News
Meta gives, and Meta takes away. At the end of last week, Meta announced its new generative AI model for speech. With the announcement came the news that it would not be making the model public at this time.
Meta described the generative model, known as Voicebox AI, as a "breakthrough," stating that in the future it could give virtual assistants natural-sounding voices in six different languages.
However, on the same day it made the announcement, it also published a research paper that explained it would be irresponsible to release the model publicly.
“There are many exciting use cases for generative speech models, but because of the potential risks of misuse, we are not making the Voicebox model or code publicly available currently," a related blog reads. It adds that as with other powerful new AI innovations, the company recognizes “this technology brings the potential for misuse and unintended harm ....”
The company built safeguards into the model to mitigate these risks, in the form of a classifier that can distinguish between authentic speech and audio generated speech.
Meta outlined several use cases for the model, including the possibility for visually impaired people to hear messages from friends in their natural voices. It also opens up the possibility for creators to edit audio tracks for video or to create better virtual assistants, and while it didn’t specifically mention it in the post, the model has applications for the metaverse.
The fact that Meta is holding back on the release is notable as it indicates a potential acknowledgement about its potential public applications.
It’s unclear when the first practical results of this will enter the public domain, but at the current rate of development, it shouldn’t take long.
Oracle's Data Sovereignty Moves
A major challenge facing organizations is data sovereignty, or the geographical and regulatory area in which an organization's data is stored.
Given the different regulations surrounding data usage, for any company operating worldwide, the location of data matters. Oracle this week introduced two new EU cloud regions in part as a nod to these requirements. The new cloud regions allow companies to store data in the location of their choosing.
For the EU, the big issue is protecting European citizens' data from access requests by organizations outside of the EU.
The two new cloud regions are in Frankfurt, Germany and Madrid, which both now offer the same cloud services as Oracle’s public cloud regions. Both EU Sovereign Cloud regions are supported by EU-based personnel and operated by separate EU legal entities. As a result of the new service, Oracle now offers:
- Location: Organizations have a choice of where they store their data.
- Access restrictions: Makes organizations the sole arbiter of who can access the stored data.
- Operations and support teams: The option to have local staff operating the data centers.
- Regulatory requirements: Organizations can apply the compliance mechanism they are following.
The new offering is available to all 27 EU member states and to non-EU companies operating inside its borders.
YOOBIC Brings Generative AI to Frontline Teams
Employee experience platform provider YOOBIC is the latest vendor to pull generative AI into its portfolio with the launch this week of YOOBIC NEO, a suite of tools designed to help organizations engage, train and interact with frontline teams.
The tool provides HR and learning and development managers better insights into the training and development needs of frontline workers. The new AI features are natively integrated with existing workflows to support personalized employee experiences for every deskless worker, based on behavior, performance goals and learning patterns.
NEO will also offer a wide range of AI prompts and customized commands to support content creation. People can use the tool to define tone, format and length of posts. In sum, it offers a way of creating scalable content that is targeted at very specific audiences in the enterprise.
The release also included a new L&D Designer, which according to the company, will support L&D leaders in designing personalized curricula, tailored to job requirements and knowledge gaps and based on past performance and learning behaviors of the individual.
The release of Neo follows the release of YOOBIC ONE in January. The new platform unifies task management, communications and training all into one mobile platform. NEO feeds into this by adding generative AI to the mix.
Galileo Identifies Problem Data in LLMs
Generative AI has raised other potential problems and Galileo, which develops machine learning-driven data intelligence for unstructured data, aims to fix one of the biggest with the launch of its proprietary data quality intelligence platform in May. By incorporating just a few lines of code during the model training process, the platform can automatically identify problematic data that negatively impacts model performance and offer solutions for data science teams to rectify the issue.
This month the company launched a new suite of tools called Galileo LLM Studio, which offers organizations a way to test the accuracy of their Large Language Models (LLMs) to help make them as effective as possible.
According to a statement from the company, Studio debugs the data that is being used to build the models and reduces the number of hallucinations in the build (hallucinations are the confident, yet fully fabricated responses AI provides).
The new Studio is specifically designed for high-performance data science teams and provides a single place where these teams can access tools, notably Galileo Prompt Inspector and the Galileo LLM Debugger.
Conference
Sep
16
HR Tech Conference Las Vegas 2025
Register
Conference
Oct
27
Gartner HR Symposium/Xpo Orlando 2025
Register
The first of these enables users to identify model hallucinations while the Prompt Inspector allows users to fine-tune LLMs with their own proprietary data.
While there are obvious advantages to cleaning the data being used in the development of these models, what Studio aims to do is to automate this process to eradicate inaccurate predictions and identify errors earlier. Data scientists will also be able to compare the prompts they are using to identify which offers the optimal outcomes.
More than ever, Yash Sheth, Galileo co-founder and chief product officer, pointed out in the statement about the release, data has become key to achieving business goals. He also said that bringing LLMs into the workplace will depend on the quality of data that is used to build them.
"Adapting LLMs to specific real-world applications depends on data more than ever before. Today, an organization’s data is its only differentiator," he said in the statement. He also hopes to expand beyond natural language processing to other AI domains like computer vision.
Galileo is growing quickly. It raised $18 million in a Series A round late last year, bringing the total raised to date to $23.1 million.
Snowflake and Comet Partner for Model Building
Sticking with data and its use in model building, MLOps platform Comet has announced a partnership with data cloud management company Snowflake. The two will work together to deliver solutions to help data scientists build better models with greater speed.
The partnership will integrate Comet’s solutions with Snowflake’s single, unified platform so developers can track and version their Snowflake queries and datasets within their own Snowflake environment.
According to the company, enabling organizations to reproduce models requires the ability to build different versions of the code along with versioning of the code, hyper parameters and data. The easiest way to do this, the companies claim, is to keep data inside Snowflake, which prevents silos from developing as well as avoids the adjacent complexities and problems that result from such silos.
Comet is backed by $69.8 million in venture funding. Its platform helps software teams manage the neural networks, training datasets and other files used in AI development projects. For its part, Snowflake is a cloud computing–based data cloud company founded in July 2012. It launched publicly in October 2014 after two years in stealth mode.
Have a tip to share with our editorial team? Drop us a line: