Audio & Speech

Today's Stories

1. 1

   Sean Parker rebuilds Stability AI around music with $76 million

   Sean Parker and Prem Akkaraju are rebuilding Stability AI as a music toolmaker; in late August it announced $76 million from Sony, Warner, and Universal, which also licensed their catalogs for training. The music labels' backing and catalog licensing indicate a shift from litigation to collaboration for AI music tools, suggesting Stability may gain an edge with professional users.

   Parker says an upcoming update will let users hum a melody or beatbox a drum pattern to steer the AI, so the test is whether that feature ships and resonates with music professionals.

2. 2

   Microsoft, LG put Voice Live AI into ThinQ ON hub

   Microsoft and LG Electronics unveiled a Speech-to-Speech AI agent at the Microsoft Industry Summit in Seoul, embedding Microsoft's Voice Live system into LG's ThinQ ON smart home hub. The agent uses direct speech comprehension instead of a transcribe-then-synthesize pipeline, letting users interrupt mid-response and switch intent instantly, so commands feel less like talking to a database.

   LG plans a firmware update for existing ThinQ ON users within 2026, so the launch hinges on delivery timing, with the hub at approximately $172 in Korea.

3. 3

   Suno adds Speech for spoken audio with music

   Suno, the AI music generator, launched a feature called Speech that turns typed text into spoken words with matching background music in one track; product chief Jack Brody says it was tested with a small group for a month. Suno is positioning Speech for poems, meditations, and bedtime stories, though the beta still has bugs, and the company has not said how it trained the model.

   The launch comes as AI music generators face copyright criticism — major record labels have sued Suno and a Munich court recently ruled against it, rejecting fair use — so the feature's reception may hinge on how those legal fights unfold.

4. 4

   AI covers: RVC (MIT, still updated) leads as rivals archiving

   An explainer describes AI song covers as original melody, lyrics and backing kept intact, with only the vocals swapped, in three steps: vocal separation, voice conversion, then mixing. It lists RVC as MIT-licensed and actively updated, Seed-VC (GPL-3.0) archived in November 2025 and so-vits-svc (AGPL-3.0) archived in November 2023. For anyone actually building this workflow, tool licensing, not sound quality, is becoming the deciding factor, since two of the three listed voice-conversion tools are no longer maintained.

   Vocal separation accuracy is what most shapes the final result, with harmonies, choruses, heavy reverb, and noisy recordings the hardest cases. Also watch the rights side: publishing or monetizing a cover can require permission from both the song's and the voice's rights holders.

5. 5

   Language discrimination lifts multilingual speech AI to 10.4% error

   In a controlled English/French HuBERT setting, two interventions — an auxiliary language classifier and per-language k-means targets — cut phone-ABX error from 11.6% to 10.4%, below the monolingual 10.8%. Multilingual speech models have generally trailed monolingual ones on each language, so this suggests the extra cost of multilingual learning is not unavoidable, and cross-language sharing is largely preserved.

   Most of the gain came when language discrimination was introduced in the first training iteration; later or repeated interventions helped less and increased language-wise segregation, so timing appears to be the deciding factor.

6. 6

   Microsoft's voice AI tops accuracy rankings with 100ms latency

   Microsoft AI unveiled MAI-Transcribe-2-Streaming. It ranks first on Artificial Analysis, supports 60 languages, and returns interim results in just over 100 milliseconds. With interim results in under 100 milliseconds, voice agents can respond before the other party finishes speaking. Automated phone calls and chat agents could feel like live conversation.

   Whether the low latency and voice cloning hold up in real-world use. Both voice models can replicate a voice from a few seconds of audio and have safety measures built in. In one test, about half of 4,000 participants thought a voice belonged to a real person.