Category All posts

Geneva, Switzerland – 26 January 2022. Today the Moving Picture, Audio and Data Coding by Artificial Intelligence (MPAI) standards developing organisation has concluded its 16^th General Assembly, the first of 2022, approving its 2022 work program.

The work program includes the development of reference software, conformance testing and performance assessment for 2 application standards (Context-based Audio Enhancement and Multimodal Conversation), reference software, conformance assessment for 1 infrastructure standard (AI Framework), and the establishment of the MPAI Store, a non-profit foundation with the mission to distribute verified implementations of MPAI standards, as specified in another MPAI infrastructure standard (Governance of the MPAI Ecosystem).

An important part of the work program addresses the development of performance assessment specifications for the 2 application standards. The purpose of performance assessment is to enable MPAI-appointed entities to assess the grade of reliability, robustness, replicability and fairness of implementations. While performance will not be mandatory for an implementation to be posted to the MPAI Store, users downloading an implementation will be informed of its status.

Another section of the work program concerns the development of extensions of existing standards. Company Performance Prediction (part of Compression and Understanding of Industrial Data) will include more risks in addition to seismic and cyber; Multimodal Conversation will enhance the features of some of its use cases, e.g., by applying them to the interaction of a human with a connected autonomous vehicle; and Context-based Audio Enhancement will enter the domain of separation of useful sounds from the environment.

An important part of the work program is assigned to developing new standards for the areas that have been explored in the last few months, such as:

1. Server-based Predictive Multiplayer Gaming (MPAI-SPG) using AI to train a network that com­pensates data losses and detects false data in online multiplayer gaming.
2. AI-Enhanced Video Coding (MPAI-EVC) improving existing video coding with AI tools for short-to-medium term applications.
3. End-to-End Video Coding (MPAI-EEV) exploring on the promising area of AI-based “end-to-end” video coding for longer-term applications.
4. Connected Autonomous Vehicles (MPAI-CAV) using AI for such features as Environment Sensing, Autonomous Motion, and Motion Actuation.

Finally, MPAI welcomes new activities proposed by its members to its work program:

1. Avatar Representation and Animation (MPAI-ARA) targeting the specification of avatar descriptors.
2. Neural Network Watermarking (MPAI-NNW) developing measures of the impact of adding ownership and licensing information inside a neural network.

MPAI develops data coding standards for applications that have AI as the core enabling technology. Any legal entity supporting the MPAI mission may join MPAI, if able to contribute to the development of standards for the efficient use of data.

Visit the MPAI web site, contact the MPAI secretariat for specific information, subscribe to the MPAI Newsletter and follow MPAI on social media: LinkedIn, Twitter, Facebook , Instagram and YouTube.

Most important: join MPAI, share the fun, build the future.

“Digital human”  has recently become a trendy expression and different meanings can be attached to it. MPAI says that it is “a digital object able to receive text/audio/video/commands (“Information”) and generate Information that is congruent with the received Information”.

MPAI has been developing several standards for “digital humans” and plans on extending them and developing more.

Let’s have an overview.

In Conversation with Emotion a digital human perceives text or speech from and video of a human. It then generates text or speech that is congruent with content and emotion of the perceived media, and displays itself as an avatar whose lips move in sync with its speech and according to the emotion embedded in the synthetic speech.

In Multimodal Question Answering a digital human perceives text or speech from a human asking a question about an object held by the human, and the video of the human holding the object. In response it generates text or speech that is a response to the human question and is congruent with the perceived media data including the emotional state of the human.

Adding an avatar whose lips move in sync with the generated speech could have a more satisfactory rendering of the speech generated by the digital human.

In Automatic Speech Translation a digital human is told to translate speech or text generated by the human into a specified language and to preserve or not the speech features of the input speech, in case the input is speech. The digital human then generates translated text, if the input is text and translated speech preserving or not the input speech features, if the input is speech.

Adding an avatar whose lips move in sync with the generated speech and according to its embedded emotion could have a more satisfactory rendering of the digital human speech.

In Emotion Enhanced Speech a digital human is told to add an emotion to an emotion-less speech by giving

1. A model utterance: the digital human extracts and adds the speech features of the model utterance to the emotion-less speech segment.
2. An emotion taken from the MPAI standard list of emotions: the digital human adds the speech features obtained by combining the speech features proper of the selected emotion to the speech features of the emotion-less speech to the emotion-less speech.

