Ahmed Bouzid wrote an interesting post on frequent unrealistic expectations from Voice First (“Voice First Sucks!“) and has started another one listing meaningful Voice First Use Cases. In that second post, he lists 6 dimensions that need to be considered when considering the effectiness and acceptability of Use Cases for Voice / Speech.
I was kindly invited to comment and I jumped on the opportunity 🙂
Like Ahmed, I, too, find the lack of knowledge among far too many people in the Voice First Community of how ASR / Speech Recognition actually works frustrating (if not criminal!). I go on about this here.
Voice / Speech is an interface of its own and yes it is serial and invisible (and often relentless). Nevertheless, at the same time, there are already plenty of use cases where it can help avoid taxing the user’s memory (e.g. not having to list all options upfront and expecting the user to remember them all and choose 1) and, if the ASR has been trained well (i.e. on sufficiently large amounts of real-world and representative data), the user doesn’t even have to speak loudly nor enunciate clearly. Heck, the user doesn’t even need to be patient, if they are allowed to barge in and interrupt the system’s prompts / instructions / factoids. And thankfully, Voice interfaces won’t get offended if you ask them to repeat something 2, 3, 10 times, so users don’t have to be that focused either. Good Voice interfaces that is 🙂
Time Sensitivity is a bastard though! You pause for too long or at the wrong point in your utterance and you may find yourself down a path you hadn’t envisaged or wanted (Enter misrecognitions). But even that can be modulated by playing around with the various timeout settings (if you have access to them!) and ensuring there are enough implicit confirmations of what you want at critical points.
This is all part of Voice User Interface Design (or, as I call it, Explainable VUI Design); decisions you have to take early on, before the detailed design and certainly before you launch to millions of users.
Here is my own choice: Max Tegmark‘s Life 3.0 (which, by chance, appeared 2 more times in this list!).
My book choice: Max Tegmark’s Life 3.0
If I were to name a book that deeply impacted me from any decade, it would have been very difficult. Nevertheless, the following 3 books from 1991 – 1993 spring to mind:
The latest episode of Data Futurology features myself talk with Felipe Flores about how I got into AI, Machine Learning and Deep Learning (“When I discovered Artificial Neural Networks back in 1993, I thought I had discovered God!”) and later Data Science and Big Data Speech Analytics for Voice User Interface (VUI) and Voice First Design (for Speech IVRs and Voice Assistants).
In the podcast, I give an overview of all the different steps involved inVUI Design, the often tricky interaction with the different stakeholders (Coders, Business Units, Marketeers) and the challenges of working in an Agile PM framework, when skipping detailed Voice Design is suicide (and my worst nightmare!). I showcase some of the interesting and outright funny things you discover, when you analyse real-world human-machine dialogues taken from Speech IVR systems (hint, most original swearwords!) and I pinpoint the main differences between a Speech IVR and a Voice Assistant skill / action / capsule. I also contemplate on where Voice is heading now with the ubiquity of Voice First and the prevalance of skills that were developed by software engineers, web designers and Marketing copywriters, without VUI Design expertise, Linguistics Training or knowledge of how Speech Recognition works.
Data Futurology Podcast EP122 – Optimising Conversational AI Dialogue Systems with Dr. Maria Aretoulaki
I even provide tips on how to survive working for a Start-up and strategies on how to stay focused and strong when you run your own SME Business. Below is a brief checklist of some of the core qualities required and lessons I have learned working for Start-ups:
#122 – Optimising Conversational AI Dialogue Systems with Dr. Maria Aretoulaki
It was a delight speaking to Felipe! It made me think about and consolidate some of the – occasionally hard – lessons I have learned along the way about Voice Design, NLP, AI, ML and running a Business, so that others hopefully prepare for and ideally also avoid any heartaches!
