The Imix (also known as ImMix) Video Cube is one of the first computer non-linear editing systems that was a full broadcast quality online video finishing machine. After its release in 1994, Imix released a more advanced version, the Imix Turbo Cube, which boasted 4 channels of real time layered visual effects. It was a hardware computer system controlled by an Apple Macintosh computer.
Robot Monk Xian'er
Robot Monk Xian'er (Chinese: 贤二机器僧) is a humanoid robot based on the cartoon character Xian'er. It was developed by a team of monks, volunteers and AI experts from Beijing Longquan Monastery in Beijing, China. He can follow human instructions to make body movements, read scriptures and play Buddhist music. He can chat and respond to people's emotional and spiritual questions with Buddhist wisdom. As a chatbot, Robot Monk Xian'er is available on certain public platforms including WeChat and Facebook. Over the years, master Xuecheng, the abbot of Beijing Longquan Monastery, replied to thousands of questions on Sina Weibo. These questions and their answers become the data source of the chatbot.
TCEC Season 14
The 14th season of the Top Chess Engine Championship took place between 17 November 2018 and 24 February 2019. Stockfish was the defending champion, having defeated Komodo in the previous season's superfinal. The season is notable for two things: the emergence of two strong, new engines, the Komodo variant Komodo Monte Carlo tree search (MCTS) and the neural network engine Leela Chess Zero, and the dramatic superfinal. Komodo MCTS and Leela fought their way from Division 4 and Division 3 respectively to the Premier Division, with Leela further qualifying for the superfinal against Stockfish. The superfinal was a topsy-turvy affair with the lead changing hands several times. It finished as the closest superfinal TCEC has ever seen, with Stockfish winning by a single game, 50.5–49.5 (+10 =81 -9). == Overview == === Structure === The season comprised five divisions: from the lowest Division 4 to the Premier Division. The top two engines of each division promote to the division above, while the bottom two engines relegate. The top two engines of the Premier Division contest a 100-game superfinal. The lengths of the opening books used increases as the divisions progress. The superfinal itself used a custom opening book designed by Jeroen Noomen. === Rules === The TCEC draw and win rules were slightly modified for Season 14. The game is now adjudicated as drawn if, after move 30, both engines have evals ±0.08 for five consecutive moves, and there are neither pawn moves nor a capture. Win adjudication now occurs if both engines have an eval of ±10 for five consecutive moves. Following the controversy over DeusX's participation last season, the uniqueness rule for neural networks was modified such that at least two of the following three hallmarks must be unique: The code for training the neural network The neural network (and weights file) itself The engine that executes this network This change meant DeusX did not meet the uniqueness criteria and therefore did not participate. Aside from this change, the season used the standard rules of the TCEC. == Results == === Division 4 === New entrant Komodo MCTS dominated Division 4, winning by a clear four points, although it did lose a game to second-place finisher rofChade. Fellow new entrant Scorpio NN performed badly and finished last, drawing only one game and losing the rest. === Division 3 === The neural network engine Leela Chess Zero had just missed promotion to Division 2 in the previous season. Since its relatively weak performance last season was partly due to hardware problems, and since it had shown a lot of improvement in strength, it was the hot favourite in this division. Leela lived up to its billing by comprehensively defeating everyone else. In a portent of future divisions however, Leela surprisingly dropped a game to third-place Arasan. Komodo MCTS was also improving quickly, and an updated version finished second behind Leela. The gap between second and third was 6.5 points, illustrating the gulf in class. === Division 2 === Although Division 2 engines are significantly stronger than Division 3, Leela and Komodo MCTS continued to dominate the competition, and again finished first and second. Komodo MCTS only lost one game to Leela, while Leela's tendency to occasionally lose to weaker engines saw her losing a game to 4th-placed Booot. Third place finisher Xiphos gave Leela and Komodo MCTS a run for their money, and was in the running up until the final rounds when it lost a crucial game to Leela. This loss left it one point behind Komodo MCTS in the final