Google AI Developers’ Post

In an hours transcript, every extra 0.5% WER means 450 more errors. That's 450 errors that you need to take the time to resolve. That's 450 errors that could make you non-compliant. That's 450 errors that will cost you money to fix.   Artificial Analysis is an independent third party conducting benchmarks on AI systems that the market can finally trust 🤝   Their analysis independently verifies that Speechmatics has the lowest WER in the market. Their analysis shows our competitors with higher Word Error Rates, and thousands more mistakes per hour. Their analysis shows the thousands of errors you won't have to worry about with Speechmatics.   When close enough isn't good enough, choose Speechmatics.   Check out the full breakdown 👇 🔗 https://lnkd.in/gw8jXPNQ

We have updated the Mixtral 8x22B guide to SynapseAI 1.16. #IamIntel

Built a gesture-based volume control system using OpenCV, employing real-time hand tracking and distance-based algorithms to map thumb-index finger span to system volume. A seamless, hands-free audio control solution driven by computer vision

Augmenting communication with real-time captioning systems could be useful for helping people distinguish speech in noisy environments. Today we illustrate how a joint sound separation + ASR model can benefit from training with hybrid datasets that have large amounts of simulated audio complemented by small amounts of real recordings. Learn more at https://goo.gle/3UjvH9Y

This time last year AIT published an in-depth report on #HumanComputerInteraction and its evolution from the mainframes of 1950s to today’s world of Graphical User Interfaces and interactive displays used by billions. #AITOralHistory Watch our trailer here: https://lnkd.in/enJqv8B7

Can you hear who's knocking at your door? 👋 HANCE does realtime AI audio enhancement, and can be integrated in both software and hardware. Where better than intercoms? Intercoms come in all shapes and sizes, and some might be in need of better audio quality. Improving upon hardware might not be the most convenient answer. If you have suggestions of other use-cases where clear audio is needed, feel free reach out! 🤝

Let’s explore Ophir’s article and find out how, and when, laser measurements must absolutely, positively be made. We digested the article and put together an audio 🎧 summary for you to listen and get a valuable insight fast: https://bit.ly/3MicLDT

LLMs can handle tokens beyond text, such as image patches, audio segments, actions, or molecules. If a problem can be framed as modeling a sequence of discrete tokens, an LLM can be applied.

Our new paper is finally public! We introduce a new audio codec architecture which we're calling the Transformer-Audio-AutoEncoder (TAAE). If you listened to the MUSHRA test I posted a few weeks back, you already heard this in action... Here's the highlights: 🤖 Fully transformer-based, encoding 16kHz speech to a low frame-rate discrete latent representation. 👂 SoTA sound quality at just 0.4 or 0.7 kB/s, measured with objective and subjective tests. 👾 Single-token FSQ-based bottleneck, with new post-hoc decomposition that allows optional separation into residual tokens after training. 🦕 Experiments showing that reconstruction quality scales with parameter count up to 1B parameters. 🐐Open-source weights coming very soon. Huge props to all co-authors at Stability, but especially to Xubo Liu, Anton Smirnov who've both been in deep with me on this topic since the summer. Demos and arXiv link in the comments 👇👇👇👇👇

#RP2350 | #RP2040 Learn how to add audio input, output, and machine learning capabilities to your Raspberry Pi Pico 2! Works just as well on original Pi Pico, but the #machinelearning inference is significantly faster on RP2350. FULL VIDEO -> https://lnkd.in/es_r2fAj