#audio datasets

[globosetechnologysolutions1]

Simplify audio dataset collection and annotation to accelerate AI advancements in speech recognition, multilingual applications, and sentiment analysis. Focus on high-quality, diverse data to ensure reliable and impactful AI solutions.

[gts-ai]

Unlock the full potential of AI by harnessing the power of sound. High-quality audio datasets are the foundation for groundbreaking innovations in speech recognition, voice assistants, and more. By using diverse and meticulously curated audio data, your AI models can achieve superior accuracy and responsiveness. At GTS.ai, we specialize in providing the audio datasets that drive AI advancements, enabling your systems to interact more naturally and effectively with users. Elevate your AI projects with the sound data they need to innovate and succeed.

[pooja1gts]

Audio datasets are often used for tasks such as speech recognition, speaker identification, music classification, and audio event detection. They may also include associated metadata, such as the language spoken, the gender and age of the speaker, the genre of the music, and the location and time of the recording visit our website:

[girlballs]

i love seeing this argument pop up because it's like. functionally how is this different than just taking inspiration from someone

also maybe i'm an outlier in that i actually understand how generative AI works but if literally anyone was capable of actually training a model to convincingly generate novel audio off such a small dataset i'd just be really fucking impressed more than anything

[mariacallous]

On Saturday, an Associated Press investigation revealed that OpenAI's Whisper transcription tool creates fabricated text in medical and business settings despite warnings against such use. The AP interviewed more than 12 software engineers, developers, and researchers who found the model regularly invents text that speakers never said, a phenomenon often called a “confabulation” or “hallucination” in the AI field.

Upon its release in 2022, OpenAI claimed that Whisper approached “human level robustness” in audio transcription accuracy. However, a University of Michigan researcher told the AP that Whisper created false text in 80 percent of public meeting transcripts examined. Another developer, unnamed in the AP report, claimed to have found invented content in almost all of his 26,000 test transcriptions.

The fabrications pose particular risks in health care settings. Despite OpenAI’s warnings against using Whisper for “high-risk domains,” over 30,000 medical workers now use Whisper-based tools to transcribe patient visits, according to the AP report. The Mankato Clinic in Minnesota and Children’s Hospital Los Angeles are among 40 health systems using a Whisper-powered AI copilot service from medical tech company Nabla that is fine-tuned on medical terminology.

Nabla acknowledges that Whisper can confabulate, but it also reportedly erases original audio recordings “for data safety reasons.” This could cause additional issues, since doctors cannot verify accuracy against the source material. And deaf patients may be highly impacted by mistaken transcripts since they would have no way to know if medical transcript audio is accurate or not.

The potential problems with Whisper extend beyond health care. Researchers from Cornell University and the University of Virginia studied thousands of audio samples and found Whisper adding nonexistent violent content and racial commentary to neutral speech. They found that 1 percent of samples included “entire hallucinated phrases or sentences which did not exist in any form in the underlying audio” and that 38 percent of those included “explicit harms such as perpetuating violence, making up inaccurate associations, or implying false authority.”

In one case from the study cited by AP, when a speaker described “two other girls and one lady,” Whisper added fictional text specifying that they “were Black.” In another, the audio said, “He, the boy, was going to, I’m not sure exactly, take the umbrella.” Whisper transcribed it to, “He took a big piece of a cross, a teeny, small piece … I’m sure he didn’t have a terror knife so he killed a number of people.”

An OpenAI spokesperson told the AP that the company appreciates the researchers’ findings and that it actively studies how to reduce fabrications and incorporates feedback in updates to the model.

Why Whisper Confabulates

The key to Whisper’s unsuitability in high-risk domains comes from its propensity to sometimes confabulate, or plausibly make up, inaccurate outputs. The AP report says, "Researchers aren’t certain why Whisper and similar tools hallucinate," but that isn't true. We know exactly why Transformer-based AI models like Whisper behave this way.

Whisper is based on technology that is designed to predict the next most likely token (chunk of data) that should appear after a sequence of tokens provided by a user. In the case of ChatGPT, the input tokens come in the form of a text prompt. In the case of Whisper, the input is tokenized audio data.

The transcription output from Whisper is a prediction of what is most likely, not what is most accurate. Accuracy in Transformer-based outputs is typically proportional to the presence of relevant accurate data in the training dataset, but it is never guaranteed. If there is ever a case where there isn't enough contextual information in its neural network for Whisper to make an accurate prediction about how to transcribe a particular segment of audio, the model will fall back on what it “knows” about the relationships between sounds and words it has learned from its training data.

According to OpenAI in 2022, Whisper learned those statistical relationships from “680,000 hours of multilingual and multitask supervised data collected from the web.” But we now know a little more about the source. Given Whisper's well-known tendency to produce certain outputs like "thank you for watching," "like and subscribe," or "drop a comment in the section below" when provided silent or garbled inputs, it's likely that OpenAI trained Whisper on thousands of hours of captioned audio scraped from YouTube videos. (The researchers needed audio paired with existing captions to train the model.)