standings. === Division 1 === Leela and Komodo MCTS's rampage through the lower divisions continued, and they again finished first and second. In a demonstration of how much it had improved, Leela scored 20/28 in this division, the same score it had achieved in Division 2. This was also a TCEC points record for this division. However, Leela dropped a game against fourth-place finisher Chiron. Komodo MCTS, which had yet to lose a game in the lower divisions except to Leela, also conceded its first loss to third-place Fizbo. At the other end of the table, former champions Jonny and Fritz, which had not been updated, found themselves outclassed and finished second-last and last respectively; however with fellow competitor Ginkgo crashing five times (and therefore being disqualified), Jonny managed to stay in the division. The penultimate game for this division set a new TCEC moves record for a decisive game: 308 moves before Leela defeated Fritz. === Premier division === This was the strongest premier division ever, with multiple-time champions Stockfish, Komodo, and Houdini in the mix. Right from the start it became clear that Stockfish was in a league of its own, and it dominated the division, scoring wins against every other engine without losing a game. Second place however was a hotly-contested affair, with Leela, Komodo and Houdini neck-and-neck for most of the division. Houdini took the early lead, but Komodo gained second after winning two games by forfeit when its sibling Komodo MCTS crashed. This led to murmurs of a "Konspiracy". However, when both Komodo and Houdini failed to score more wins against the lower half of the field, Leela was able to take the lead. Halfway through the division the race was upended again when Leela went through a bad streak, losing three games in a row to Stockfish, Komodo, and Fire. This led to Komodo regaining second place, only for Komodo MCTS to crash yet again. By TCEC rules this meant Komodo MCTS was disqualified and all its scores were zeroed out, which put Leela back in second place. With three games left, Leela missed a win against Andscacs, which would've more or less secured her a place in the superfinal. Meanwhile, Komodo kept the division interesting by winning two of its last three games. Because Komodo had superior tiebreakers to Leela, this meant Komodo would qualify for the superfinal unless Leela managed to hold Stockfish to a draw with Black in the last game of the division. In a tense final game, Stockfish came close to winning, but missed the winning line. Leela managed to draw and qualified for the superfinal. At the other end of the table, it was quickly apparent that Ethereal and Andscacs were the weakest engines and would likely relegate. However, when Komodo MCTS was disqualified (and therefore relegated), it threw both engines a lifeline, since they could now stay in the division by beating the other. Andscacs was able to score a head-to-head win against Ethereal, but was crushed by Stockfish (+0 =2 -4) and Leela (+0 =3 -3). Ethereal didn't manage to score a win in the entire division, but did manage to score more draws than Andscacs, condemning Andscacs to relegation. === Superfinal === Going into the superfinal expectations were high for Leela: she had received a new network and had just won her first major competition when she defeated Houdini in the second TCEC cup. However, she had won the tournament without having played Stockfish (who had been surprisingly eliminated by Houdini in the semifinals). That, plus the fact that Stockfish dominated Premier Division and had never lost a match to Leela, left it unclear which engine was superior, although most spectators favored Stockfish. The superfinal turned out to be a roller-coaster. It began with Stockfish drawing first blood in game 7, and then scoring another win in game 10. Leela hit back with wins in game 11 and 13, but then lost games 20, 21, and 22. This gave Stockfish a 3-point lead. However, in the next 30 games, Leela was the only one to score wins: it first equalized by winning games 25, 27, and 29, and then took the lead by winning games 49 and 53. Stockfish won game 56, but Leela won game 63, maintaining her lead. There followed two dramatic games. In game 65, Leela built up a winning position. Stockfish showed a +153 evaluation, indicating that it had found a forced line leading to an endgame tablebase win; indeed analysis with 7-piece tablebases showed that Leela's position was winning. Under previous seasons' rules, the game would have been adjudicated as a win because Leela's evaluation was above 6.5. However under the new rules, Leela's +8.92 evaluation was not enough to adjudicate. It turned out that Leela could not see the winning line, and shuffled her pieces aimlessly, leading to a 50-move draw. In game 66, Stockfish was given a substantial advantage by the opening, but failed to make the most of it. The evaluations were leveling out to zero when the internet connection to the GPU servers was cut off. By tournament rules, this meant the game was replayed from scratch. After a further internet disconnection and restart, Stockfish handled the opening better and won, leaving Leela with a 1-point lead. In the last third of the superfinal, there followed more drama as Leela often built up strong advantages, but Stockfish showed great resourcefulness in defending inferior positions. Meanwh