In both cases an avatar can be animated by the emotion-enhanced speech.

MPAI has more digital human use cases under development:

1. Human-CAV Interaction: A digital human (a face) speaks to a group of humans gazing at the human it is responding to.

1. Mixed-reality Collaborative Spaces: A digital human (a torso) utters the speech of a participant in a virtual videoconference while its torso moves in sync with the participant’s torso.

1. Conversation about a scene: A digital human (a face) converses with a human about the objects of a scene the human is part of gazing at the human or at an object.

The year 2021 was very productive for MPAI. In January it started with a Call for Technologies on AI Framework and ended in December with 5standards approved and 7 projects in the pipeline.

How was that possible? Simple: intense collaborative work.

OK, but exactly how?

So far MPAI has not held physical meetings. MPAI does all current work online in 1, 2 or 3 one-hour sessions a day in the 13-18 UTC time frame. Purpose of this post is to describe how an MPAI work week unfolds. All times are UTC. All meetings last 1 hour.

Monday @15 – Mixed-reality Collaborative Spaces (MCS) is a project finalising the MCS Use Cases and Functional Requirements document. The use cases considered are “Avatar-Based Videoconference” and “Virtual eLearning” where avatars with levels of similarity with the person they represent hold meetings or assist to lectures. Extension of avatars to volumetric video is being considered. To know more, visit mcs.mpai.community.

Monday @16 – Governance of the MPAI Ecosystem (GME) is the name of an MPAI standard developed, approved and published by MPAI. It envisages the establishment of the MPAI Store, a non-profit commercial organisation whose role is to receive implementations of MPAI standards, verify they are secure, test that they are conforming with MPAI specifications, possibly collect the results proving that the implementation is reliable, robust, replicable and fair – attributes that MPAI labels as Performance – and post the implementation on the MPAI Store web site for users to download. MPAI is moving to the implementation of the MPAI Store. To know more, visit gme.mpai.community.

Monday @17 – AI Framework (AIF) is the name of an MPAI standard developed, approved and published by MPAI. MPAI is now developing a software implementation of the standard. To know more, visit aif.mpai.community.

Tuesday @14 – Multimodal Conversation (MMC) is the name of an MPAI standard developed, approved, and published by MPAI. MPAI is now developing software implementations of the four MPAI-MMC use cases, conformance testing and performance assessment and extending the current V1 specification to support more use cases. To know more, visit mmc.mpai.community.

Tuesday @16 – Context-based Audio Enhancement is the name of an MPAI standard developed, approved, and published by MPAI. MPAI is now developing software implementations of the four MPAI-MMC use cases, conformance testing and performance assessment and extending the current V1 specification to support more use cases. To know more, visit cae.mpai.community.

Wednesday @13 – Connected Autonomous Vehicles (CAV) is a project finalising the CAV Use Cases and Functional Requirements document. The MPAI CAV is composed of 4 technology-laden subsystems. To know more, visit cav.mpai.community.

Wednesday @15 – AI-Enhanced Video Coding (EVC) is a project investigating the enhancement or replacement of existing video coding tools in the MPEG-5 EVC codec with AI tools. When MPAI will reach an improvement of 25%, a Call for Technologies will be issued. To know more, visit evc.mpai.community. On alternate week, the session is taken by End-to-End Video Coding (EEV), a project addressing video coding without the constraint of traditional video coding architectures. To know more, visit eev.mpai.community.

Wednesday @16 – Compression and Understanding of Industrial Data (CUI) is the name of an MPAI standard developed, approved, and published by MPAI. As all 4 components of an MPAI standard – technical specification, reference software, conformance testing and performance assessment have been published, the group is preparing for a new version of the standard that includes additional risks to seismic and cyber.

Thursday @14 – Server-based Predictive Multiplayer Gaming (SPG) is a project completing the validation of the SPG model. Once the validation is completed, it will finalise the SPG Use Cases and Functional Requirements. To know more, visit spg.mpai.community.

In addition to these technical meeting sessions, MPAI holds a General Assembly to discuss and ratify any proposed results from the technical groups.

The Board of Directors typically meets twice between two General Assemblies. Two advisory groups hold meetings:

Thursday @15 – Communication Advisory Committee manages the manifold MPAI communication activities. Some of them are press releases, newsletter, online presentations and social media.

