Defence Information Infrastructure (DII) is a secure military network owned by the United Kingdom's Ministry of Defence MOD. It is used by all branches of the armed forces, including the Royal Navy, British Army and Royal Air Force as well as MOD civil servants. It reaches to deployed bases and ships at sea, but not to aircraft in flight. In 2000, the MOD began to plan the systems replacement project. In March 2005, the MOD gave a contract to the Atlas Consortium, with EDS as prime contractor, for installation and management over 10 years. That has developed into a consortium made up of DXC Technology (formerly EDS), Fujitsu, Airbus Defence and Space (formerly EADS Defence & Security) and CGI (formerly Logica). Starting in May 2016, MOD users of DII begin to migrate to the New Style of IT within the defence to be known as MODNET; again supported by ATLAS. == Overview == DII supports 2,000 MOD sites with some 150,000 terminals (desktops and laptops) and 300,000 user accounts. It is designed to offer a high level of resilience, flexibility, and security in the provision of connectivity from ‘business space to battlespace’ in MOD offices in the UK, bases overseas, at sea, and on the front line. It aims to rationalise and improve IT provision for the defence sector in the 21st century; involving a major culture change for MOD users and their ways of working through a structure of shared working areas with controlled security and access. It should provide a records management system and search facility together with a range of office services. It hosts several hundred COTS (commercial off-the-shelf) and bespoke MOD applications from a range of suppliers judged to meet the required security standards. The network handles alphanumeric data, graphics, and video. The system carries information from Restricted to above-Secret levels, but users are able to see only the data and applications for which they are authorised. == Incremental approach == In order to de-risk the programme Atlas and the MOD took an incremental approach to the development and implementation of DII, with a separate contract for each increment. The extended timeline allowed the MOD flexibility in defining its requirements. Increment 1: Contract awarded March 2005. This covered 70,000 user access devices (UADs) and 200,000 user accounts in the Restricted and Secret domains in 680 fixed locations. Increment 2a: Contract awarded December 2006. This was for an additional 44,000 UADs and 58,000 user accounts in the Restricted and Secret domains, again in fixed locations. Increment 2b: Contract awarded September 2007: This extended DII(F) into the deployed environment with the provision of UADs to support land and maritime deployed operations. Increment 2c: Signed in January 2009. This extended the DII footprint into the above-Secret domain to support a number of key operations and intelligence initiatives. Increment 3a: Contract awarded January 2010. Atlas provided 42,000 UADs operating in the Restricted and Secret domains to the remaining MOD fixed sites. This supported some 60,000 personnel, notably within the RAF, at Joint Helicopter Command and other MOD locations. Increment 3a received an MOD Chief of Defence Materiel commendation. == Costs and transparency == The Ministry of Defence informed Parliament the system would cost £2.3bn, even though it knew the cost would be at least £5.8bn. By 2008 the programme was running at least 18 months late; had delivered only 29,000 of a contracted 63,000 terminals; and had delivered none of the contracted Secret capability. In January 2010 the Parliamentary Under-Secretary of State for Defence announced that the Ministry of Defence had authorised DII increment 3a at a cost of around £540 million to provide 42,000 terminals within the RAF and at Joint Helicopter Command. He stated that the project would deliver "benefits" worth over £1.6 billion over the 10 years of the contract. That year the project was scheduled to cost at least £7bn, however, the UK government said it might attempt to reduce this sum. By 2014 the rollout of all UK terminals was complete and a refresh of the original desktops and printers to new hardware underway. The overseas rollout was coming to an end and well over half the fleet, including aircraft carrier HMS Queen Elizabeth, equipped. The final part of Secret capability deployment was scheduled to complete in summer of 2014.
