k2-fsa Open-Sources OmniVoice, a Zero-Shot TTS Supporting 600 Languages
k2-fsa (a group centered on members of the Kaldi development community) has released OmniVoice, a zero-shot text-to-speech (TTS) model supporting more than 600 languages, on Hugging Face. It adopts a Diffusion Language Model-style architecture, combining high-quality speech generation with fast inference.
Its standout feature is a pair of modes: Voice Cloning, which clones a voice from a short reference clip, and Voice Design, which lets you craft a voice using attributes such as gender, age, pitch, and dialect. It also supports non-verbal tokens like `[laughter]` for laughter, as well as pronunciation correction via pinyin and phonemes. Inference speed reaches an RTF of 0.025 (40x faster than real time), so practical bottlenecks are expected to be minimal.
The project claims the broadest language coverage of any zero-shot TTS to date, enabling speech synthesis across a wide range of languages, including minority languages. The paper is available on arXiv.
40% of Managers Enter Confidential Information into "Shadow AI" — GRAS Group Survey
A survey by the GRAS Group has revealed the growing reality of "shadow AI" — AI tools used without companies' knowledge or approval. The most striking finding: roughly 40% of managers enter confidential information into unapproved AI tools even while recognizing the danger.
According to ITmedia's coverage, while generative AI adoption in the workplace keeps expanding, corporate governance has failed to keep pace. Managers on the front lines are believed to face a genuine dilemma — feeling they have no choice if they want to work efficiently — which makes this a problem that simple prohibition won't solve.
The shadow AI problem is a global concern, not one limited to Japan, and there is now an urgent need for companies to establish AI usage policies and to provide properly approved AI tools.
Most Wall Street AI Trading Bots Are Underperforming
As AI adoption advances across the financial industry, reports indicate that many of the AI trading bots actually running on Wall Street are struggling to perform. According to an article in FA-Mag, many funds that deployed AI-driven automated trading systems have failed to deliver the expected results.
A fundamental limitation has been pointed out: because AI market predictions depend on pattern recognition over historical data, they have difficulty handling discontinuous market shocks and irrational human behavior. That said, reports also note that AI is becoming established as a useful tool for financial market analysis and risk management, strengthening the view that it is better suited to "analysis and decision support" than to "trading itself".
Armorer Arrives as a Security Foundation for AI Agents
Armorer, posted to Hacker News as a Show HN, is an open-source project that provides a secure local control plane for AI agents. It achieves process isolation via Docker and comes with agent lifecycle management, job tracking, and a unified UI/CLI.
The developer says they built it to solve two problems: the "dependency hell" of setting up AI agents such as Codex and Claude Code, and the security risk of giving powerful agents broad access to the host machine.
Particularly interesting is that when you point an existing coding agent at the Armorer repository, the agent itself can autonomously install and configure the entire stack. As AI agents continue to multiply, this highlights how security and management infrastructure for local environments is becoming a critical challenge.
---