Friday @15 – Industry and Standards Advisory Committee manages the relationships of MPAI with all external entities that are relevant for MPAI activities.

A lot has been happening in MPAI, and a lot is now happening so that a lot can happen next outside MPAI.

Today video/audio conference is a virtual space where many of us spend their working hours. Still the experience of a conference suffer from many deficiencies depending on the fact that the way audio is captured and conveyed to the virtual space is inadequate. Our brains can separate the voice of competing speakers, and remove the effect of non-ideal acoustical properties of the physical space and/or the background noise in the same in the same physical environment.

However, when the acoustic signals from the different physical environments are merged in the virtual space, the operation of our brains can well not be as efficient. The result is a reduction in intelligibility of speech causing participants not fully understanding what their interlocutors are saying. The very purpose of the conference may be harmed and, at the end of the day, participants may very well feel more stressed than when they could meet people in person.

Many of these problems can be alleviated or resolved if a microphone array is used to capture the speakers’ speech signals. The individual speech signals can be separated, the non-ideal acoustics of the space can be reduced and any background noise can be substantially suppressed.

The fourth context of Context-based Audio Enhancement (MPAI-CAE), called Enhanced Audioconference Experience (CAE-EAE), aims to provide a complete solution enabling the processing of speech signals from a microphone array to provide clear speech signals free from back­ground noise and acoustics-related artefacts to improve the auditory quality of an audioconference experience. Specifically, CAE-EAE addresses the situation where one or more speakers are active in a noisy meeting room and try to communicate using speech with one or more interlocutors over the network.

The AIM Modules (AIM) required by this use case extract the speech signals of the individual speakers from the microphone array and reduce background noise and reverberation. CAE-EAE can also extract the spatial attributes of the speakers with respect to the position of the microphone array. This information, multiplexed with the multichannel audio can be properly used at the receiver side to create a spatial representation of the speech signals.

Figure 1 depicts the AI Module structure whose operation is described below.

Figure 1 – Enhanced Audioconference Experience Reference Model

1. Analysis Transform AIM performs a time-frequency transformation to enable the operations downstream to be carried out in the frequency domain.
2. Sound Field Description AIM converts the output from the Analysis Transform AIM into the spherical frequency domain.
3. Speech Detection and Separation AIM detects the directions of active sound sources and separates the sources using the Source Model KB which provides simple acoustic source models. The separated sources can either be speech or non-speech.
4. Noise Cancellation AIM eliminates background noise and reverberation producing a Denoised Speech in the frequency domain.
5. Synthesis Transform AIM applies the inverse transform to Denoised Speech.
6. Packager AIM produces a multiplexed stream which contains separated Multichannel Speech Streams and Audio Scene Geometry.

The MPAI CAE-EAE standard can change the experience of audio/video teleconference users.

The third context of the Context-based Audio Enhancement (MPAI-CAE) standard is restoration of damaged speech.

Unlike Audio Recording Preservation where audio has clear provenance – the magnetic tape of an open reel whose analogue audio has been digitised – Speech Restoration System does not make a reference to anything analogue. It assumes that there is a file containing digital speech. For whatever reason, it may be so that portions of the file are damaged – maybe the physical medium from which the file was created was partly corrupted. However, the use case assumes that the text that the creator used to make their speech is available.

Figure 1 shows the AI Modules (AIM), i.e., the components of the system

Figure 1 – The Speech Restoration System Reference Model

The basic idea is to create a speech model using a sufficient number of undamaged audio segments. The model is then served to a neural network acting as a speech synthesiser of the original that is used to synthesise all damaged speech segments in the list of damaged speech segments using the text corresponding to the damaged segment.

The result is an entirely restored speech file where the damaged segments have been replaced by the best estimate of the speech produced by the speaker.

It is time to become an MPAI member https://mpai.community/how-to-join/join/. Join the fun – build the future!

Established on 30 September 2020, MPAI spent its first 3 months giving itself a structure to execute its mission of developing Artificial Intelligence (AI)-based data coding standards.

Its first full year of operation – 2021 – has been engaging but rewarding:

• 5 Technical Specifications (TS)  have been approved and released in the following domains:
  • Finance.
  • Human-machine communication.
  • Audio enhancement.
  • AI Framework
  • Ecosystem Governance.
• The Company Performance Assessment TS was complemented by 3 additional specifications:
  • Reference Software (RS). a conforming implementation of the TS,
  • Conformance Testing (CT), to test that an implementation is technically correct and provides an adequate user experience
  • Performance Assessment (PA), to assess implementation reliability and trustworthiness.

