Soniox STT v6 ships: voice agents that cancel their own TTS echo

Live now in 60+ languages with clean barge-in handling, while Anthropic open-sources proof-of-inference and OpenAI lifts its bug bounty to $100K.

Nowline NOV 4 2:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Your voice agent can finally ignore itself

    Soniox STT v6 (live Nov 3) adds agent voice echo cancellation: it strips your own TTS playback out of the mic feed while still catching the user who talks over it, and it lets you stream the agent's outgoing text in as live context. Real-time and async, 60+ languages, on the API today. The two ugliest bugs in a voice loop, transcribing your own bot and handling barge-in, now get solved upstream instead of in your glue code.

  • Prove the model you called is the model you paid for

    Anthropic's Frontier Red Team open-sourced FlashProve (Apache 2.0), a zkSNARK that lets a provider commit to exact weights and prove every response came from them, checkable in ~17s on a CPU without ever seeing the weights. It ships a prover, a fast C++ verifier, a Lean-verified verifier, and a protocol spec. It's a prototype (~110 tokens/sec generation on one H100, ~18.8MB proofs per 1,024 tokens, 70-95x the GPU cost of just running the model), but it's the first runnable answer to 'did they quietly swap in a cheaper model?'

  • OpenAI 5x's its bug bounty to $100K

    OpenAI raised its top payout for critical vulnerabilities to $100,000, a five-fold jump, and hired an outside red-team firm to attack its own systems using its own models. If you do security research, high-impact OpenAI findings just got far more lucrative.