WhatsApp Messenger, commonly known simply as WhatsApp, is an American social media, instant messaging (IM), and Voice over IP (VoIP) service accessible via desktop and mobile app. Owned by Meta Platforms, the service allows users to send text messages, voice messages, and video messages, make voice and video calls, and share images, documents, user locations, and other content. The service requires a cellular mobile telephone number to register. WhatsApp was launched in May 2009. In January 2018, WhatsApp released a standalone business app called WhatsApp Business which can communicate with the standard WhatsApp client. As of May 2025, the service had 3 billion monthly active users, making it the most used messenger app. The name of the app is meant to sound like "what's up". The service was created by WhatsApp Inc. of Mountain View, California, which was acquired by Facebook in February 2014 for approximately US$19.3 billion. It became the world's most popular messaging application in 2015, with 900 million users, and had more than 2 billion active users worldwide in February 2020. WhatsApp Business had approximately 200 million monthly users in 2023. By 2016, it had become the primary means of Internet communication in regions including the Americas, the Indian subcontinent, and large parts of Europe and Africa. == History == === 2009–2014 === WhatsApp was founded by Brian Acton and Jan Koum, former employees of Yahoo. Koum incorporated WhatsApp Inc. in California on February 24, 2009. A month earlier, Koum had purchased an iPhone, and he and Acton decided to create an app for the App Store. The idea started off as an app that would display statuses in a phone's Contacts menu, showing if a person was at work or on a call. Their discussions often took place at the home of Koum's Russian friend Alex Fishman in West San Jose. They realized that to take the idea further, they would need an iPhone developer. Fishman visited RentACoder.com, found Russian developer Igor Solomennikov, and introduced him to Koum. Koum named the app WhatsApp to sound like "what's up" and it was published on the Apple App Store and BlackBerry App World in May and June 2009 respectively. However, when early versions of WhatsApp kept crashing, Koum considered giving up and looking for a new job. Acton encouraged him to wait for a "few more months". In June 2009, when the app had been downloaded by only a handful of Fishman's Russian-speaking friends, Apple launched push technology, allowing users to be pinged even when not using the app. Koum updated WhatsApp so that everyone in the user's network would be notified when a user's status changed. This new facility, to Koum's surprise, was used by users to ping "each other with jokey custom statuses like, 'I woke up late' or 'I'm on my way.'" Fishman said, "At some point it sort of became instant messaging". WhatsApp 2.0, released for iPhone in August 2009, featured a purpose-designed messaging component; the number of active users suddenly increased to 250,000. Although Acton was working on another startup idea, he decided to join the company. In October 2009, Acton persuaded five former friends at Yahoo! to invest $250,000 in seed funding, and Acton became a co-founder and was given a stake. He officially joined WhatsApp on November 1. Koum then hired a friend in Los Angeles, Chris Peiffer, to develop a BlackBerry version, which arrived two months later. Subsequently, WhatsApp for Symbian OS was added in May 2010, and for Android OS in August 2010. In 2010 Google made multiple acquisition offers for WhatsApp, which were all declined. To cover the cost of sending verification texts to users, WhatsApp was changed from a free service to a paid one. In December 2009, the ability to send photos was added to the iOS version. By early 2011, WhatsApp was one of the top 20 apps in the U.S. Apple App Store. In April 2011, Sequoia Capital invested about $8 million for more than 15% of the company, after months of negotiation by Sequoia partner Jim Goetz. By February 2013, WhatsApp had about 200 million active users and 50 staff members. Sequoia invested another $50 million at a $1.5 billion valuation. Some time in 2013 WhatsApp acquired Santa Clara–based startup SkyMobius, the developers of Vtok, a video and voice calling app. As of December 2013, the service had 400 million monthly active users. That year, the company had $148 million in expenses and a net loss of $138 million. === 2014–2015 === On February 19, 2014, one year after the venture capital financing round at a $1.5 billion valuation, Facebook, Inc. (now Meta Platforms) agreed to acquire the company for US$19 billion, its largest acquisition to date. At the time, it was the largest acquisition of a venture-capital-backed company in history. Sequoia Capital received an approximate 5,000% return on its initial investment. Facebook paid $4 billion in cash, $12 billion in Facebook shares, and an additional $3 billion in restricted stock units granted to WhatsApp's founders Koum and Acton. Employee stock was scheduled to vest over four years subsequent to closing. Days after the announcement, WhatsApp users experienced a loss of service, leading to anger across social media. The acquisition was influenced by the data provided by Onavo, Facebook's research app for monitoring competitors and trending usage of social activities on mobile phones, as well as startups that were performing "unusually well". The acquisition caused many users to try, or move to, other message services. Telegram claimed that it acquired 8 million new users, and Line, 2 million. At a keynote presentation at the Mobile World Congress in Barcelona in February 2014, Facebook CEO Mark Zuckerberg said that Facebook's acquisition of WhatsApp was closely related to the Internet.org vision. A TechCrunch article said about Zuckerberg's vision:The idea, he said, is to develop a group of basic internet services that