A goal can be declared as reached only if the next goal is known, and the purpose of this post is to disclose exactly that.

The AI Framework (AIF), depicted in Figure 1, is a cornerstone of the MPAI architecture.

Figure 1 – The AI Framework (AIF) Reference Model and its Components

• The AIF
  • Is Operating System-independent.
  • Has a local and distributed component-based Zero-Trust architecture.
  • Can create AI Workflows (AIW) made of elementary units called AI Modules (AIM).
  • Can access validated AIWs and AIMs by interfacing to the MPAI Store.
  • Can execute in a range of computing environments: from MCUs to HPCs.
  • Can interact with other AIFs operating in proximity.
  • Supports Machine Learning functionalities.
• Its AIMs
  • Encapsulate components to abstract them from the development environment.
  • Call the Controller via standard interfaces.
  • Can be AI-based or data processing-based.
  • Can be in software or in hardware.

MPAI has already developed 3 application oriented Technical Specifications: MPAI-CAE (Enhanced audio), MPAI-CUI (Company Performance Prediction) and MPAI-MMC (Multimodal human-machine conversation). It total there are 10 AIWs and some 20 AIMs (several of them are used in different AIWs).

An active MPAI generates an ecosystem with the following actors:

1. MPAI develop standards.
2. Implem­enters develop MPAI standard implementations
3. Users access such im­plemen­tations.

MPAI is all about facilitating a market of AI applications. Releasing standards enables a market but does not ensure that the market is functional. How can a user be sure that an implementation is secure, technically correct, unbiased? Note that by “user” we do not necessarily mean an end user, but also an app developer (i.e., AIW) who may need an AIM and does not have the resources or the competence to answer the 3 questions.

In its Governance of the MPAI Ecosystem TS, MPAI has envisaged two more players:

4. Performance Assessors who assess that implementations are reliable and trustworthy.
5. The MPAI Store where uploaded implementations are:
   1. Checked for security
   2. Tested for conformance
   3. Posted to the Store with a clear indication of level of performance.

Note that MPAI appoints Performance Assessors, and establishes and controls the MPAI Store, a not-for-profit commercial entity.

Figure 2 depicts the operation of the MPAI Ecosystem.

Figure 2 – The MPAI Ecosystem and its Governance

In 2020 MPAI has developed 3 application oriented TSs:

Compression and Understanding of Industrial Data (MPAI-CUI) with 1 use case.

Multimodal Conversation (MPAI-MMC) with 5 use cases.

Context-based Audio Enhancement (MPAI-CAE) with 4 use cases.

Figure 3 depicts the reference model of the Company Performance Prediction Use Case.

	AI-based Company Performance Prediction measures the performance of a Company by providing Default Probability, Organisational Model Index, and Business Discontinuity Probability of the Company within a given Prediction Horizon using the Company’s Governance, Financial and Risk data
Figure 3 – The Company Performance Prediction CUI-CPP) Reference Model

MPAI-CUI includes the Reference Software (RS), Conformance Testing (CT) and Performance Assessment (PA) Specifications of the AI-based Company Performance Prediction (CPP).

Multi-modal conversation (MPAI-MMC) uses AI to enable human-machine conversation emul­ating human-human conversation in completeness and intensity. It includes 5 Use Cases: Conversation with Emotion, Multimodal Question Answering, Unidirectional Speech Translation, Bidirectional Speech Translation and One-to-Many Unidirectional Speech Translation.

The figures below show the reference models of the MPAI-MMC Use Cases.