There's also a phenomenon called “overfitting” in AI models where information (in this case, text found in audio transcriptions) encountered more frequently in the training data is more likely to be reproduced in an output. In cases where Whisper encounters poor-quality audio in medical notes, the AI model will produce what its neural network predicts is the most likely output, even if it is incorrect. And the most likely output for any given YouTube video, since so many people say it, is “thanks for watching.”

In other cases, Whisper seems to draw on the context of the conversation to fill in what should come next, which can lead to problems because its training data could include racist commentary or inaccurate medical information. For example, if many examples of training data featured speakers saying the phrase “crimes by Black criminals,” when Whisper encounters a “crimes by [garbled audio] criminals” audio sample, it will be more likely to fill in the transcription with “Black."

In the original Whisper model card, OpenAI researchers wrote about this very phenomenon: "Because the models are trained in a weakly supervised manner using large-scale noisy data, the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination). We hypothesize that this happens because, given their general knowledge of language, the models combine trying to predict the next word in audio with trying to transcribe the audio itself."

So in that sense, Whisper "knows" something about the content of what is being said and keeps track of the context of the conversation, which can lead to issues like the one where Whisper identified two women as being Black even though that information was not contained in the original audio. Theoretically, this erroneous scenario could be reduced by using a second AI model trained to pick out areas of confusing audio where the Whisper model is likely to confabulate and flag the transcript in that location, so a human could manually check those instances for accuracy later.

Clearly, OpenAI's advice not to use Whisper in high-risk domains, such as critical medical records, was a good one. But health care companies are constantly driven by a need to decrease costs by using seemingly "good enough" AI tools—as we've seen with Epic Systems using GPT-4 for medical records and UnitedHealth using a flawed AI model for insurance decisions. It's entirely possible that people are already suffering negative outcomes due to AI mistakes, and fixing them will likely involve some sort of regulation and certification of AI tools used in the medical field.

[ao3scrapesearch]

Hey so I'm not very tech savvy but I was wondering if adding random silly lines or just something that makes no sense between paragraphs/sentences on our fics can poison AI if the fics are scraped?? I tried something by adding some random lines with white text between paragraphs of my fic which don't show up on default ao3 mode but they are a part of the text nonetheless. Of course that'll involve more efforts on part of the writers to add lines and format the white text using html and workskins but if it does turn out to be effective it might make ao3 less lucrative for AI scraping if a major amount of works contain this and it'll make it harder for AI training. It does have drawbacks that it'll only work on default mode so anyone using dark skin on ao3 might have to switch to be able to read properly and it'll make works less accessible to readers who use text to audio if there are random lines in between but what other options are we left with if even archive locking our works doesn't work??

You absolutely could, but there are limitations to that.

For one, like you said, you're making your work inaccessible to certain readers. That's fully within your rights, though I think most of us strive not to exclude people using screen readers.

Second, from what I know, when you download a dataset like this and intend to use it to train an AI model, you first go through the dataset looking for obvious junk data and toss that out. So if you're putting something that is clearly not real fanfic in there, any decent data analyst is probably going to spot it and toss your fic. If that's your goal, that's a win for you. Personally, if I'm making the effort to inject poison data, my goal is to be included in the training data used so I can trash the model, so I don't want it to be obvious.

Third, I don't see anything explicitly in AO3's TOS against adding data poison in this way, but I don't see them endorsing doing that either. It feels like a grey area to me, and I'm not sure you're allowed to do it, so I am not recommending anyone do this. Rest of this post is theoretical.

So theoretically, how I would do it is putting the junk data at the end of the fic/chapter. Hide it like you're saying, by changing the font and/or background color of the section with CSS. Then put a nice, clear message right after the chapter ends and the junk data starts, something like, "Hey, readers! This chapter is over. Turn off your screen reader and move to the next chapter now." That gives your real humans a warning and stops them from being confused or wasting their time. Then dump your poison. You can also write something in the beginning A/N, I believe. I know this most recent scraper never ever pulled data from the author's notes, so the AI wouldn't see anything you put in that section.

Scrapers are typically pulling your work without the workskin enabled, so for formatting, you're really just trying to make it look nice for your real readers so they don't have to see your poison.

As far as actual poison, my suggestions:

Your own writing or writing you have explicit permission to use, so you're not breaking anyone's copyright. Easy mode: jumbled paragraphs of your own past works for any fandom except the one your current fic is for.

As mentioned above, don't put absolute nonsense in there. If it's bad enough, it'll be spotted and filtered out. Like, if it's not even real words, anyone feeding it to AI is probably going to catch that and toss your data out, excluding it from the model. It might be fine if it's all real words, but not in any sensible order. Not sure on that. But don't just insert keysmashes if you want your data to be used in the AI training.

Terrible crackfic would be good. So would writing for a completely different fandom and different tags. The writing should not fit well with the tags you use for the fics. (So if the real fic is tagged Fluff and Alternative Universe - Coffee Shop, your poison should not include that. Make the poison a hurt no comfort canon-compliant fic or something else different.)

Keep in mind you should not be putting E-rated data poison in a G-rated fic. Real humans may still see this no matter how much you hide it, particularly if they download a PDF copy of your fic. If it's content that requires a warning per AO3's rules (explicit content, graphic violence, etc), you do still have to tag for that, even if it's designed to be invisible to humans.