would be free of charge to use – "a 911 for the internet". These could be a social networking service like Facebook, a messaging service, maybe search and other things like weather. Providing a bundle of these free of charge to users will work like a gateway drug of sorts – users who may be able to afford data services and phones these days just don't see the point of why they would pay for those data services. This would give them some context for why they are important, and that will lead them to pay for more services like this – or so the hope goes. Three days after announcing the Facebook purchase, Koum said they were working to introduce voice calls. He also said that new mobile phones would be sold in Germany with the WhatsApp brand, and that their ultimate goal was to be on all smartphones. In August 2014, WhatsApp was the most popular messaging app in the world, with more than 600 million users. By early January 2015, WhatsApp had 700 million monthly users and over 30 billion messages every day. In April 2015, Forbes predicted that between 2012 and 2018, the telecommunications industry would lose $386 billion because of "over-the-top" services like WhatsApp and Skype. That month, WhatsApp had over 800 million users. By September 2015, it had grown to 900 million; and by February 2016, one billion. On November 30, 2015, the Android WhatsApp client made links to Telegram unclickable and not copyable. Multiple sources confirmed that it was intentional, not a bug, and that it had been implemented when the Android source code that recognized Telegram URLs had been identified. (The word "telegram" appeared in WhatsApp's code.) Some considered it an anti-competitive measure; WhatsApp offered no explanation. === 2016–2019 === On January 18, 2016, WhatsApp's co-founder Jan Koum announced that it would no longer charge users a $1 annual subscription fee, in an effort to remove a barrier faced by users without payment cards. He also said that the app would not display any third-party ads, and that it would have new features such as the ability to communicate with businesses. On May 18, 2017, the European Commission announced that it was fining Facebook €110 million for "providing misleading information about WhatsApp takeover" in 2014. The Commission said that in 2014 when Facebook acquired the messaging app, it "falsely claimed it was technically impossible to automatically combine user information from Facebook and WhatsApp." However, in the summer of 2016, WhatsApp had begun sharing user information with its parent company, allowing information such as phone numbers to be used for targeted Facebook advertisements. Facebook acknowledged the breach, but said the errors in their 2014 filings were "not intentional". In September 2017, WhatsApp's co-founder Brian Acton left the company to start a nonprofit group, later revealed as the Signal Foundation, which developed the WhatsApp competitor Signal. He explained his reasons for leaving in an interview with Forbes a year later. WhatsApp also
Probabilistic database
Most real databases contain data whose correctness is uncertain. In order to work with such data, there is a need to quantify the integrity of the data. This is achieved by using probabilistic databases. A probabilistic database is an uncertain database in which the possible worlds have associated probabilities. Probabilistic database management systems are currently an active area of research. "While there are currently no commercial probabilistic database systems, several research prototypes exist..." Probabilistic databases distinguish between the logical data model and the physical representation of the data much like relational databases do in the ANSI-SPARC Architecture. In probabilistic databases this is even more crucial since such databases have to represent very large numbers of possible worlds, often exponential in the size of one world (a classical database), succinctly. == Terminology == In a probabilistic database, each tuple is associated with a probability between 0 and 1, with 0 representing that the data is certainly incorrect, and 1 representing that it is certainly correct. === Possible worlds === A probabilistic database could exist in multiple states. For example, if there is uncertainty about the existence of a tuple in the database, then the database could be in two different states with respect to that tuple—the first state contains the tuple, while the second one does not. Similarly, if an attribute can take one of the values x, y or z, then the database can be in three different states with respect to that attribute. Each of these states is called a possible world. Consider the following database: (Here {b3, b3′, b3′′} denotes that the attribute can take any of the values b3, b3′ or b3′′) Assuming that there is uncertainty about the first tuple, certainty about the second tuple, and uncertainty about the value of attribute B in the third tuple. Then the actual state of the database may or may not contain the first tuple (depending on whether it is correct or not). Similarly, the value of the attribute B may be b3, b3′ or b3′′. Consequently, the possible worlds corresponding to the database are as follows: === Types of Uncertainties === There are essentially two kinds of uncertainties that could exist in a probabilistic database, as described in the table below: By assigning values to random variables associated with the data items, different possible worlds can be represented. == History == The first published use of the term "probabilistic database" was probably in the 1987 VLDB conference paper "The theory of probabilistic databases", by Cavallo and Pittarelli. The title (of the 11 page paper) was intended as a bit of a joke, since David Maier's 600 page monograph, The Theory of Relational Databases, would have been familiar at that time to many of the conference participants and readers of the conference proceedings.