The latest Boundless Podcast Episode is out! It features myself in a deep conversation with Richard Foster-Fletcher about Big Data and Speech Analytics for Voice First & Voice Assistants (but not only), BigTech, AI Ethics and the need for a new legal framework for Responsible AI and Explainable Machine Learning. Below is a snippet of the conversation:
Boundless Podcast Episode EP48 – snippet
“My data is being used and exploited, and I can do nothing about it. We need to modernise the legal system. Apart from all the ethical, moral discussions that need to be made, we need a legal system that takes into consideration the fact that intelligence doesn’t need to be visible to be acting against me.”
I wouldn’t call myself a technophobe. Quite the opposite. I was learning how to program (BASIC!) back in 1986 in the Summer after getting my degree in English Language & Literature; I was teaching computers how to translate between human languages and chatting online with mainframes in University buildings 2 kms away back in 1991; I was programming Artificial Neural Networks and using Parallel computers back in 1993; I was reading internet newsgroups and downloading music – very much legally! – on a web browser (Netscape!) in 1994; I was designing Voice interfaces already in 1996 and voice-controlled home assistants back in 1998. I have even been using LinkedIn since 2005.
Yet, I am very sceptical and pessimistic about our uninhibited sharing of personal data and sensitive information all day every day on multiple channels, to multiple audiences, much of it willingly, much more unwillingly, in the name of sharing and human connection, service and product personalisation and ultimately, far too often, monetisation.
What will that mean for our legacy as individuals? Who will own, control and curate all the generated data after our death? Who can clone us and why? Will there be a second life after death? Will there ever be freedom from a dictator or will there ever be any point in bequeathing anything in a will? These and many more questions are discussed in this episode. I had lots of fun recording this! Thank you so much to Richard for creating this🙏
You can listen to a snippet of the conversation here (Download)
You can listen to the full episode here or wherever you get your podcasts.
Alternatively, you can listen to it on YouTube (in 2 parts):
I recently read an article on a new approach to Common Sense understanding, which uses a combination of traditional, Good Old-Fashioned AI (GOFAI) symbolic and the latest data-intensive Machine Learning (ML) / Deep Learning neural network approaches to deal with the hard problem of human reasoning. Here’s a link to the article (with thanks to Phillip Hunter for the pointer!):
My favourite quote from the article is:
‘statistics ≠ understanding'”
That’s because (another favourite quote):
“common sense, like natural language, remains fundamentally fuzzy”
I was delighted to read about this research, especially because almost 30 years ago, when I was doing my PhD at the University of Manchester, I, too, realised that the only promising way to capture this fuzziness, ambiguity and complexity of language and meaning is through a hybrid approach, combining hand-crafted “rules” (human annotations, i.e. symbolic processing) with the automatic weight distribution and semi-supervised learning of a neural network (connectionist processing).
My PhD Thesis
Thus, I used text annotations generated by humans, which encoded morphosyntactic / grammatical, lexical-semantic and discourse pragmatic features of each sentence in a news article.
sentence annotations with discourse pragmatic features
I would then feed them into a basic feed-forward backpropagation neural network (ANN) that would calculate the degree of “importance” of each sentence in the whole article and generate a YES or NO answer to the question whether that specific sentence would be included (not necessarily verbatim) in the final summary of that news article.
ANN decides the degree of importance of a sentence in a summary
It was a neat idea, very imperfectly executed, as both the data set was not that large for today’s standards (1,100 sentences representing 55 news articles) and the ANN barely had 3 layers and the single hidden layer only had 30 units (so very skin-deep learning!).
My PhD thesis only scratched the surface. It’s awesome to see a similarly hybrid approach now gaining momentum! We now have both the huge data collections and the sophisticated Deep Learning algorithms to try out different things and better copy and simulate human intelligence in AI systems and, hence, achieve deeper understanding and generate more relevant and useful responses and actions. This will also contribute to more Explainable AI and, by extension, more Explainable Conversational AI for transparency and reusability in Voice User Interface (VUI) Design.
“it uses a neural network to recognize the colors, shapes, and materials of the objects and a symbolic system to understand the physics of their movements and the causal relationships between them. It outperformed existing models across all categories of questions.”
The motivation behind this approach was two-fold:
Deep Learning is superb at pattern matching, but bad at actually understanding the data it processes.