Use unique writing, so even if someone later using it for AI catches it once, they can't just search for the exact wording you used in one fic and easily filter out all the rest of your poison. Again, this is if you want to be included in the AI training to throw the model off.

Again, theoretically, if I were going to do this, this is the CSS code I might use for my poison section of the fic:

#workskin .fuckai { background: #333333; color: #333333; font-size: 1%; }

It would theoretically look like a weird grey gap to mobile users or be nearly invisible to desktop users, even if it contained, say... 1,000 additional words.

Finally, scrapers are trying to grab millions of fics from AO3 when they do it. They're not looking closely at 13 million fics. They're only searching for the most obvious junk. So the only reason you would want to hide it like that is to make a better experience for your real readers. You don't need to hide it to get it into a scraper's AI model.

[olderthannetfic]

AI and queerness: Thank you! AI is great for small domain tasks with a huge number of variables that classic algorithms would struggle to consider all and would be near impossible for humans to fine tune properly (audio speech synthesis or cancer cell diagnostic for example). It sucks at anything else. AI will only replicate and amplify the biases in the training data. You don't even have to maliciously slant training data. Subconscious biases will show up just as much. (1/2)

AI and queerness (2/2). That's why it bugs me when people talk about AI as if it will solve anything. No, hiring processes won't get less biased because a rational and logical AI makes the choices. It makes the choices that seem statistically sensible based on the biased training data given by the company! Also LLMs, they don't tell you what's true, they tell you what sounds statistically nice. That's why they impress. They sound good and confident. It would be a liar, if it was sentient!

And even aside from the stuff that's biased in the way we usually mean, there's the false positives for the big thing vs. the small thing issue.

If I like the more popular thing, I might have to scroll through 10% junk to find it vs. 90% junk for the unpopular thing. AI isn't going to prioritize queer things unless it's explicitly programmed to do so and trained on an appropriate dataset because the world overall is very straight.

(Can you tell I've been shopping for romance novels on kindle? Haha.)

[lyingbard]

A Sample for the Knowing Ones

I realized I hadn't posted an audio sample from this version yet so here you go! You can definitely hear some Joshua in there which is promising. Hopefully this encourages you to download LyingBard and give it a shot.

Just to recap for those new around here, this TTS is trained on a much smaller dataset and with a much worse model than LyreBird actually used, so it sounds simliar but differentce. One day I'd like to take another crack at this and see if I can get something that sound just like the real deal, but that'll have to wait.

[globosetechnologysolutions1]

An audio dataset is a set of sound files used as a training resource for machine learning models, mostly in speech recognition, voice assistants, and sentiment analysis. 

[pooja1gts]

From Soundwaves to Insights: Unleashing the Potential of Audio Datasets in AI

Introduction

Audio datasets has become an increasingly valuable resource in the field of artificial intelligence (AI). The ability to analyse and extract meaningful insights from soundwaves opens up a wide range of applications, including speech recognition, music analysis, acoustic event detection, and environmental monitoring. However, harnessing the full potential of audio datasets in AI requires overcoming several challenges, such as data collection, annotation, and preprocessing. In this article, we delve into the world of audio datasets, exploring their significance, the techniques used for their creation, and the ways in which they can be leveraged to drive innovation in AI.

The Importance of Audio Datasets in AI

1.The Power of Sound: This section highlights the unique value of audio data in AI applications. It explores how sound carries valuable information that complements other types of data, such as text and images. We discuss the advantages of audio data in capturing nuances of human communication, emotion, and environmental context. Furthermore, we explore the role of audio datasets in advancing speech recognition, audio classification, and sound source separation tasks, showcasing their potential impact in various domains.

2. Challenges in Audio Data Collection: Collecting high-quality audio datasets poses several challenges. This subheading focuses on the intricacies of audio data collection, including considerations such as recording equipment, environmental conditions, and ethical considerations. We discuss techniques for capturing diverse audio sources, such as microphones, acoustic sensors, and even smartphones. Additionally, we address the need for large-scale and diverse datasets to ensure robust AI models capable of generalising to real-world scenarios.

Preprocessing and Annotation of Audio Datasets

1. Audio Preprocessing: Audio data often requires preprocessing to enhance its quality and extract meaningful features. This section explores techniques such as noise reduction, signal normalisation, and audio segmentation to prepare audio datasets for AI applications. We discuss the challenges of handling background noise, reverberation, and varying recording conditions. Additionally, we explore the role of feature extraction methods, such as spectrograms and mel-frequency cepstral coefficients (MFCCs), in representing audio data effectively for subsequent analysis and modelling.

2.Annotation and Labelling: Annotating audio datasets with relevant labels and metadata is essential for supervised learning and model training. This subheading delves into the various methods used for audio annotation, including manual labelling, automatic speech recognition, and crowd-sourcing. We discuss the challenges of annotating audio data, such as dealing with multiple speakers, overlapping speech, and complex audio events. Furthermore, we explore the potential of weakly supervised and semi-supervised approaches in alleviating the annotation burden while maintaining dataset quality.