Autonomic computing
Autonomic computing (AC) is distributed computing resources with self-managing characteristics, adapting to unpredictable changes while hiding intrinsic complexity to operators and users. Initiated by IBM in 2001, this initiative ultimately aimed to develop computer systems capable of self-management, to overcome the rapidly growing complexity of computing systems management, and to reduce the barrier that complexity poses to further growth. == Description == The AC system concept is designed to make adaptive decisions, using high-level policies. It will constantly check and optimize its status and automatically adapt itself to changing conditions. An autonomic computing framework is composed of autonomic components (AC) interacting with each other. An AC can be modeled in terms of two main control schemes (local and global) with sensors (for self-monitoring), effectors (for self-adjustment), knowledge and planner/adapter for exploiting policies based on self- and environment awareness. This architecture is sometimes referred to as Monitor-Analyze-Plan-Execute (MAPE). Driven by such vision, a variety of architectural frameworks based on "self-regulating" autonomic components has been recently proposed. A similar trend has recently characterized significant research in the area of multi-agent systems. However, most of these approaches are typically conceived with centralized or cluster-based server architectures in mind and mostly address the need of reducing management costs rather than the need of enabling complex software systems or providing innovative services. Some autonomic systems involve mobile agents interacting via loosely coupled communication mechanisms. Autonomy-oriented computation is a paradigm proposed by Jiming Liu in 2001 that uses artificial systems imitating social animals' collective behaviours to solve difficult computational problems. For example, ant colony optimization could be studied in this paradigm. == Problem of growing complexity == Forecasts suggested that the computing devices in use would grow at 38% per year and the average complexity of each device was increasing. This volume and complexity was managed by highly skilled humans; but the demand for skilled IT personnel was already outstripping supply, with labour costs exceeding equipment costs by a ratio of up to 18:1. Computing systems have brought great benefits of speed and automation but there is now an overwhelming economic need to automate their maintenance. In a 2003 IEEE Computer article, Kephart and Chess warn that the dream of interconnectivity of computing systems and devices could become the "nightmare of pervasive computing" in which architects are unable to anticipate, design and maintain the complexity of interactions. They state the essence of autonomic computing is system self-management, freeing administrators from low-level task management while delivering better system behavior. A general problem of modern distributed computing systems is that their complexity, and in particular the complexity of their management, is becoming a significant limiting factor in their further development. Large companies and institutions are employing large-scale computer networks for communication and computation. The distributed applications running on these computer networks are diverse and deal with multiple tasks, ranging from internal control processes to presenting web content to customer support. Additionally, mobile computing is pervading these networks at an increasing speed: employees need to communicate with their companies while they are not in their office. They do so by using laptops, personal digital assistants, or mobile phones with diverse forms of wireless technologies to access their companies' data. This creates an enormous complexity in the overall computer network which is hard to control manually by human operators. Manual control is time-consuming, expensive, and error-prone. The manual effort needed to control a growing networked computer-system tends to increase quickly. 80% of such problems in infrastructure happen at the client specific application and database layer. Most 'autonomic' service providers guarantee only up to the basic plumbing layer (power, hardware, operating system, network and basic database parameters). == Characteristics of autonomic systems == A possible solution could be to enable modern, networked computing systems to manage themselves without direct human intervention. The Autonomic Computing Initiative (ACI) aims at providing the foundation for autonomic systems. It is inspired by the autonomic nervous system of the human body. This nervous system controls important bodily functions (e.g. respiration, heart rate, and blood pressure) without any conscious intervention. In a self-managing autonomic system, the human operator takes on a new role: instead of controlling the system directly, he/she defines general policies and rules that guide the self-management process. For this process, IBM defined the following four types of property referred to as self-star (also called self-, self-x, or auto-) properties. Self-configuration: Automatic configuration of components; Self-healing: Automatic discovery, and correction of faults; Self-optimization: Automatic monitoring and control of resources to ensure the optimal functioning with respect to the defined requirements; Self-protection: Proactive identification and protection from arbitrary attacks. Others such as Poslad and Nami and Sharifi have expanded on the set of self-star as follows: Self-regulation: A system that operates to maintain some parameter, e.g., Quality of service, within a reset range without external control; Self-learning: Systems use machine learning techniques such as unsupervised learning which does not require external control; Self-awareness (also called Self-inspection and Self-decision): System must know itself. It must know the extent of its own resources and the resources it links to. A system must be aware of its internal components and external links in order to control and manage them; Self-organization: System structure driven by physics-type models without