Symbolic Logic can capture and model really well human reasoning, correlations and interdependencies between actions, events, objects, people, and “common sense” or even objective rigorous Physics understanding.
Thus, the natural next step is to combine both and leverage the strengths of each in a Hybrid model:
“Deep Learning excels at scalability and pattern recognition; symbolic systems are better at abstraction and reasoning.”
It feels like yesterday, when in December 2008 I registered it with Companies House and became Company Director (with multiple hats).
My very first client project was for the NHS Business Authority on their EHIC Helpline (which hopefully will survive the Brexit negotiations). Back then, whenever I was telling anyone what my company does (VUI Design for Speech IVRs), I was greeted by blank stares of confusion or incomprehension. It did feel a bit lonely at times!
Many more clients and thousands of long hours, long days and working weekends since, here we are in December 2019 and I suddenly find myself surrounded by VUI Designers and Voice Strategists who have now seen the potential and inescapable nature of speech interfaces and have followed on my footsteps. I feel vindicated, especially since I started in Voice back in 1996 with my Post-Doc in Spoken Dialogue Management at the University of Erlangen! 😎 (Yet another thing I’m hugely grateful to the EU for!)
We started with Voice-First VUI Design back in 1996, well before Samsung’s BIXBY (2017), Google’s ASSISTANT (2016), Amazon’s ALEXA (2014), Apple’s SIRI (2010) and even before the world started using GOOGLE for internet searches (1998)!
It’s quite frustrating when I realise that many of these newcomers have never heard of an IVR (Interactive Voice Response) system before, but they will eventually learn. 🤓 In the past 25 years it was the developers who insisted could design conversational interfaces without any (Computational) Linguistics, Natural Language Processing (NLP) or Speech Recognition (ASR) background and didn’t need, therefore, a VUI Designer. And we were an allegedly superfluous luxury and rarity in those times. In the past couple of years it’s the shiny Marketing people, who make a living from their language mastery, and the edgy GUI Designers, who excell in visual design and think they can design voice interfaces too, but still know nothing about NLP or ASR.
What they don’t know is that, by modifying, for instance, just the wording of what your system says (prompt tuning), you can achieve dramatically better speech recognition and NLU accuracy, because the user is covertly “guided” to say what we expect (and have covered in the grammar). The same holds for tuned grammars (for out-of-vocabulary words), word pronunciations (for local and foreign accents), tuned VUI designs (for error recovery strategies) and tuned ASR engine parameters (for timeouts and barge-ins). It’s all about knowing how the ASR software and our human brain language software works.
Excited to see what the next decade is going to bring for DialogCONNECTION and the next quarter of a century for Voice! Stay tuned!
I was really interested to read the BDVA – Big Data Value Association‘s and euRobotics‘ recent report on “Strategic
Research, Innovation and Deployment Agenda for an AI PPP: A focal point
for collaboration on Artificial Intelligence, Data and Robotics“, which you can find here.
Of particular relevance to me was the Section on Physical and Human Action and Interaction
(pp. 39-41), which describes the dependencies, challenges and expected
outcome of coordinated action on NLP, NLU and multimodal dialogue
processing. The associated challenges are:
Natural interaction in unstructured contexts, which is the default in the case of voice assistants for instance, as they are expected to hold a conversation on any of a range of different topics and act on them
Improved natural language understanding, interaction and dialogue covering all European languages and age ranges, thus shifting the focus from isolated recognition to the interpretation of the semantic and cultural context, and the user intention
Development of verbal and non-verbal interaction models for people and machines, underlining the importance of gestures and emotion recognition and generation (and not only in embodied artificial agents)
Co-development of technology and regulation to assure safe interaction in safety-critical and unstructured environments, as the only way to assure trust and, hence, widespread citizen and customer adoption
The development of confidence measures for interaction and the interpretation of actions, leading to explanable AI and, hence, improved and more reliable decision-making
You can find the excellent and very comprehensive report here.