Conclusion:

In conclusion, audio datasets hold immense potential in driving innovation and advancement in AI. This article has shed light on the importance of audio data, exploring its unique value and the challenges associated with collecting, preprocessing, and annotating audio datasets.

As audio data continues to play a vital role in AI, it is crucial to invest in further research and development to overcome challenges and ensure the availability of high-quality, diverse, and well-annotated audio datasets. By doing so, we can unleash the true potential of soundwaves and pave the way for exciting advancements in AI-driven audio analysis and understanding.

[counttwinkula]

when i was a kid i got really into mashups. it started with american edit which was an album that took (almost) every song off american idiot and mashed it up with all sorts of stuff. i still love a good mashup, but the culture got kind of weird—why do they all have to be titled "x but it's y"—and also it's hard to access mashups because they're always getting uploaded and taken down due to copyright issues (this is also the fate of many of the best cupcakke remixes)

my mom makes collages. it's one of her main hobbies, but she's only shown them a handful of times. collage, photomontage, and assemblage are art forms that really rose to prominence in the early 20th century—see hannah höch—in part due to the increased availability of print media. now that these pieces of visual culture were being mass produced it was way more conceivable to tear them, cut them up, manipulate them, and thereby create a new work of art

mashups and collage are both sort of rogue art forms, because they both deal in repurposing a piece of culture that has been put out into the world. mashups are easier to regulate in our ecosystem of DMCA takedowns and the general use of more mainstream media, whereas it's pretty hard to prevent people from taking any print media, cutting it up, and gluing it back together in a different way

there's also the fact that if i'm listening to "toxic but it's change (in the house of flies)" i can immediately identify britney spears and deftones in that mix, whereas if i'm looking at any given collage i am far less likely to be able to identify the many pieces of which it has been composed. i'm not sure that the editors of cosmopolitan, nor the photographers, models, designers, directors of photography, etc. would necessarily appreciate my mom taking their magazine covers and using them in her art projects, but the thing about these "remix" art forms is that the original creators do not get to consent to the remixing of their contribution

i feel like i have to preface what i'm about to say by disclaiming that i am not "pro-AI" by any stretch of the imagination. well that's not true. if you stretch your imagination by way of "we piss on the poor" then you'll misunderstand/misconstrue me. i've never used chatGPT and never plan on it. i find most uses of AI embarrassing, uninteresting, and annoying. (i hate that we call it AI but it's the term that people recognize so here we are.) on the flip side, i find most arguments against AI embarrassing, lacking in nuance, and annoying. one of my research interests is fear, particularly its political uses, and i think there's no question that the reactions against AI have all the identifying features of a moral panic

did you know that when photography was invented there was a similar moral panic? people proclaimed that this was the death of painting. and you know what's crazy? it wasn't! without photography challenging the place of painting, we wouldn't have any of the many movements of modern art that arose in the 20th century. collage arose within cubism and dada, and the very technologies that enabled this art form is what led walter benjamin to write his famous "the work of art in the age of mechanical reproduction"—which you can (currently) read courtesy of MIT

tl;dr: benjamin is pro-reproduction of art! you may also like to know, if you didn't already, that benjamin was a marxist and antifascist

i feel like i'm meandering so let me get to my point: this fixation on AI generated images as partaking in "theft" is bogus. the incorporation of images into a dataset without the consent of the creator is only theft if the creation of a collage is also theft. this relies on stable ideas of artistic ownership that challenge art forms which we wouldn't problematize when they appear in other contexts. a music producer taking an audio file of someone's song and turning it into a new song does not remove access to the old song, nor is it passing the new product off as a replacement for the existing work

i am never going to argue that the role AI-generated images are currently playing in our media landscape is not without its ethical problems! i am particularly offended by the way many people are posting AI-generated images and not labeling them as such. i think that's really weird. i think the lack of transparency about AI does present harm. i also respect fears of human artists losing work due to corporate reliance on AI, but i think that the immense focus on the "humanity" of this work is misplaced over the economic harm it poses, a danger shared by industries beyond the creation of art/visual culture (other users have voiced this far more articulately than i dare to)

the one concern that i understand regarding this notion of theft is that it would ostensibly make it possible for users to reproduce works in the style of an artist and sell what are essentially forgeries. this is a shitty thing to do, i am not going to argue with that. the problem is that this is not something unique to the generation of images through AI; i've been around long enough to hear of major companies stealing artists' work for products, as well as people ripping others off through practices of tracing or reposting

recently i saw people getting up in arms about the use of AI to generate images in the style of hayao miyazaki. not only does a figure as acclaimed as miyazaki not need random netizens standing up for him, it's been popular to make art in the style of his movies for years and years. (remember when breath of the wild came out? there was so much miyazaki-style BOTW fanart.) i don't see how individual artists learning how to recreate his style manually should be any different from the generation of images emulating his work, except for fetishized notions of labor and individuality expressed in the human act of artistic creation

returning to the issue of theft: can we really be mad that the catalogue of film stills from miyazaki's œuvre have been added to a data set without his consent? or works of individual artists emulating him? how is this process of absorbing a body of visual stimuli and recreating something modeled after them any different when performed by a machine instead of a human? is the fan artist not also digesting every miyazaki (and miyazaki-style) image they have seen in the act of emulating his style? setting aside the agency of the artist in how they replicate style, are they not reorganizing existing stylistic elements into a new work, like the machine is doing? how is any of this any different regarding the consent of the original artist than the creation of a collage?