	Conversation with Emotion (CWE) enables a human to holds an audio-visual conver­sation using audio and video with a computational system that is impersonated by a synthetic voice and an animated face, both expressing emotion appropriate to the emotional state of the human.
Figure 4 – Conversation with Emotion
	Multimodal Question Answering (MQA) enables a user to request information using speech concerning an object the user displays and to receive the requested information from a computational system via synthetic speech.
Figure 5 – Multimodal Question Answering
	Unidirectional Speech Translation (UST) allows a user to select a language different from the one s/he uses and to get a spoken utterance translated into the desired language with a synthetic voice that optionally preserves the personal vocal traits of the spoken utterance.
Figure 6 – Unidirectional Speech Translation
	Bidirectional Speech Translation (BST) allows a human to hold a dialogue with another human. Both speech their own language and their translated speech is a synthetic speech that optionally preserves their personal vocal traits.
Figure 7 – Bidirectional Speech Translation
	One-to-Many Speech Translation (MST) enables a human to select a number of languages and have their speech translates to the selected languages using a synthetic speech that optionally preserves their personal vocal traits.
Figure 8 – One-to-Many Speech Translation

Currently, only the MPAI-MMC TS is available. Thereforethe

The 4 use cases considered are: Emotion Enhanced Speech, Audio Recording Preservation, Speech Restoration System and Enhanced Audioconference.

The figures below shows the reference models of the MPAI-CAE Use Cases. Note that an Implementation is supposed to run in the MPAI-specified AI Framework (MPAI-AIF).

	Emotion-Enhanced Speech (EES) enables a user to indicate a model utterance or an Emotion to obtain an emotionally charged version of a given utterance. In many use cases, emotional force can usefully be added to speech which by default would be neutral or emotionless,
Figure 9 – Emotion Enhanced Speech
	Audio Recording Preservation (ARP) Use Case enables a user to create of digital copies  of a digitised audio of open-reel magnetic tapes suitable for long-term preservation and for correct play back of the digitised recording (restored, if necessary).
Figure 10 – Audio Recording Preservation
	Speech Restoration System (SRS) enables a user to restore a Damaged Segment of an Audio Segment containing only speech from a single speaker. No filtering or signal processing is involved. Instead, replacements for the damaged vocal elements are synthesised using a speech model.
Figure 11 – Speech Restoration System
	Enhanced Audioconference Experience (EAE) enables a user to improve the auditory quality of audioconference experience by processing speech signals recorded by microphone arrays and  provide speech signals free from back­ground noise and acoustics-related artefacts .
Figure 12 – Enhanced Audioconference Experience

Currently, only the MPAI-CAE TS is available. Therefore

MPAI has 7 projects at different levels of development. For each of these a Goal is assigned.

One year ago today, MPAI could take stock of 3 months of work: an established organisation with the mission of developing Artificial Intelligence (AI)-based data coding standards, a first identification of a program of work, progression of several work items and a published Call for Technologies for one of them.

What can MPAI say today? That it has lived a very intense year and that it can declare itself satisfied of what it has achieved.

The first is that it has refined its method of work to make it solid but also capable to overcome problems plaguing other Standards Developing Organisations (SDO). A standard project goes through 8 stages, progression to a new stage requiring approval by the MPAI General Assembly. A Call for Technology is issued with Functional and Commercial Requirements.

The second is that it has developed 3 Technical Specifications (TS) that use AI to enable the industry to accelerate deployment of AI-based applications. Two words for each of them:

Context-based Audio Enhancement (MPAI-CAE) – supports 4 use cases:

1. Emotion-Enhanced Speech (EES) allows a user to give a machine a sentence uttered without emotion and obtain one that it is uttered with a given emotion, say, happy, or sad, or cheerful etc., or uttered with the colour of a specific model utterance.
2. Audio Recording Preservation (ARP) allows a user to preserve old audio tapes by providing a high-quality digital version and a digital version restored using AI with a documented set of irregularities found in the tape.
3. Speech Restoration System (SRS) allows a user to automatically recover damaged speech segments using a speech model obtained from the undamaged part of the speech.
4. Enhanced Audioconference Experience (EAE) improves a participant’s audioconference experience by using a microphone array and extracting the spatial attributes of the speakers with respect to the position of the microphone array to allow spatial representation of the speech signals at the receiver.

Compression and Understanding of Industrial Data (MPAI-CUI) support one use case: Company Performance Prediction. This gives the financial risk assessment industry new, powerful and extensible means to predict the performance of a company several years into the future in terms of company default probability, business discontinuity probability and adequacy index of company organisational model.