I recently attended The Association for Conversational Interaction Design (ACIXD) Brown bag “Challenges of Implementing Voice Control for Space Applications” presented by the NASA Authority in the field, George Salazar. George Salazar is Human Computing Interface Technical Discipline Lead at NASA with over 30 years of experience and innovation in Space applications. Among a long list of achievements, he was involved in the development of the International Space Station internal audio system and has been awarded several awards, including a John F. Kennedy Astronautics Award, a NASA Silver Achievement Medal and a Lifetime Achievement Award for his service and commitment to STEM. His acceptance speech for that last one brought tears to my eyes! An incredibly knowledgeable and experienced man with astounding modesty and willingness to pass his knowledge and passion to younger generations.
George Salazar’s Acceptance Speech
Back to Voice Recognition.
Mr Salazar explained how space missions slowly migrated over the years from ground control (with dozens of engineers involved) to vehicle control and from just 50 to 100s of buttons. This put the onus of operating all those buttons to the 4-5 person space crew, which in turn brought in speech recognition as an invaluable interface that would make good sense in such a complex environment.
Screenshot from George Salazar’s ACIxD presentation
Factors affecting ASR accuracy in Space
He described how they have tested different Speech Recognition (ASR) software to see which fared the best, both speaker-independent and speaker-dependent. As he noted, they all claim 99% accuracy officially but that is never the case in practice! He listed many factors that affect recognition accuracy, including:
background noise (speaker vs background signal separation)
multiple speakers speaking simultaneously (esp. in such a noisy environment)
intraspeaker speech variation due to psychological factors (as being in space can, apparently, make you depressed, which in turn affects your voice!), but presumably also to physiological factors (e.g. just having a cold)
Astronaut gender (low pitch in males vs high pitch in females): ASR software was designed for males, so male astronauts always had better error rates!
The effects of microgravity (physiological effects) on the voice quality, as already observed on the first flight (using templates from ground testing as the baseline), are impossible to separate from the environment and crew stress and can lead to a 10-30% error increase!
Screenshot from George Salazar’s ACIxD presentation
Even radiation can affect the ASR software, but also the hardware (computing power). As a comparison, AMAZON Alexa uses huge computer farms, whereas in Space they rely on slow “radiation-hardened” processors: they can handle the radiation, but are actually 5-10 times slower than commercial processors!
Screenshot from George Salazar’s ACIxD presentation
Solutions to Space Challenges
To counter all these negative factors, a few different approaches and methodologies have been employed:
on-orbit retrain capability: rendering the system adaptive to changes in voice and background noise, resulting in up to 100% accuracy
macro-commanding: creating shortcuts to more complex commands
redundacy as fallback (i.e. pressing a button as a second modality)
Screenshot from George Salazar’s ACIxD presentation
Critical considerations
One of the challenges that Mr Salazar mentioned in improving ASR accuracy is overadaptation or skewing the system to a single astronaut.
In addition, he mentioned the importance of Dialog Design in NASA’s human-centered design (HCD) Development approach. The astronauts should always be able to provide feedback to the system, particularly for error correction (Confusability leads to misrecognitions).
Screenshot from George Salazar’s ACIxD presentation
Screenshot from George Salazar’s ACIxD presentation
In closing, Mr Salazar stressed that speech recognition for Command and Control in Space applications is viable, especially in the context of a small crew navigating a complex habitat.
Moreover, he underlined the importance of trust that the ASR system needs to inspire in its users, as in this case the astronauts may literally be placing their lives onto its performance and accuracy.
Screenshot from George Salazar’s ACIxD presentation
Q & A
After Mr Salazar’s presentation, I couldn’t help but pose a couple of questions to him, given that I consider myself to be a Space junkie (and not in the sci-fi franchise sense either!).
So, I asked him to give us a few examples of the type of astronaut utterances and commands that their ASR needs to be able to recognise. Below are some such phrases:
zoom in, zoom out
tilt up, tilt down
pan left
please repeat
and their synonyms. He also mentioned the case of one astronaut who kept saying “Wow!” (How do you deal with that!)