and if this issue of replicating style bothers you, as the aforementioned examples of people producing AI-generated forgeries touches upon, i question the viability of "style" as something worth protecting through systems such as copyright. others have argued better than i can that copyright laws predominantly protect major corporations, who often weaponize these laws against the small artists opponents of AI claim to wish to protect. ownership of a style is, to be frank, a ludicrous concept, and one we cannot regulate in a manner which only protects small artists when they are appropriating styles associated with established corporations. how do we even begin to isolate "style" into characteristics in a meaningful way to prevent these issues? by and large, i think we once again encounter the perennial problem of internet leftism, where many purported leftists do not wish to challenge received notions of things such as "ownership" or "intellectual property" and find rationales based upon the status quo which they can dress up to seem progressive when taken at face value, but which crumble upon further investigation

i am as bored, annoyed by, and wary of the incursion of AI into every sphere, the way certain industries and fields take for granted that we are all interested in the implementation of this technology. at best, i am ambivalent about AI. that said, this is a tool that exists, and while we can desire and push for regulations around its use, we're simply not going to be able to put it back in the box. and frankly, some of the arguments surrounding what is wrong with AI image-generation are misguided and ill-conceived

[jcmarchi]

Arsham Ghahramani, PhD, Co-founder and CEO of Ribbon – Interview Series

New Post has been published on https://thedigitalinsider.com/arsham-ghahramani-phd-co-founder-and-ceo-of-ribbon-interview-series/

Arsham Ghahramani, PhD, Co-founder and CEO of Ribbon – Interview Series

Arsham Ghahramani, PhD, is the co-founder and CEO of Ribbon. Based in Toronto and originally from the UK, Ghahramani has a background in both artificial intelligence and biology. His professional experience spans a range of domains, including high-frequency trading, recruitment, and biomedical research.

Ghahramani began working in the field of AI around 2014. He completed his PhD at The Francis Crick Institute, where he applied early forms of generative AI to study cancer gene regulation—long before the term “generative AI” entered mainstream use.

He is currently leading Ribbon, a technology company focused on dramatically accelerating the hiring process. Ribbon has raised over $8 million in funding, supported over 200,000 job seekers, and continues to grow its team. The platform aims to make hiring 100x faster by combining AI and automation to streamline recruitment workflows.

Let’s start at the beginning — what inspired you to found Ribbon, and what was the “aha” moment that made you realize hiring was broken?

I met my co-founder Dave Vu while we were both at Ezra–he was Head of People & Talent, and I was Head of Machine Learning. As we rapidly scaled my team, we constantly felt the pressure to higher quickly, yet we lacked the right tools to streamline the process. I was early to AI (I completed my PhD in 2014, long before AI became mainstream), and I had an early understanding of the impacts of AI on hiring. I saw firsthand the inefficiencies and challenges in traditional recruitment and knew there had to be a better way. That realization led us to create Ribbon.

You’ve worked in machine learning roles at Amazon, Ezra, and even in algorithmic trading. How did that background shape the way you approached building Ribbon?

At Ezra, I worked on AI health tech, where the stakes couldn’t be higher–if an AI system is biased, it can be a matter of life or death. We spent a lot of time and energy making sure that our AI was unbiased, as well as developing methods to detect and mitigate bias. I brought over those techniques to Ribbon, where we use these techniques to monitor and reduce bias in our AI interviewer, ultimately creating a more equitable hiring process.

How did your experience as a candidate and hiring manager influence the product decisions you made early on?

Finding a job is a grueling process for junior candidates. I remember, not too long ago, being a junior candidate applying to many jobs. It’s only become harder since then. At Ribbon, we have deep empathy for job seekers. Our Voice AI is often the first point of contact between a company and a candidate, so we work hard to make this experience positive and rewarding. One of the ways we do that is by ensuring candidates chat with the same AI throughout the entire hiring process. This consistency helps build trust and comfort—unlike traditional processes where candidates are passed between multiple people, our AI provides a steady, familiar presence that helps candidates feel more at ease as they move through interviews and assessments.

Ribbon’s AI conducts interviews that feel more human than scripted bots. Tell us more about Ribbon’s adaptive interview flow. What kind of real-time understanding is happening behind the scenes?

We have built five in-house machine learning models and combined them with four publicly available models to create the Ribbon interview experience. Behind the scenes, we are constantly evaluating the conversation and combining this with context from the company, careers pages, public profiles, resumes, and more. All of this information comes together to create a seamless interview experience. The reason we combine so much information is that we want to give the candidate an experience as close to a human recruiter as possible.

You highlight that five minutes of voice can match an hour of written input. What kind of signal are you capturing in that audio data, and how is it analyzed?