Speech synthesis
Speech synthesis is the artificial production of human speech. A computer system used for this purpose is called a speech synthesizer, and can be implemented in software or hardware products. A text-to-speech (TTS) system converts normal language text into speech; other systems render symbolic linguistic representations like phonetic transcriptions into speech. The reverse process is speech recognition. Synthesized speech can be created by concatenating pieces of recorded speech that are stored in a database. Systems differ in the size of the stored speech units; a system that stores phones or diphones provides the largest output range, but may lack clarity. For specific usage domains, the storage of entire words or sentences allows for high-quality output. Alternatively, a synthesizer can incorporate a model of the vocal tract and other human voice characteristics to create a completely "synthetic" voice output. The quality of a speech synthesizer is judged by its similarity to the human voice and by its ability to be understood clearly. An intelligible text-to-speech program allows people with visual impairments or reading disabilities to listen to written words on a home computer. The earliest computer operating system to have included a speech synthesizer was Unix in 1974, through the Unix speak utility. In 2000, Microsoft Sam was the default text-to-speech voice synthesizer used by the narrator accessibility feature, which shipped with all Windows 2000 operating systems, and subsequent Windows XP systems. A text-to-speech system (or "engine") is composed of two parts: a front-end and a back-end. The front-end has two major tasks. First, it converts raw text containing symbols like numbers and abbreviations into the equivalent of written-out words. This process is often called text normalization, pre-processing, or tokenization. The front-end then assigns phonetic transcriptions to each word, and divides and marks the text into prosodic units, like phrases, clauses, and sentences. The process of assigning phonetic transcriptions to words is called text-to-phoneme or grapheme-to-phoneme conversion. Phonetic transcriptions and prosody information together make up the symbolic linguistic representation that is output by the front-end. The back-end—often referred to as the synthesizer—then converts the symbolic linguistic representation into sound. In certain systems, this part includes the computation of the target prosody (pitch contour, phoneme durations), which is then imposed on the output speech. == History == Long before the invention of electronic signal processing, some people tried to build machines to emulate human speech. There were also legends of the existence of "Brazen Heads", such as those involving Pope Silvester II (d. 1003 AD), Albertus Magnus (1198–1280), and Roger Bacon (1214–1294). In 1779, the German-Danish scientist Christian Gottlieb Kratzenstein won the first prize in a competition announced by the Russian Imperial Academy of Sciences and Arts for models he built of the human vocal tract that could produce the five long vowel sounds (in International Phonetic Alphabet notation: [aː], [eː], [iː], [oː] and [uː]). There followed the bellows-operated "acoustic-mechanical speech machine" of Wolfgang von Kempelen of Pressburg, Hungary, described in a 1791 paper. This machine added models of the tongue and lips, enabling it to produce consonants as well as vowels. In 1837, Charles Wheatstone produced a "speaking machine" based on von Kempelen's design, and in 1846, Joseph Faber exhibited the "Euphonia". In 1923, Paget resurrected Wheatstone's design. In the 1930s, Bell Labs developed the vocoder, which automatically analyzed speech into its fundamental tones and resonances. From his work on the vocoder, Homer Dudley developed a keyboard-operated voice-synthesizer called The Voder (Voice Demonstrator), which he exhibited at the 1939 New York World's Fair. Franklin S. Cooper and his colleagues at Haskins Laboratories