunciation assessment. Voice recognition (speaker identification) refers to identifying the speaker, rather than speech contents. Recognizing the speaker can simplify the task of translating speech in systems trained on a specific person's voice. It can also be used to authenticate the speaker as part of a security process. == History == Applications for speech recognition developed over many decades, with progress accelerated due to advances in deep learning and the use of big data. These advances are reflected in an increase in academic papers, and greater system adoption. Key areas of growth include vocabulary size, more accurate recognition for unfamiliar speakers (speaker independence), and faster processing speed. === Pre-1970 === 1952 – Bell Labs researchers, Stephen Balashek, R. Biddulph, and K. H. Davis, built Audrey for single-speaker digit recognition. Their system located the formants in the power spectrum of each utterance. 1960 – Gunnar Fant developed and published the source–filter model of speech production. 1962 – IBM's 16-word "Shoebox" machine's speech recognition debuted at the 1962 World's Fair. 1966 – Linear predictive coding, a speech coding method, was proposed by Fumitada Itakura of Nagoya University and Shuzo Saito of Nippon Telegraph and Telephone. 1969 – Funding at Bell Labs came to a halt for several years after the company's head engineer, John R. Pierce, wrote an open letter criticizing speech recognition research. This defunding lasted until Pierce retired and James L. Flanagan took over. Raj Reddy was the first person to work on continuous speech recognition, as a graduate student at Stanford University in the late 1960s. Previous systems required users to pause after each word. Reddy's system issued spoken commands for playing chess. Around this time, Soviet researchers invented the dynamic time warping (DTW) algorithm and used it to create a recognizer capable of operating on a 200-word vocabulary. DTW processed speech by dividing it into short frames (e.g. 10 ms segments) and treating each frame as a unit. Speaker independence, however, remained unsolved. === 1970–1990 === 1971 – DARPA funded a five-year speech recognition research project, Speech Understanding Research, seeking a minimum vocabulary size of 1,000 words. The project considered speech understanding a key to achieving progress in speech recognition, which was later disproved. BBN, IBM, Carnegie Mellon (CMU), and Stanford Research Institute participated. 1972 – The IEEE Acoustics, Speech, and Signal Processing group held a conference in Newton, Massachusetts. 1976 – The first ICASSP was held in Philadelphia, which became a major venue for publishing on speech recognition. During the late 1960s, Leonard Baum developed the mathematics of Markov chains at the Institute for Defense Analysis. A decade later, at CMU, Raj Reddy's students James Baker and Janet M. Baker began using the hidden Markov model (HMM) for speech recognition. James Baker had learned about HMMs while at the Institute for Defense Analysis. HMMs enabled researchers to combine sources of knowledge, such as acoustics, language, and syntax, in a unified probabilistic model. By the mid-1980s, Fred Jelinek's team at IBM created a voice-activated typewriter called Tangora, which could handle a 20,000-word vocabulary. Jelinek's statistical approach placed less emphasis on emulating human brain processes in favor of statistical modelling. (Jelinek's group independently discovered the application of HMMs to speech.) This was controversial among linguists since HMMs are too simplistic to account for many features of human languages. However, the HMM proved to be a highly useful way for modelling speech and replaced dynamic time warping as the dominant speech recognition algorithm in the 1980s. 1982 – Dragon Systems, founded by James and Janet M. Baker, was one of IBM's few competitors. === Practical speech recognition === The 1980s also saw the introduction of the n-gram language model. 1987 – The back-off model enabled language models to use multiple-length n-grams, and CSELT used HMM to recognize languages (in software and hardware, e.g. RIPAC). At the end of the DARPA program in 1976, the best computer available to researchers was the PDP-10 with 4 MB of RAM. It could take up to 100 minutes to decode 30 seconds of speech. Practical products included: 1984 – the Apricot Portable was released with up to 4096 words support, of which only 64 could be held in RAM at a time. 1987 – a recognizer from Kurzweil Applied Intelligence 1990 – Dragon Dictate, a consumer product released in 1990. AT&T deployed the Voice Recognition Call Processing service in 1992 to route telephone calls without a human operator. The technology was developed by Lawrence Rabiner and others at Bell Labs. By the early 1990s, the vocabulary of the typical commercial speech recognition system had exceeded the average human vocabulary. Reddy's former student, Xuedong Huang, developed the Sphinx-II system at CMU. Sphinx-II was the first to do speaker-independent, large vocabulary, continuous speech recognition, and it won DARPA's 1992 evaluation. Handling continuous speech with a large vocabulary was a major milestone. Huang later founded the speech recognition group at Microsoft in 1993. Reddy's student Kai-Fu Lee joined Apple, where, in 1992, he helped develop the Casper speech interface prototype. Lernout & Hauspie, a Belgium-based speech recognition company, acquired other companies, including Kurzweil Applied Intelligence in 1997 and Dragon Systems in 2000. L&H was used in Windows XP. L&H was an industry leader until an accounting scandal destroyed it in 2001. L&H speech technology was bought by ScanSoft, which became Nuance in 2005. Apple licensed Nuance software for its digital assistant Siri. ==== 2000s ==== In the 2000s, DARPA sponsored two speech recognition programs: Effective Affordable Reusable Speech-to-Text (EARS) in 2002, followed by Global Autonomous Language Exploitation (GALE) in 2005. Four teams participated in EARS: IBM; a team led by BBN with LIMSI and the University of Pittsburgh; Cambridge University; and a team composed of ICSI, SRI, and the University of Washington. EARS funded the collection of the Switchboard telephone speech corpus, which contained 260 hours of recorded conversations from over 500 speakers. The GALE program focused on Arabic and Mandarin broadcast news. Google's first effort at speech recognition came in 2007 after recruiting Nuance researchers. Its first product, GOOG-411, was a telephone-based directory service. Since at least 2006, the U.S. National Security Agency has employed keyword spotting, allowing analysts to index large volumes of recorded conversations and identify speech containing "interesting" keywords. Other government research programs focused on intelligence applications, such as DARPA's EARS program and IARPA's Babel program. In the early 2000s, speech recognition was dominated by hidden Markov models combined with feed-forward artificial neural networks (ANN). Later, speech recognition was taken over by long short-term memory (LSTM), a recurrent neural network (RNN) published by Sepp Hochreiter & Jürgen Schmidhuber in 1997. LSTM RNNs avoid the vanishing gradient problem and can learn "Very Deep Learning" tasks that require memories of events that happened thousands of discrete time steps earlier, which is important for speech. Around 2007, LSTMs trained with Connectionist Temporal Classification (CTC) began to outperform. In 2015, Google reported a 49 percent error‑rate reduction in its speech recognition via CTC‑trained LSTM. Transformers, a type of neural network based solely on attention, were adopted in computer vision and language modelling, and then to speech recognition. Deep feed-forward (non-recurrent) networks for acoustic modelling were introduced in 2009 by Geoffrey Hinton and his students at the University of Toronto, and by Li Deng and colleagues at Microsoft Research. In contrast to the prioer incremental improvements, deep learning decreased error rates by 30%. Both shallow and deep forms (e.g., recurrent nets) of ANNs had been explored since the 1980s. Howev