People generally speak quite fast! Most job application processes are very tedious, tasking you with filling out many different forms and multiple-choice questions. We’ve found that 5 minutes of natural conversation equates to around 25 multiple-choice questions. The information density of voice conversation is hard to beat. On top of that, we are collecting other factors, such as language proficiency and communication skills.

Ribbon also acts as an AI-powered scribe with auto-summaries and scoring. What role does interpretability play in making this data useful—and fair—for recruiters?

Interpretability is at the core of Ribbon’s approach. Every score and analysis we generate is always tied back to its source, making our AI deeply transparent.

For example, when we score a candidate on their skills, we’re referencing two things:

The original job requirements and

The exact moment in the interview that the candidate mentioned a skill.

We believe that the interpretability of AI systems is deeply important because, at the end of the day, we are helping companies make decisions, and companies like to make decisions based on concrete data. Something we believe is critical for both fairness and trust in AI-driven hiring.

Bias in AI hiring systems is a big concern. How is Ribbon designed to minimize or mitigate bias while still surfacing top candidates?

Bias is a critical issue in AI hiring, and we take it very seriously at Ribbon. We’ve built our AI interviewer to assess candidates based on measurable skills and competencies, reducing the subjectivity that often introduces bias. We regularly audit our AI systems for fairness, utilize diverse and balanced datasets, and integrate human oversight to catch and correct potential biases. Our commitment is to surface the best candidates fairly, ensuring equitable hiring decisions.

Candidates can interview anytime, even at 2 AM. How important is flexibility in democratizing access to jobs, especially for underserved communities?

Flexibility is key to democratizing job access. Ribbon’s always-on interviewing allows candidates to participate at any time convenient for them, breaking down traditional barriers such as conflicting schedules or limited availability, which is especially beneficial for working parents and those with non-traditional hours. In fact, 25% of Ribbon interviews happen between 11 pm and 2 am local time.

This is especially crucial for underserved communities, where job seekers often face additional constraints. By enabling round-the-clock access, Ribbon helps ensure everyone has a fair chance to showcase their skills and secure employment opportunities.

Ribbon isn’t just about hiring—it’s about reducing friction between people and opportunities. What does that future look like?

At Ribbon, our vision extends beyond efficient hiring; we want to remove friction between individuals and the opportunities they’re suited for. We foresee a future where technology seamlessly connects talent with roles that align perfectly with their abilities and ambitions, regardless of their background or network. By reducing friction in career mobility, we enable employees to grow, develop, and find fulfilling opportunities without unnecessary barriers. Faster internal mobility, lower turnover, and ultimately better outcomes for both individuals and companies.

How do you see AI transforming the hiring process and broader job market over the next five years?

AI will profoundly reshape hiring and the broader job market in the next five years. We expect AI-driven automation to streamline repetitive tasks, freeing recruiters to focus on deeper candidate interactions and strategic hiring decisions. AI will also enhance the precision of matching candidates to roles, accelerating hiring timelines and improving candidate experiences. However, to realize these benefits fully, the industry must prioritize transparency, fairness, and ethical considerations, ensuring that AI becomes a trusted tool that creates a more equitable employment landscape.

Thank you for the great interview, readers who wish to learn more should visit Ribbon.

[disparition]

Whether you approach arts and media as a creator, a fan, a professional, or a hobbyist, you are probably very well aware of the rapid growth of AI in many areas of creative life and the conversation surrounding it. Whether we are talking about fiction or text of any kind, visual arts like painting and drawing, music composition and performance, sound design and audio editing - in all of these fields AI is something we now have to deal with, and there are many unanswered questions. How do we protect our livelihoods and practices as artists? What control can we have over our own work and its incorporation into machine learning tools and datasets? How do we as lovers of art and music determine if what we are reading, watching, or listening to is made by humans? Or how do we denote the particular degree of involvement of humans in a work of art?

It’s a complicated issue because it can be hard to say exactly where to draw the line. For example, well before the current conversation about AI, composers like Brian Eno were incorporating generative and algorithmic elements into their work. Does this mean such work is not “essentially human made”? Does it matter if the non-human components of a piece are created by randomness, or by natural elements gathered in e.g. a field recording, vs. being created by a computer program?

Those kinds of questions are more about the philosophical side of the issue. There are more pressing questions to do with how AI affects our ability to actually survive as artists. And, as the technology grows more powerful and the distinction between AI-generated or AI-enhanced material and human made material or basic documentation of reality itself, what methods do we use to signify that a given piece of media is or is not AI made, and what are the exact qualifications of that? Should the label “AI art” apply to work that is entirely generated by a computer, or should that label also be applied to any art that uses AI tools in any context - e.g. the new tools in Photoshop that make it easier to remove a specific object from an image and fill it in with background. What about the world of video games, where AI has been used for decades for things like pathfinding but is new and controversial when it comes to using it for story/dialog elements or certain visual assets?

We are, I think, still very much in the early days of all this, so it’s hard to come up with firm answers as the field changes so rapidly, but this is an essential conversation to be held now. I’ve been talking to a friend who put together a survey for an organization called Verified Human, who are looking specifically into the issue of how do we determine - going forward - whether a piece of art or media is “essentially human made”, and the question of how that should be communicated.