built the pattern playback in the late 1940s and completed it in 1950. There were several different versions of this hardware device; only one currently survives. The machine converts pictures of the acoustic patterns of speech in the form of a spectrogram back into sound. Using this device, Alvin Liberman and colleagues discovered acoustic cues for the perception of phonetic segments (consonants and vowels). === Electronic devices === The first computer-based speech-synthesis systems originated in the late 1950s. Noriko Umeda et al. developed the first general English text-to-speech system in 1968, at the Electrotechnical Laboratory in Japan. In 1961, physicist John Larry Kelly, Jr and his colleague Louis Gerstman used an IBM 704 computer to synthesize speech, an event among the most prominent in the history of Bell Labs. Kelly's voice recorder synthesizer (vocoder) recreated the song "Daisy Bell", with musical accompaniment from Max Mathews. Coincidentally, Arthur C. Clarke was visiting his friend and colleague John Pierce at the Bell Labs Murray Hill facility. Clarke was so impressed by the demonstration that he used it in the climactic scene of his screenplay for his novel 2001: A Space Odyssey, where the HAL 9000 computer sings the same song as astronaut Dave Bowman puts it to sleep. Despite the success of purely electronic speech synthesis, research into mechanical speech-synthesizers continues. Linear predictive coding (LPC), a form of speech coding, began development with the work of Fumitada Itakura of Nagoya University and Shuzo Saito of Nippon Telegraph and Telephone (NTT) in 1966. Further developments in LPC technology were made by Bishnu S. Atal and Manfred R. Schroeder at Bell Labs during the 1970s. LPC was later the basis for early speech synthesizer chips, such as the Texas Instruments LPC Speech Chips used in the Speak & Spell toys from 1978. In 1975, Fumitada Itakura developed the line spectral pairs (LSP) method for high-compression speech coding, while at NTT. From 1975 to 1981, Itakura studied problems in speech analysis and synthesis based on the LSP method. In 1980, his team developed an LSP-based speech synthesizer chip. LSP is an important technology for speech synthesis and coding, and in the 1990s was adopted by almost all international speech coding standards as an essential component, contributing to the enhancement of digital speech communication over mobile channels and the internet. In 1975, MUSA was released, and was one of the first Speech Synthesis systems. It consisted of a stand-alone computer hardware and a specialized software that enabled it to read Italian. A second version, released in 1978, was also able to sing Italian in an "a cappella" style. Dominant systems in the 1980s and 1990s were the DECtalk system, based largely on the work of Dennis Klatt at MIT, and the Bell Labs system; the latter was one of the first multilingual language-independent systems, making extensive use of natural language processing methods. Handheld electronics featuring speech synthesis began emerging in the 1970s. One of the first was the Telesensory Systems Inc. (TSI) Speech+ portable calculator for the blind in 1976. Other devices had primarily educational purposes, such as the Speak & Spell toy produced by Texas Instruments in 1978. Fidelity released a speaking version of its electronic chess computer in 1979. The first video game to feature speech synthesis was the 1980 shoot 'em up arcade game, Stratovox (known in Japan as Speak & Rescue), from Sun Electronics. The first personal computer game with speech synthesis was Manbiki Shoujo (Shoplifting Girl), released in 1980 for the PET 2001, for which the game's developer, Hiroshi Suzuki, developed a "zero cross" programming technique to produce a synthesized speech waveform. Another early example, the arcade version of Berzerk, also dates from 1980. The Milton Bradley Company produced the first multi-player electronic game using voice synthesis, Milton, in the same year. In 1976, Computalker Consultants released their CT-1 Speech Synthesizer. Designed by D. Lloyd Rice and Jim Cooper, it was an analog synthesizer built to work with microcomputers using the S-100 bus standard. Synthesized voices typically sounded male until 1990, when Ann Syrdal, at AT&T Bell Laboratories, created a female voice. Ray Kurzweil predicted in 2005 that as the cost-performance ratio caused speech synthesizers to become cheaper and more accessible, more people would benefit from the use of text-to-speech programs. === Artificial intelligence === In September 2016, DeepMind released WaveNet, which demonstrated that deep learning models are capable of modeling raw waveforms and generating speech from acoustic features like spectrograms or mel-spectrograms, starting the field of deep learning speech synthesis. Although WaveNet was initially considered to be computationally expensive and slow to be used in consumer products at the time, a year after its