Multimodal Conversation (MPAI-MMC) – supports 5 use cases:

1. Conversation with Emotion (CWE) allows a user to have a full conversation with a machine impersonated by a synthetic speech and an animated face. The machine understands the emotional state of the user and its speech and face are congruent with that emotional state.
2. Multimodal Question Answering (MQA) allows a user to ask a machine via speech information about an object held in their hand and obtain a verbal response from the machine.
3. Unidirectional Speech Translation (UST) allows a user to express a verbal sentence in a language and obtain a verbal translation in another language that preserves the user’s vocal featurers.
4. Bidirectional Speech Translation (BST) allows two users to have a dialogue each using their own language and hearing the other user’s translated voice with that user’s native speech features.
5. One-to-Many Speech Translation (MST); allows a user to select a set of languages and have their speech translated to the selected languages with the possibility to decide whether to preserve or not their speech features in the translations.

The first MPAI Call for Technologies was issued on 16 December 2020 and concerned the AI Framework (MPAI-AIF) for creation and execution of AI Workflows (AIW) composed of AI Modules (AIM). These may have been developed in any environment using any proprietary framework for any operating system, be AI- and non-AI-based, implemented in hardware or software or in a hybrid hardware and software combination, for execute in MCUs up to HPC in local and distributed environments, and in proximity with other AIFs, irrespective of the AIM provider. The three TSs mentioned above rely on the MPAI-AIF TS for their implementations.

Finally, in 2021 MPAI has develop the Governance of the MPAI Ecosystem (MPAI-GME) TS. This lays down the rules governing the MPAI Ecosystem composed

1. MPAI developing standards.
2. Implem­enters developing implementations
3. The MPAI established and controlled not-for-profit MPAI Store where implementations are uploaded, checked for security, and tested for conformance.
4. MPAI-appointed performance assessors who assess that implementations are reliable and trustworthy.
5. Users who can access secure MPAI standard im­plemen­tations guar­an­teed for Conformance and Performance.

In 2020, MPAI has developed the four MPAI components – Technical Specifications, Reference Software, Conformance Testing and Performance Assessment – for Compression and Understanding of Industrial Data (MPAI-CUI). These, together with the other TSs, are published on the MPAI web site.

These are firm results for standards that industry can take up, but MPAI has carried out substantial more work preparing for the future:

MPAI-CAV: Connected Autonomous Vehicles

MPAI-EEV: AI-based End-to-End Video Coding

MPAI-EVC: AI-Enhanced Video Coding

MPAI-MCS: Mixed-reality Collaborative Spaces

MPAI-SPG: Server-based Predictive Multiplayer Gaming

This huge work has been carried out by a network of technical groups that MPAI thanks for their efforts and results.

Want to know more? Read “Towards Pervasive and Trustworthy Artificial Intelligence”, the book that illustrates the results achieved by MPAI in its 15 months of operation and the plans for the next 12 months.

The work has just begun. Become an MPAI member. Join the fun – build the future!

Another context of MPAI Audio Enhancement is preservation. Many audio archives urgently need to digitise their records, especially analogue magnetic tapes, because of their life expectancy is short if compared to paper records. International institutions (e.g., International Association of Sound and Audio-visuals Archives, IASA; World Digital Library, WDL; Europeana) have defined guidelines, sometimes only partly compatible, but appropriate international standards are lacking.

The Audio Recording Preservation (ARP) use case of the MPAI-CAE standard (CAE-ARP) opens the way to effectively respond to methodological questions of reliability with respect to audio recordings as documentary sources, while clarifying the concept of “historical faithfulness”. The magnetic tape carrier may hold important information: multiples splices; annotations (by the composer or by the technicians) and/or display several types of irregularities (e.g., corruptions of the carrier, tape of different colour or chemical composition).

AI can have a significant impact on cultural heritage because it can make its safeguarding sustainable by drastically changing the way it is preserved, accessed, added value. Audio archives, an important part of this heritage require important resources in term of people, time, and funding.

An important example of how AI can drastically reduce the resources necessary to preserve and make accessible analogue recordings is provided by CAE-ARP providing a workflow for managing open-reel tape audio recordings. It focuses on audio read from magnetic tapes, digitised and fed into a preservation system together with the data from a video camera pointed to the head reading the magnetic tape. The output of the restoration process is composed by a preservation master file that contains the high-resolution audio signal and several other information types created by the preservation process. The goal is to cover the whole “philologically informed” archival process of an audio document, from the active preservation of sound documents to the access to digitised files.

Figure 1 depicts the CAE-ARP workflow. Its operation is concisely described below.