The idea of this survey is to gather as many points of view as possible for this conversation. I know a lot of the people that follow me here are artists and lovers of art, of all different kinds, and to me it is absolutely essential that creators of all kinds be involved in this conversation from early on. Please take a few minutes and fill out the survey via the link below, and if you are interested in helping out, please spread the word. And if you have any questions or would like to discuss, please feel free to contact me directly. Thank you.

https://forms.gle/BrSbGyq9wAyzwTa88

[jonesmartin-reviews]

Unlock the Power of AI: Give Life to Your Videos with Human-Like Voice-Overs

Video has emerged as one of the most effective mediums for audience engagement in the quickly changing field of content creation. Whether you are a business owner, marketer, or YouTuber, producing high-quality videos is crucial. However, what if you could improve your videos even more? Presenting AI voice-overs, the video production industry's future.

It's now simpler than ever to create convincing, human-like voiceovers thanks to developments in artificial intelligence. Your listeners will find it difficult to tell these AI-powered voices apart from authentic human voices since they sound so realistic. However, what is AI voice-over technology really, and why is it important for content creators? Let's get started!

AI Voice-Overs: What Is It? Artificial intelligence voice-overs are produced by machine learning models. In order to replicate the subtleties, tones, and inflections of human speech, these voices are made to seem remarkably natural. Applications for them are numerous and include audiobooks, podcasts, ads, and video narration.

It used to be necessary to hire professional voice actors to create voice-overs for videos, which may be costly and time-consuming. However, voice-overs may now be produced fast without sacrificing quality thanks to AI.

Why Should Your Videos Have AI Voice-Overs? Conserve time and money. Conventional voice acting can be expensive and time-consuming. The costs of scheduling recording sessions, hiring a voice actor, and editing the finished product can mount up rapidly. Conversely, AI voice-overs can be produced in a matter of minutes and at a far lower price.

Regularity and Adaptability You can create consistent audio for all of your videos, regardless of their length or style, by using AI voice-overs. Do you want to alter the tempo or tone? No worries, you may easily change the voice's qualities.

Boost Audience Involvement Your content can become more captivating with a realistic voice-over. Your movies will sound more polished and professional thanks to the more natural-sounding voices produced by AI. Your viewers may have a better overall experience and increase viewer retention as a result.

Support for Multiple Languages Multiple languages and accents can be supported with AI voice-overs, increasing the accessibility of your content for a worldwide audience. AI is capable of producing precise and fluid voice-overs in any language, including English, Spanish, French, and others.

Available at all times AI voice generators are constantly active! You are free to produce as many voiceovers as you require at any one time. This is ideal for expanding the production of content without requiring more human resources.

What Is the Process of AI Voice-Over Technology? Text-to-speech (TTS) algorithms are used in AI voice-over technology to interpret and translate written text into spoken words. Large datasets of human speech are used to train these systems, which then learn linguistic nuances and patterns to produce voices that are more lifelike.

The most sophisticated AI models may even modify the voice according to context, emotion, and tone, producing voice-overs that seem as though they were produced by a skilled human artist.

Where Can AI Voice-Overs Be Used? Videos on YouTube: Ideal for content producers who want to give their work a polished image without investing a lot of time on recording.

Explainers and Tutorials: AI voice-overs can narrate instructional films or tutorials, making your material interesting and easy to understand.

Marketing Videos: Use expert voice-overs for advertisements, product demonstrations, and promotional videos to enhance the marketing content for your brand.

Podcasts: Using AI voice technology, you can produce material that sounds like a podcast, providing your audience with a genuine, human-like experience.

E-learning: AI-generated voices can be included into e-learning modules to provide instructional materials a polished and reliable narration.

Selecting the Best AI Voice-Over Program Numerous AI voice-over tools are available, each with special features. Among the well-liked choices are:

ElevenLabs: renowned for its customizable features and AI voices that seem natural.

HeyGen: Provides highly human-sounding, customisable AI voices, ideal for content producers.

Google Cloud Text-to-Speech: A dependable choice for multilingual, high-quality voice synthesis.

Choose an AI voice-over tool that allows you to customize it, choose from a variety of voices, and change the tone and tempo.

AI Voice-Overs' Prospects in Content Production Voice-overs will only get better as AI technology advances. AI-generated voices could soon be indistinguishable from human voices, giving content producers even more options to improve their work without spending a lot of money on voice actors or spending a lot of time recording.

The future is bright for those who create content. AI voice-overs are a fascinating technology that can enhance the quality of your films, save money, and save time. Using AI voice-overs in your workflow is revolutionary, whether you're making marketing materials, YouTube videos, or online courses.

Are You Interested in AI Voice-Overs? Read my entire post on how AI voice-overs may transform your videos if you're prepared to step up your content production. To help you get started right away, I've also included suggestions for some of the top AI voice-over programs on the market right now.