explicit pressure or involvement from outside the system; Self-creation (also called Self-assembly, Self-replication): System driven by ecological and social type models without explicit pressure or involvement from outside the system. A system's members are self-motivated and self-driven, generating complexity and order in a creative response to a continuously changing strategic demand; Self-management (also called self-governance): A system that manages itself without external intervention. What is being managed can vary dependent on the system and application. Self -management also refers to a set of self-star processes such as autonomic computing rather than a single self-star process; Self-description (also called self-explanation or Self-representation): A system explains itself. It is capable of being understood (by humans) without further explanation. IBM has set forth eight conditions that define an autonomic system: The system must know itself in terms of what resources it has access to, what its capabilities and limitations are and how and why it is connected to other systems; be able to automatically configure and reconfigure itself depending on the changing computing environment; be able to optimize its performance to ensure the most efficient computing process; be able to work around encountered problems by either repairing itself or routing functions away from the trouble; detect, identify and protect itself against various types of attacks to maintain overall system security and integrity; adapt to its environment as it changes, interacting with neighboring systems and establishing communication protocols; rely on open standards and cannot exist in a proprietary environment; anticipate the demand on its resources while staying transparent to users. Even though the purpose and thus the behaviour of autonomic systems vary from system to system, every autonomic system should be able to exhibit a minimum set of properties to achieve its purpose: Automatic: This essentially means being able to self-control its internal functions and operations. As such, an autonomic system must be self-contained and able to start-up and operate without any manual intervention or external help. Again, the knowledge required to bootstrap the system (Know-how) must be inherent to the system. Adaptive: An autonomic system must be able to change its operation (i.e., its configuration, state and functions). This will allow the system to cope with temporal and spatial changes in its operational context either long term (environment customisation/optimisation) or short term (exceptional conditions such as malicious attacks, faults, etc.). Aware: An autonomic system must be able to monitor (sense) its operational context as well as its internal state in order to be able to asses
Really Simple Licensing
Really Simple Licensing (RSL) is an open content licensing standard that allows web publishers to set terms for web crawlers gathering training data for generative AI use. It was launched on September 10, 2025 and is managed by the nonprofit RSL Collective, co-founded by RSS co-creator Eckart Walther and former Ask.com CEO Doug Leeds. Participating companies at launch include Reddit, Yahoo, and Medium. Publishers can implement the RSL standard by adding licensing terms to their robots.txt files.
Automated medical scribe
Automated medical scribes (also called artificial intelligence scribes, AI scribes, digital scribes, virtual scribes, ambient AI scribes, AI documentation assistants, and digital/virtual/smart clinical assistants) are tools for transcribing medical speech, such as patient consultations and dictated medical notes. Many also produce summaries of consultations. Automated medical scribes based on large language models (LLMs, commonly called "AI", short for "artificial intelligence") increased drastically in popularity in 2024. There are privacy and antitrust concerns. Accuracy concerns also exist, and intensify in situations in which tools try to go beyond transcribing and summarizing, and are asked to format information by its meaning, since LLMs do not deal well with meaning (see weak artificial intelligence). Medics using these scribes are generally expected to understand the ethical and legal considerations, and supervise the outputs. The privacy protections of automated medical scribes vary widely. While it is possible to do all the transcription and summarizing locally, with no connection to the internet, most closed-source providers require that data be sent to their own servers over the internet, processed there, and the results sent back (as with digital voice assistants). Some retailers say their tools use zero-knowledge encryption (meaning that the service provider can't access the data). Others explicitly say that they use patient data to train their AIs, or rent or resell it to third parties; the nature of privacy protections used in such situations is unclear, and they are likely not to be fully effective. Most providers have not published any safety or utility data in academic journals, and are not responsive to requests from medical researchers studying their products. == Privacy == Some providers unclear about what happens to user data. Some may sell data to third parties. Some explicitly send user data to for-profit tech companies for secondary purposes, which may not be specified. Some require users to sign consents to such reuse of their data. Some ingest user data to train the software, promising to anonymize it; however, deanonymization may be possible (that is, it may become obvious who the patient is). It is intrinsically impossible to prevent an LLM from correlating its inputs; they work by finding similar patterns across very large data sets. Some information on the patient will be known from other sources (for instance, information that they were injured in an incident on a certain day might be available from the news media; information that they attended specific appointment locations at specific times is probably available to their cellphone provider/apps/data brokers; information about when they had a baby is probably implied by their online shopping records; and they