Multi Autonomous Ground-robotic International Challenge
The Multi Autonomous Ground-robotic International Challenge (MAGIC) is a 1.6 million dollar prize competition for autonomous mobile robots funded by TARDEC and the DSTO, the primary research organizations for Tank and Defense research in the United States and Australia respectively. The goal of the competition is to create multi-vehicle robotic teams that can execute an intelligence, surveillance and reconnaissance mission in a dynamic urban environment. The challenge required competitors to map a 500 m x 500 m challenge area in under 3.5 hours and to correctly locate, classify and recognise all simulated threats. The challenge event was conducted in Adelaide, Australia, during November 2010. == Competitors == Initially 12 teams were selected for the competition in November 2009, of which 10 teams received funding. These included: MAGICian – Adelaide/Perth, Australia (UWA, ECU, Flinders, Thales) Strategic Engineering – Adelaide, Australia (U. Adelaide) Northern Hunters – Canada (Royal Military College of Canada) Chiba Team – Japan (Chiba University) Cappadocia – Ankara, Turkey (ASELSAN, Ohio State University) RASR – Gaithersburg, Md. (Robotics Research, LLC; QinetiQ; Embry-Riddle Aeronautical University) Team Cornell – US (Cornell University) Team Michigan – Ann Arbor, Mich. (University of Michigan) Virginia Tech – US (Virginia Tech) University of Pennsylvania – Philadelphia (University of Pennsylvania) Numinence – Brisbane, Australia (Numinence Pty Ltd, La Trobe University) UNSW – Sydney, Australia (UNSW) The first downselection trial required teams to map an indoor area and outdoor area, and to demonstrate distributing and handing over tasks between robots. During the first downselection trial, the top six teams were selected: Cappadocia – Ankara, Turkey MAGICian – Adelaide/Perth, Australia RASR – Gaithersburg, Md. Team Michigan – Ann Arbor, Mich. University of Pennsylvania – Philadelphia Chiba Team – Japan Before the finals were held, Chiba Team withdrew from the competition, leaving five competitors. == Event == Ultimately the overall goal of fully autonomous operations without human intervention was not achieved, however, the Secretary for Defence stated "The competing vehicles demonstrated new advances in robotics technology, which are very promising for their potential deployment in combat zones where they can replace our troops in carrying out life-threatening tasks" and considered the competition a success. == Results == The official results of the competition were: First – Team Michigan ($750,000 prize) Second – University of Pennsylvania ($250,000 prize) Third – RASR ($100,000 prize) Fourth – MAGICian & Cappadocia The "Old Ram Shed Challenge" was a single-day competition held after the completion of MAGIC. It was smaller in scale, allowing all of the teams to demonstrate their systems during a single day. The University of Pennsylvania won this challenge, having found a greater number of the target objects than the other teams. == Technology == Key technology used by all teams was computer vision, sensor fusion, human-robot interaction, and simultaneous localization and mapping (SLAM). Team Michigan, a collaboration between the University of Michigan's APRIL Lab and Soar Technology, Inc., had the largest fleet of 14 robots, developed their own Inertial Measurement Unit, and created their skid steer robot chassis out of Baltic birch plywood. Additionally, they had minimal reliance on GPS and used bandwidth limited 900 MHz radios for all telemetry, imaging, and status communications between all robots and the ground station. The code was written primarily in Java and each robot was equipped with an actuated 2D LIDAR, along with a unique 2D barcode for inter-robot recognition. The University of Pennsylvania team consisted of only four members. All code was written using Matlab. The robots were equipped with omnidirectional vision. RASR used the Foster-Miller TALON vehicle. MAGICian used the WAMbot robots developed by The University of Western Australia, Edith Cowan University and Thales Australia. Code was written in C++ and Java. The robots were equipped with SICK laser scanners. See the September/October 2012 special issue of the Journal of Field Robotics for contest highlights, technical approaches taken by several of the teams, and an explanation of the evaluation metrics used by organizers.