[Go Here to Read the Complete Article]

[yongzs1218]

📍Hong Kong Palace Museum

Gallery 1: Entering the Forbidden City: Architecture, Collection, and Heritage

P1: Bell (bó zhōng) with stand--"Nán lǚ" pitch

Qing dynasty, Qianlong period, 1761

Bell: gilded copper alloy; stand: lacquer and gold on wood

P2: Chime (tè qìng) with stand--"Huáng zhōng" pitch

Qing dynasty, Qianlong period, 1761

Chime: jade (nephrite), gold; stand: lacquer and gold on wood

Gallery 2: From Dawn to Dusk: Life and Art in the Forbidden City

P3: Table screen with drawers for storing the Bight Pillars of the Orchid Pavilion

Qianlong period (1736～1795)

Zǐtán wood

Gallery 3: Brilliance: Ming Dynasty Ceramic Treasures from the Palace Museum, 1368～1644

P4: Architectural fitting in the shape of a dragon

Ming dynasty (1368～1644)

Earthenware with glazes

Gallery 4: The Hong Kong Jockey Club Series: Stories Untold - Figure Paintings of the Ming Dynasty from the Palace Museum

P5: Portrait of Táo Yuānmíng

Wáng Zhòngyù (active late 14th century)

Ming dynasty, late 14th century

Hanging scroll, ink on paper

Gallery 5: The Quest for Originality: Contemporary Design and Traditional Craft in Dialogue

P6:

1.(lower): Duck-shaped incense burner

Han dynasty (206 BCE～220 CE)

Bronze

2a.(left): Water dropper in the shape of a mythical beast

Qing dynasty (1644～1911)

Bronze with gold and silver inlay

2b.(right): Water dropper in the shape of a mythical beast

Ming dynasty (1368～1644)

Jade (nephrite)

3.(upper): Ram-shaped lamp

Han dynasty (206 BCE～220 CE)

Bronze

P7: Vase with spiral pattern

Imperial Workshops

Qing dynasty, Qianlong mark and period (1736～1795)

Blown glass

Gallery 6: Passion for Collecting: Founding Donations to the Hong Kong Palace Museum

P8: Headdress with phoenixes chasing a pearl

Liao dynasty (907～1125)

Gilt copper

Gift of Mengdiexuan Collection (Ms. Betty Lo Yan-yan and Mr. Kenneth Chu Wai-kee)

Gallery 7: The Hong Kong Jockey Club Series: Dwelling in Tranquillity - Reinventing Traditional Gardens

P9: Sky, Water and Land

Keith Lam

Generative visual and sound dataset, LED screen, stainless steel, bronze, four-channel audio installation

"In spring's serene embrace, tranquil water flows with grace; sky, water, and land above and below, boundless expanse azure aglow." -- Fan Zhongyan (989～1052), Yueyang Pavilion, Song dynasty

This artwork is inspired by the pavilions in Chinese gardens.

Gallery 8: Bank of China (Hong Kong) Presents: The Origins of Chinese Civilisation

P10: Relief with tigers and a human face

Shimao culture (4300～3800 BP)

Stone

📍香港故宮文化博物館

展廳1：紫禁萬象——建築、典藏與文化傳承

P1：甬紐橋口鎛鐘——南呂律（附架座）

清乾隆二十六年（1761年）

鎛鐘：銅��金；架座：木胎、金漆

P2：雲龍紋特磬——黃鐘律（附架座）

清乾隆二十六年（1761年）

特磬：碧玉描金；架座：木胎、金漆

展廳2：紫禁一日——清代宮廷生活與藝術

P3：蘭亭八柱八屜插屏

清乾隆（1736～1795年）

紫檀木

展廳3：流光彰色——故宮博物院藏明代陶瓷珍品

P4：琉璃龍吻

明（1368～1644年）

陶

展廳4：香港賽馬會呈獻系列：故事新說——故宮博物院藏明代人物畫名品

P5：陶淵明像

明，十四世紀晚期，王仲玉（活躍於14世紀晚期）

紙本墨筆立軸

展廳5：器惟求新——當代設計對話古代工藝

P6：

1.（下）鴨式薰爐

漢（公元前206～公元220年）

青銅

2a.（左）異獸形硯滴

清（1644～1911年）

銅胎錯金銀

2b.（右）瑞獸形硯滴

明（1368～1644年）

青玉

3.（上）羊式燭臺

漢（公元前206～公元220年）

青銅

P7：乾隆款纏絲玻璃撇口瓶

清乾隆（1736～1795年），清宮內務府造辦處

玻璃

展廳6：樂藏與共——香港故宮文化博物館首批受贈藏品展

P8：雙鳳戲珠紋冠

遼（907～1125年）

紅銅鎏金

夢蝶軒主人（盧茵茵女士和朱偉基先生）捐贈

展廳7：香港賽馬會呈獻系列：山林市城——遊歷舊園新景

P9：一碧萬頃

林欣傑，生成式視覺與音效數據組、LED屏幕、不鏽鋼、銅、四聲道聲音裝置。

「至若春和景明，波瀾不驚，上下天光，一碧萬頃」。

——宋•范仲淹（989～1052年）《岳陽樓記》

創作靈感源自中國園林建築中的「亭」。

展廳8：中國銀行（香港）呈獻：中華文明溯源（2024年9月25日～2025年2月7日）

P10：神人對虎紋石雕

石峁文化（距今4300～3800年）

石