I asked whether the system ever had to deal with ambiguity in trying to determine which component to tilt, pan or zoom. He answered that, although they do carry out plenty of confusability studies, the context is quite deterministic: the astronaut selects the monitor by going to the monitor section and speaking the associated command. Thus, there is no real ambiguity as such.
Screenshot from George Salazar’s ACIxD presentation
My second question to Mr Salazar was about the type of ASR they have gone for. I understood that the vocabulary is small and contained / unambiguous, but wasn’t sure whether they went for speaker-dependent or speaker-independent recognition in the end. He replied that the standard now is speaker-independent ASR, which however has been adapted to a small group of astronauts (i.e. “group-dependent“). Hence, all the challenges of distinguishing between different speakers with different pitch and accents, all against the background noise and the radiation and microgravity effects! They must be really busy!
It was a great pleasure to listen to the talk and an incredible and rare honour to get to speak with such an awe-inspiring pioneer in Space Engineering!
This week I bumped into Echodadand “the world’s harshest critics of Alexa skills“, i.e. his 2 kids. Hat tip for the idea by the way!! 😀
I read with particular interest his Medium post on “How Voice Interfaces are infiltrating society and changing our children“, a mockingly deceptive title for a list of witty repartees to all the usual arguments against Voice Assistants and speech interfaces being used by children. Highly enjoyable! 🙂
He addresses the perennial argument about children unlearning good manners, when they interact with VAs too much, and talks about his own experience with his kids assuring us that, actually, children know the difference between a human and a Voice Assistant. Phew, so there is no need to panic that they are going to suddenly be more and more rude to their teachers and grandparents and, hence, no need to reprogram voice interfaces to insist on “Pleases” and “Thank-yous” (Phew, again). I am fully in agreement there.
Similarly, he discusses the argument that children unlearn the importance and necessity of patience in the face of inescapable delays and life’s frustrations, when they can’t instantly get what they want (an answer to an obscure question about the animal kingdom or their favourite YouTube video). This seasoned dad knows that patience is something all kids have always struggled with and have to continually practise at that age. It is part of learning and growing up. And I can attest to that, too. (Oh the tantrums!)
Echodad & “the world’s harshest critics of Alexa skills”
However, in contrast to him, I am not at all comfortable or happy with the idea that my kid would go straight to Alexa or Google or Bixby for an answer to his Space or Transport questions and not even bother approach me. I may well take recourse to voice search myself for the answer, but I want to be the Gatekeeper to the Information Castle and not a passive bystander in his quests for knowledge. At least that’s what I consider as one of my – traditional – roles as a parent and I would like him to think that way too, at least until he is older. And it’s not just a hunch: having studied Psycholinguistics and Child Developmental Psychology at University in an earlier life (last century though!), I know the paramount importance of parental input, continuous feedback and playful interaction in the child’s quality and rate of learning anything, but particularly language.
In this respect, Voice Assistants, Voicebots, Chatbots and Voice Search itself have changed parent-child roles and interaction forever, and neither I nor many people can predict the long-term effect of this societal development. Kids are already and will become more and more independent, for sure, and that is not a negative thing in itself, but will that at the same time make me – as a parent – less dependable and approachable? Will he automatically assume I’m lacking in encyclopaedic knowledge and he that lessen his respect for me? And will the kid get the same quantity and quality of validation cues from Alexa or Google Assistant, as he would have got of me? More soberingly, will it teach kids that they don’t need to rely on or even invest much in relationships with people, because interacting with technology is easier, more fun and safer and more efficient? It’s not just the relationship to the parents we should be worried about, but all other family, work, peer and romantic relationships as well.
I am all for the use of voice interfaces and spoken dialogue systems by adults, e.g. myself, naturally, and my own clients; VUIs have been my bread and butter for decades! Nonetheless, at the risk of sounding like a regressive technophobe, I am vehemently against its use by children and will keep mine shielded for as long as I can manage! It’s already hard to keep him from doing a regular Google voice search for “pictures for buses” (sic). 🤔 😅