Cross-entropy method
The cross-entropy (CE) method is a Monte Carlo method for importance sampling and optimization. It is applicable to both combinatorial and continuous problems, with either a static or noisy objective. The method approximates the optimal importance sampling estimator by repeating two phases: Draw a sample from a probability distribution. Minimize the cross-entropy between this distribution and a target distribution to produce a better sample in the next iteration. Reuven Rubinstein developed the method in the context of rare-event simulation, where tiny probabilities must be estimated, for example in network reliability analysis, queueing models, or performance analysis of telecommunication systems. The method has also been applied to the traveling salesman, quadratic assignment, DNA sequence alignment, max-cut and buffer allocation problems. == Estimation via importance sampling == Consider the general problem of estimating the quantity ℓ = E u [ H ( X ) ] = ∫ H ( x ) f ( x ; u ) d x {\displaystyle \ell =\mathbb {E} _{\mathbf {u} }[H(\mathbf {X} )]=\int H(\mathbf {x} )\,f(\mathbf {x} ;\mathbf {u} )\,{\textrm {d}}\mathbf {x} } , where H {\displaystyle H} is some performance function and f ( x ; u ) {\displaystyle f(\mathbf {x} ;\mathbf {u} )} is a member of some parametric family of distributions. Using importance sampling this quantity can be estimated as ℓ ^ = 1 N ∑ i = 1 N H ( X i ) f ( X i ; u ) g ( X i ) {\displaystyle {\hat {\ell }}={\frac {1}{N}}\sum _{i=1}^{N}H(\mathbf {X} _{i}){\frac {f(\mathbf {X} _{i};\mathbf {u} )}{g(\mathbf {X} _{i})}}} , where X 1 , … , X N {\displaystyle \mathbf {X} _{1},\dots ,\mathbf {X} _{N}} is a random sample from g {\displaystyle g\,} . For positive H {\displaystyle H} , the theoretically optimal importance sampling density (PDF) is given by g ∗ ( x ) = H ( x ) f ( x ; u ) / ℓ {\displaystyle g^{}(\mathbf {x} )=H(\mathbf {x} )f(\mathbf {x} ;\mathbf {u} )/\ell } . This, however, depends on the unknown ℓ {\displaystyle \ell } . The CE method aims to approximate the optimal PDF by adaptively selecting members of the parametric family that are closest (in the Kullback–Leibler sense) to the optimal PDF g ∗ {\displaystyle g^{}} . == Generic CE algorithm == Choose initial parameter vector v ( 0 ) {\displaystyle \mathbf {v} ^{(0)}} ; set t = 1. Generate a random sample X 1 , … , X N {\displaystyle \mathbf {X} _{1},\dots ,\mathbf {X} _{N}} from f ( ⋅ ; v ( t − 1 ) ) {\displaystyle f(\cdot ;\mathbf {v} ^{(t-1)})} Solve for v ( t ) {\displaystyle \mathbf {v} ^{(t)}} , where v ( t ) = argmax v 1 N ∑ i = 1 N H ( X i ) f ( X i ; u ) f ( X i ; v ( t − 1 ) ) log f ( X i ; v ) {\displaystyle \mathbf {v} ^{(t)}=\mathop {\textrm {argmax}} _{\mathbf {v} }{\frac {1}{N}}\sum _{i=1}^{N}H(\mathbf {X} _{i}){\frac {f(\mathbf {X} _{i};\mathbf {u} )}{f(\mathbf {X} _{i};\mathbf {v} ^{(t-1)})}}\log f(\mathbf {X} _{i};\mathbf {v} )} If convergence is reached then stop; otherwise, increase t by 1 and reiterate from step 2. In several cases, the solution to step 3 can be found analytically. Situations in which this occurs are When f {\displaystyle f\,} belongs to the natural exponential family When f {\displaystyle f\,} is discrete with finite support When H ( X ) = I { x ∈ A } {\displaystyle H(\mathbf {X} )=\mathrm {I} _{\{\mathbf {x} \in A\}}} and f ( X i ; u ) = f ( X i ; v ( t − 1 ) ) {\displaystyle f(\mathbf {X} _{i};\mathbf {u} )=f(\mathbf {X} _{i};\mathbf {v} ^{(t-1)})} , then v ( t ) {\displaystyle \mathbf {v} ^{(t)}} corresponds to the maximum likelihood estimator based on those X k ∈ A {\displaystyle \mathbf {X} _{k}\in A} . == Continuous optimization—example == The same CE algorithm can be used for optimization, rather than estimation. Suppose the problem is to maximize some function S {\displaystyle S} , for example, S ( x ) = e − ( x − 2 ) 2 + 0.8 e − ( x + 2 ) 2 {\displaystyle S(x)={\textrm {e}}^{-(x-2)^{2}}+0.8\,{\textrm {e}}^{-(x+2)^{2}}} . To apply CE, one considers first the associated stochastic problem of estimating P θ ( S ( X ) ≥ γ ) {\displaystyle \mathbb {P} _{\boldsymbol {\theta }}(S(X)\geq \gamma )} for a given level γ {\displaystyle \gamma \,} , and parametric family { f ( ⋅ ; θ ) } {\displaystyle \left\{f(\cdot ;{\boldsymbol {\theta }})\right\}} , for example the 1-dimensional Gaussian distribution, parameterized by its mean μ t {\displaystyle \mu _{t}\,} and variance σ t 2 {\displaystyle \sigma _{t}^{2}} (so θ = ( μ , σ 2 ) {\displaystyle {\boldsymbol {\theta }}=(\mu ,\sigma ^{2})} here). Hence, for a given γ {\displaystyle \gamma \,} , the goal is to find θ {\displaystyle {\boldsymbol {\theta }}} so that D K L ( I { S ( x ) ≥ γ } ‖ f θ ) {\displaystyle D_{\mathrm {KL} }({\textrm {I}}_{\{S(x)\geq \gamma \}}\|f_{\boldsymbol {\theta }})} is minimized. This is done by solving the sample version (stochastic counterpart) of the KL divergence minimization problem, as in step 3 above. It turns out that parameters that minimize the stochastic counterpart for this choice of target distribution and parametric family are the sample mean and sample variance corresponding to the elite samples, which are those samples that have objective function value ≥ γ {\displaystyle \geq \gamma } . The worst of the elite samples is then used as the level parameter for the next iteration. This yields the following randomized algorithm that happens to coincide with the so-called Estimation of Multivariate Normal Algorithm (EMNA), an estimation of distribution algorithm. === Pseudocode === // Initialize parameters μ := −6 σ2 := 100 t := 0 maxits := 100 N := 100 Ne := 10 // While maxits not exceeded and not converged while t < maxits and σ2 > ε do // Obtain N samples from current sampling distribution X := SampleGaussian(μ, σ2, N) // Evaluate objective function at sampled points S := exp(−(X − 2) ^ 2) + 0.8 exp(−(X + 2) ^ 2) // Sort X by objective function values in descending order X := sort(X, S) // Update parameters of sampling distribution via elite samples μ := mean(X(1:Ne)) σ2 := var(X(1:Ne)) t := t + 1 // Return mean of final sampling distribution as solution return μ == Related methods == Simulated annealing Genetic algorithms Harmony search Estimation of distribution algorithm Tabu search Natural Evolution Strategy Ant colony optimization algorithms
Deep learning speech synthesis
Deep learning speech synthesis refers to the application of deep learning models to generate natural-sounding human speech from written text (text-to-speech) or spectrum (vocoder). Deep neural networks are trained using large amounts of recorded speech and, in the case of a text-to-speech system, the associated labels and/or input text. == Formulation == Given an input text or some sequence of linguistic units Y {\displaystyle Y} , the target speech X {\displaystyle X} can be derived by X = arg max P ( X | Y , θ ) {\displaystyle X=\arg \max P(X|Y,\theta )} where θ {\displaystyle \theta } is the set of model parameters. Typically, the input text will first be passed to an acoustic feature generator, then the acoustic features are passed to the neural vocoder. For the acoustic feature generator, the loss function is typically L1 loss (Mean Absolute Error, MAE) or L2 loss (Mean Square Error, MSE). These loss functions impose a constraint that the output acoustic feature distributions must be Gaussian or Laplacian. In practice, since the human voice band ranges from approximately 300 to 4000 Hz, the loss function will be designed to have more penalty on this range: l o s s = α loss human + ( 1 − α ) loss other {\displaystyle loss=\alpha {\text{loss}}_{\text{human}}+(1-\alpha ){\text{loss}}_{\text{other}}} where loss human {\displaystyle {\text{loss}}_{\text{human}}} is the loss from human voice band and α {\displaystyle \alpha } is a scalar, typically around 0.5. The acoustic feature is typically a spectrogram or Mel scale. These features capture the time-frequency relation of the speech signal, and thus are sufficient to generate intelligent outputs. The Mel-frequency cepstrum feature used in the speech recognition task is not suitable for speech synthesis, as it reduces too much information. == History == In September 2016, DeepMind released WaveNet, which demonstrated that deep learning-based models are capable of modeling raw waveforms and generating speech from acoustic features like spectrograms or mel-spectrograms. Although WaveNet was initially considered to be computationally expensive and slow to be used in consumer products at the time, a year after its release, DeepMind unveiled a modified version of WaveNet known as "Parallel WaveNet," a production model 1,000 faster than the original. This was followed by Google AI's Tacotron 2 in 2018, which demonstrated that neural networks could produce highly natural speech synthesis but required substantial training data—typically tens of hours of audio—to achieve acceptable quality. Tacotron 2 used an autoencoder architecture with attention mechanisms to convert input text into mel-spectrograms, which