might mention lifestyle changes to their doctor and on a forum or blog). The software may correlate such information with the "anonymized" clinical consultation record, and, asked about the named patient, provide information which they only told their doctor privately. Because a patient's record is all about the same patient, it is all unavoidably linked; in very many cases, medical histories are intrinsically identifiable. Depending on how common a condition and what other data is available, K-anonymity may be useless. Differential privacy could theoretically preserve privacy. Data broker companies like Google, Amazon, Apple and Microsoft have produced or bought up medical scribes, some of which use user data for secondary purposes, which has led to antitrust concerns. Transfer of patient records for AI training has, in the past, prompted legal action. Open-source programs typically do all the transcription locally, on the doctor's own computer. Open-source software is widely used in healthcare, with some national public healthcare bodies holding hack days. === Data resale and commercialization === Several AI medical scribe providers include terms in their service agreements that allow the reuse, sale, or commercialization of de-identified or user-submitted data. Although such data are generally described as anonymized or aggregated, these practices have raised ethical concerns among clinicians and privacy advocates regarding secondary uses of medical information beyond clinical documentation. Freed, an AI transcription and scribe platform, states in its Terms of Use that it may "collect, use, publish, disseminate, sell, transfer, and otherwise exploit" de-identified and aggregated data derived from user inputs. OpenEvidence similarly states that it may "collect, use, transfer, sell, and disclose non-personal information and customer usage data for any purpose including commercial uses." Doximity, which offers an AI-enabled medical scribe as part of its physician platform, grants itself a "nonexclusive, irrevocable, worldwide, perpetual, unlimited, assignable, sublicensable, royalty-free" license to "copy, prepare derivative works from, improve, distribute, publish, ... analyze, index, tag, [and] commercialize" content submitted by users, subject to its privacy policy. Because these terms allow broad secondary use—including sale, licensing, model-training, derivative works, and commercial exploitation of de-identified or user-submitted data—some commentators have recommended that clinicians review data-handling provisions carefully when adopting AI-scribe tools, particularly in clinical environments where patient privacy and regulatory compliance are critical. === Encryption === Multifactor authentication for access to the data is expected practice. Typically, Diffie–Hellman key exchange is used for encryption; this is the standard method commonly used for things like online banking. This encryption is expensive but not impossible to break; it is not generally considered safe against eavesdroppers with the resources of a nation-state. If content is encrypted between the client and the service provider's remote server (transport cryptography), then the server has an unencrypted copy. This is necessary if the data is used by the service provider (for instance, to train the software). Zero-knowledge encryption implies that the only unencrypted copy is at the client, and the server cannot decrypt the data any more easily than a monster-in-the-middle attacker. == Platforms == Scribes may operate on desktops, laptop, or mobile computers, under a variety of operating systems. These vary in their risks; for instance, mobiles can be lost. The underlying mobile or desktop operating systems are also part of the trusted computing base, and if they are not secure, the software relying on them cannot be secure either. Some AI medical scribe platforms are designed to operate as cloud-based applications that generate structured clinical documentation from clinician–patient conversations. These systems may offer features such as real-time transcription, document generation, and integration with electronic health record (EHR) systems. == Confabulation, omissions, and other errors == Like other LLMs, medical-scribe LLMs are prone to hallucinations, where they make up content based on statistically associations between their training data and the transcription audio. LLMs do not distinguish between trying to transcribe the audio and guessing what words will come next, but perform both processes mixed together. They are especially likely to take short silences or non-speech noises and invent some sort of speech to transcribe them as. LLM medical scribes have been known to confabulate racist and otherwise prejudiced content; this is partly because the training datasets of many LLMs contain pseudoscientific texts about medical racism. They may misgender patients. A survey found that most doctors preferred, in principle, that scribes be trained on data reviewed by medical subject experts. Relevant, accurate training data increases the probability of an accurate transcription, but does not guarantee accuracy. Software trained on thousands of real clinical conversations generated transcripts with lower word error rates. Software trained on manually-transcribed training data did better than software trained with automatically transcribed training data such as YouTube captions. Autoscribes omit parts of the conversation classes as irrelevant. The may wrongly classify pertinent information as irrelevant and omit it. They may also confuse historic and current symptoms, or otherwise misclassify information. They may also simply wrongly transcribe the speech, writing something incorrect instead. If clinicians do not carefully check the recording, such mistakes could make their way into their medical records and cause patient harms. == Patient consent == Professional organizations generally require that scribes be used only with patient consent; some bodies may require written consent. Medics must also abide by local surveillance laws, which may criminalize recording pri
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.