Audio & Speech

Today's Stories

1. 1

   YuE2 music AI rivals Suno as Sony's Hakken mines papers

   A weekly roundup covers five new generative AI technologies, including YuE2, a music generator compared with Suno, and Sony's Hakken, an AI that predicts unknown scientific relationships from vast paper collections. Music generation and scientific discovery are both moving from experimental demos toward tools with named, competing products, so buyers in those fields are likely to get more choices.

   The roundup names the tools but the body gives no benchmark numbers, prices or release dates, so the test is whether YuE2 and Hakken publish results that can be compared directly with existing products.

2. 2

   Tencent's Gander interrupts users in just 8 percent of cases, beats GPT-Realtime

   Tencent's Hunyuan Speech team and university researchers introduced Gander, which splits real-time conversation from background tasks between a "cerebellum" and a swappable "brain." On Full-Duplex-Bench v3, Gander interrupted users in 8 percent of cases versus 13.5 percent for GPT-Realtime. Gander's lower interruption rate suggests splitting conversation from reasoning can make voice assistants feel less disruptive than systems like GPT-Realtime, though its task accuracy fell slightly below that of the weakest competitor.

   The researchers call the work early, and how to scale Gander up remains an open question. Tencent plans to publish the weights and training data once it completes the open source release process.

3. 3

   McDonald's revives Archy AI at drive-thrus after 2024 shutdown

   McDonald's unveiled ArchIQ and its Archy voice assistant in limited testing at a handful of US drive-thrus, after shutting down a similar system in 2024 that had run for three years. The first system handled drive-thru orders at about a hundred US locations and failed because it didn't flag its own mistakes, not because it was inaccurate overall.

   The QSR Drive-Thru Report found AI order accuracy was 81%, rising to 95% when the system transferred to an employee, so success hinges on when Archy pauses and hands off.

4. 4

   Treble raises $18 million for voice simulation platform

   Treble raised $18 million in a Series A extension led by Paladin Capital Group, with KOMPAS VC, Frumtak Ventures, EIC and Omega ehf participating. It raised $12 million in 2024, taking its total to over $40 million. Treble's simulation platform gives voice AI labs, wearables and robotics makers a way to generate synthetic audio data and test models, an alternative to internet-scraped recordings that Paladin sees as increasingly valuable infrastructure.

   Treble aims to expand into physical AI, including robotics, automotive and drones, so the test is whether it can win those customers beyond Amazon and Logitech. Watch how its Hugging Face speech-recognition benchmark is used.

5. 5

   Spotify set to change course on AI music labeling

   Spotify and industry groups are moving to label, restrict or ban AI-generated tracks, after the New York Times reported about a dozen AI-made songs hit various Billboard charts since 2022. The leading music streaming platform had been sitting on the fence, saying it was working on a shared industry standard for disclosing AI use, but nothing had been implemented so far.

   Whether Spotify actually rolls out the labeling and blocking tools hinges on the shared industry standard moving from talk to implementation, and on whether other platforms follow suit.

6. 6

   Google launches Gemini 3.8 Live for real-time voice

   Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15 and began offering them the same day, rolling them out in Search Live, Gemini Live, Google Workspace, the Gemini API and Gemini Enterprise. The two models let voice assistants keep talking while tasks run in the background, which could make AI voice agents more usable for hands-on or customer-facing work.

   Extended Thinking's #1 ranking on Artificial Analysis's Speech to Speech Quality Index uses that index's own scoring; Workspace access requires a paid plan, and API usage pricing is not stated in the official blog.