Figure 1 – Audio recording preservation

1. The Audio Analyser and Video Analyser AIMs analyse the Preservation Audio File (a high-quality audio signal) and the Preservation Audio-Visual File (video of the reading head).
2. All detected Audio and Image irregularities are sent to the Tape Irregularity Classifier AIM, which selects those most relevant for restoration and access.
3. The Tape Audio Restoration AIM uses the irregularities to correct potential errors occurred at the time the audio signal was analogue-to-digital converted.
4. The Restored Audio File, the Editing List (used to produce the Restored Audio File, the Irregularity Images, and the Irregularity File containing information about the irregularities) are inserted in the Packager.
5. The Packager produces the Access Copy Files. These are used, as the name implies, to access the audio content and the Preservation Master Files, with the original inputs and data produced during the analysis, used for preservation.

The ARP workflow described above is complex and involves different audio and video competences. Therefore, the MPAI approach of subdividing complex systems in smaller components is well-suited to advance different algorithms and functionalities typically involving different professionals or companies.

Currently, ARP is limited to mono audio recordings on open-reel magnetic tape, The goal is to extend it to more complex recordings and additional analogue carriers such as audiocassettes or vinyl.

Standards for audio exist: MPEG-1 Audio layer II and layer III (so called MP3) and a slate of AAC standards serving all tastes offer efficient ways to store and transmit different types of mono, stereo and multichannel audio . MPEG-H offers ways to transmit and present 3D audio experiences.

Never before, if not at the level of company products, however, was there a standard whose goal is not to preserve audio quality at low bitrates, but to improve it or, as the name of the standard – “Context-based Audio Enhancement”, acronym MPAI-CAE – says, enhance it.

Of course there are probably as many ways to enhance audio as there are target users, so what does audio enhancement mean and how can a standard be produced for such a goal?

The magic word that changes the perspective is the word “context”. The MPAI-CAE standard identifies contexts in which audio can be enhanced. The next clarification comes from the fact that the standard is not monolithic, in other words, it identifies several contexts to which the standard can be applied.

Context #1: imagine that you have a sentence that you would like to be able to pronounce with a particular emotional charge: say, happy, or sad, or cheerful etc. or as if it were pronounced with the colour of a specific model utterance. If we were in a traditional encoder-decoder setting, there would be little to standardise. If you have the know how, you do it. If you don’t, you ask someone who has that know how to do it for you.

So, why should there be a standard for context #1?

To answer the question, I need to go back to a definition that I found years ago in the Encyclopaedia Britannica:

Standardisation, in industry: setting of guidelines that permit large production runs of component parts that are readily fitted to other parts without adjustment.

In practice the definition means that if there is a standard for nuts and bolts, and you have a standard nut, you can find someone who has the bolt to which your nut fits.

MPAI-CAE Context #1 is a straightforward application of the Encyclopaedia Britannica definition because it defines the components that can be assembled to make a system that lets you do one of the following:

1. It receives your vocal utterance without colour and pronounces it using the speech features of the model utterance
2. It receives your vocal utterance without colour, the indication of one or more emotions, the indication of a language and pronounces it with the particular emotion(s) and the “intonation” of the specified language.

There is one point that I must make clear. I said that the standard “defines the components” of the system, but I should have said that the “defines the interfaces of the components”. This is no different than the “nuts and bolts standard”. That standard defines neither the nuts nor the bolts. It defines the threading, i.e., the “interface” between the nut and the bolts.

Lets now go to a block diagram

 Figure 1 – Reference Model of Emotion Enhanced Speech

Here we see how the MPAI standardisation model works.

1. Speech Feature Analyser2 is a very sophisticated technology component that must be able to extract your speech features which are very specific of you and embedded deeply in your vocal utterances.
2. Emotion Feature Inserter is an even more sophisticated technology component because it must be able to take the Features of your Emotionless Speech, the Emotion, say, “cheerful” (whose semantics is defined by MPAI-CAE standard), and the Language, and generate Speech Features that convey your personal speech features, the cheerful Emotion, and the specifics of the selected language.
3. The Emotion Inserter, another very sophisticated component, receives the Speech Features from the Emotion Feature Inserter together with your Emotionless Speech and produces an emotionally charged vocal utterance according to your wishes.

A similar process unfolds for the upper branch of the diagram where is used. a model utterance.

In principle, each of the identified components – that MPAI calls AI Modules (AIM) – can be re-used in other context. We will see how that is done because this is just the first MPAI-CAE context. There will be soon opportunities to introduce other contexts,

In research, a technology that had attracted the interest of researchers decades ago and stayed at that level for a long time, may suddenly come into focus. This is the case of the collection of different technologies called Artificial Intelligence (AI). Although this moniker might suggest that machines are able to replicate the main human trait, in practice such techniques boil down to algorithmically sophisticated pattern matching enabled by training on large collections of input data.  Embedded today in a range of applications, AI has started affecting the life of millions of people and is expected to do so even more in the future.

AI provides tools to “get inside” the meaning of data to an extent not reached by previous technologies. The word “data” is used to indicate anything that represents information in digital form ranging from the US Library of Congress to a sequenced DNA, to the output of a video camera or an array of microphones, to the data generated by a company. Through AI, the number of bits required to represent information can be reduced, “anomalies” in the data discovered, and a machine can spot patterns that might not be immediately evident to humans.

AI is already among us doing useful things. There is keen commercial interest in implementing more AI-centric processes unleashing its full potential. Unfortunately, the way a technology leaves the initial narrow scientific scope to become mainstream and pervasive for products, services and applications is usually not linear nor fast. However, exceptions exist. Looking back to the history of MPEG, we can see digital media standards not only accelerated the mass availability of products enabled by new technologies, but also generated new products never thought of before.

In fact, the MPEG phenomenon was revolutionary because its standards were conceived to be industry neutral, and the process unfolded successfully because it had been designed around this feature. The revolution, however, was kind of “limited” because MPEG was confined to “media” (even though it tried to escape from that walled garden).

Here we talk about AI-centric data coding standards, which do not have such limitations. AI tools are flexible and can reasonably be adapted to any type of data. Therefore, as digital media standards have positively influenced industry and billions of people, so AI-based data coding standards are expected to have a similar, if not stronger impact. Research shows that AI-based data coding is generally more efficient than existing technologies for, e.g., data compression and description.

These considerations have led a group of companies and institutions to establish the Moving Picture, Audio and Data Coding by AI – MPAI – as an international, unaffiliated not-for-profit Standards Developing Organisation (SDO).

However, standards are useful to people and industry if they enable open markets. Still, the industry might invest hundreds of millions into the development of a standard, only to find that it is not practically usable or it is only accessible to a lucky few. In this case rather than enabling markets, the standard itself causes market distortion. This is a rather new situation for official standards, caused by the industry’s recent inability to cope with tectonic changes induced by technology and market. As a result, developing a standard today may appear like a laudable goal, but the current process can actually turn into a disappointment for industry. A standards development paradigm more attuned to the current situation is needed.

Therefore, to compensate for some standards organisations’ shortcomings in their handling of patents, the MPAI scope extends beyond the development of standards for a technology area to include Intellectual Property Rights guidelines.

Let’s briefly compare how the incumbent Data Processing (DP) technology and AI work. When they apply DP, humans study the nature of the data and design a priori methods to process it. When they apply AI, prior understanding of the data is not paramount – a suitably “prepared” machine is subjected to many possible inputs so that it can “learn” from the actual data what the data “means”.

In a sense, the results of bad training are similar in humans and machines. As an education with “bad” examples can make “bad” humans, a “bad”, i.e., insufficient, sectorial, biased etc. education makes machines do a “bad” job. The conclusion is that, when designing a standard for an AI-based application, the technical specification is not sufficient. So, MPAI’s stated goal to make AI applications interoperable and hence pervasive through standards is laudable, but the result is possibly perverse if ungoverned “bad” AI applications pollute a society relying on them.

For these reasons, MPAI has been designed to operate beyond the typical remit of a standards-developing organisation – albeit it fulfills this mission quite effectively, with five full-fledged standards developed in 15 months of operation. An essential part of the MPAI mission consists of providing the users with quantitative means to make informed decisions about which implementations should be preferred for a given task.

Thanks to MPAI, implementers have available standards that can be used to provide trustworthy products, applications and services, and users can make informed decisions as to which one is best suited to their needs. This will result in a more widespread acceptance of AI-based technology, paving the way for its benefits to be fully reaped by the society.

To know more you should read the book “Towards Pervasive and Trustworthy Artificial Intelligence” available from Amazon https://www.amazon.com/dp/B09NS4T6WN/