were then converted to waveforms using a separate neural vocoder. When trained on smaller datasets, such as 2 hours of speech, the output quality degraded while still being able to maintain intelligible speech, and with just 24 minutes of training data, Tacotron 2 failed to produce intelligible speech. In 2019, Microsoft Research introduced FastSpeech, which addressed speed limitations in autoregressive models like Tacotron 2. FastSpeech utilized a non-autoregressive architecture that enabled parallel sequence generation, significantly reducing inference time while maintaining audio quality. Its feedforward transformer network with length regulation allowed for one-shot prediction of the full mel-spectrogram sequence, avoiding the sequential dependencies that bottlenecked previous approaches. The same year saw the release of HiFi-GAN, a generative adversarial network (GAN)-based vocoder that improved the efficiency of waveform generation while producing high-fidelity speech. In 2020, the release of Glow-TTS introduced a flow-based approach that allowed for fast inference and voice style transfer capabilities. In March 2020, the free text-to-speech website 15.ai was launched. 15.ai gained widespread international attention in early 2021 for its ability to synthesize emotionally expressive speech of fictional characters from popular media with minimal amount of data. The creator of 15.ai (known pseudonymously as 15) stated that 15 seconds of training data is sufficient to perfectly clone a person's voice (hence its name, "15.ai"), a significant reduction from the previously known data requirement of tens of hours. 15.ai is credited as the first platform to popularize AI voice cloning in memes and content creation. 15.ai used a multi-speaker model that enabled simultaneous training of multiple voices and emotions, implemented sentiment analysis using DeepMoji, and supported precise pronunciation control via ARPABET. The 15-second data efficiency benchmark was later corroborated by OpenAI in 2024. == Semi-supervised learning == Currently, self-supervised learning has gained much attention through better use of unlabelled data. Research has shown that, with the aid of self-supervised loss, the need for paired data decreases. == Zero-shot speaker adaptation == Zero-shot speaker adaptation is promising because a single model can generate speech with various speaker styles and characteristic. In June 2018, Google proposed to use pre-trained speaker verification models as speaker encoders to extract speaker embeddings. The speaker encoders then become part of the neural text-to-speech models, so that it can determine the style and characteristics of the output speech. This procedure has shown the community that it is possible to use only a single model to generate speech with multiple styles. == Neural vocoder == In deep learning-based speech synthesis, neural vocoders play an important role in generating high-quality speech from acoustic features. The WaveNet model proposed in 2016 achieves excellent performance on speech quality. Wavenet factorised the joint probability of a waveform x = { x 1 , . . . , x T } {\displaystyle \mathbf {x} =\{x_{1},...,x_{T}\}} as a product of conditional probabilities as follows p θ ( x ) = ∏ t = 1 T p ( x t | x 1 , . . . , x t − 1 ) {\displaystyle p_{\theta }(\mathbf {x} )=\prod _{t=1}^{T}p(x_{t}|x_{1},...,x_{t-1})} where θ {\displaystyle \theta } is the model parameter including many dilated convolution layers. Thus, each audio sample x t {\displaystyle x_{t}} is conditioned on the samples at all previous timesteps. However, the auto-regressive nature of WaveNet makes the inference process dramatically slow. To solve this problem, Parallel WaveNet was proposed. Parallel WaveNet is an inverse autoregressive flow-based model which is trained by knowledge distillation with a pre-trained teacher WaveNet model. Since such inverse autoregressive flow-based models are non-auto-regressive when performing inference, the inference speed is faster than real-time. Meanwhile, Nvidia proposed a flow-based WaveGlow model, which can also generate speech faster than real-time. However, despite the high inference speed, parallel WaveNet has the limitation of needing a pre-trained WaveNet model, so that WaveGlow takes many weeks to converge with limited computing devices. This issue has been solved by Parallel WaveGAN, which learns to produce speech through multi-resolution spectral loss and GAN learning strategies.