IBM released ALTK-Evolve, which derives prompt-level playbooks from agent training trajectories without weight updates or human annotation. IBM reports held-out tau2-bench gains for GPT-4.1 and GPT-5.5, including GPT-5.5 airline accuracy rising from 72.5% to 82.5% and Pass^4 from 55% to 70%.
Why it matters: The method offers a comparatively cheap way to make repeated workflows more reliable by turning past trajectories into structured guidance, but it shifts more context and hidden policy into the prompt layer.
Salesforce AI Research described Launa-Omni, a model that maps speech and language into a shared latent space and uses an in-house text-to-speech component for two-way conversation. It was trained for Salesforce and Slack experiences, and Salesforce published evaluation results rather than a public API or model release.
Why it matters: The architecture targets the hard part of enterprise voice: combining low-latency conversation with instruction following and task accuracy, but availability remains a separate question from reported evaluation quality.
xAI made grok-voice-3.0 available for Speech to Speech at the same price as 2.0, repointed grok-voice-latest to the new model without changing the API, and deprecated grok-voice-2.0.
Why it matters: Applications pinned to 2.0 now carry migration risk, while applications using the latest alias changed models automatically. Voice teams need to test output quality, latency, safety and turn handling before the deprecated slug becomes an incident.
AWS made Nova 2.5 Sonic generally available as a speech-to-speech model for real-time voice agents. It retains mixed voice and text sessions, asynchronous tools, seven-language voices and a 256K context window while AWS reports better reasoning, instruction following, tool accuracy and latency than Nova 2 Sonic.
Why it matters: A single streaming model can keep conversation active while tools run, reducing the handoffs in customer-service and assistant systems. The consequential boundary is action authority: more reliable tool calls still need transaction confirmation, interruption…
QwenCloud released Qwen-Audio-3.1-Realtime-Plus for duplex speech conversations with eight additional system voices, a 262,144-token context window, function calling, web search and voice cloning through its realtime integration protocol.
Why it matters: Voice agents can now combine long sessions, external tools and cloned voices in one hosted path, increasing usefulness and the consequences of mistaken identity, tool execution or consent handling.
OpenAI released GPT-Live-1 in the API as a full-duplex voice layer that can listen and speak at the same time, handle interruptions, adapt tone and pace, operate through noise and silence, and delegate reasoning or tool calls to a backend model. The voice layer is listed at $0.05 per minute before backend model and tool charges.
Why it matters: Voice-agent performance now depends on two coordinated systems: the live conversational layer and the backend agent that reasons or acts. Failures in speaker detection, interruption handling or confirmation can therefore trigger wrong or mistimed actions…
AWS made Agentic CX Designer generally available in nine regions. Its no-code canvas combines agentic dialogue with deterministic flow steps for eligibility, approvals, routing, integrations and compliance, with design, testing and deployment in Amazon Connect.
Why it matters: Contact-centre teams can keep flexible conversation and high-consequence business rules in one inspectable workflow instead of delegating the whole interaction to an unconstrained agent.
Google made two dedicated Gemini speech-to-text endpoints generally available: a file model with language detection across more than 85 languages, diarization, word-level timestamps, and vocabulary biasing, plus a bidirectional WebSocket model for streaming transcription with interim and final events.
Why it matters: The release gives voice and media teams a stable Gemini API route for batch and live transcription without using a general conversational model. The two endpoints expose materially different latency, duration, diarization, timestamp, and streaming boundaries.
Hume released a benchmark covering more than 40 voice systems, 15 evaluation dimensions, over 60 metrics and more than one million human ratings. It tests recognition, generation and response to emotion, identity, style and acoustic context.
Why it matters: Word error rate and text-only dialogue scores miss whether a system hears and responds to the nonverbal information that makes spoken interaction useful.
Google introduced continuous speech-to-speech translation across more than 70 languages in Gemini Live, with public API and AI Studio preview access plus narrower previews in Meet and Google Translate.
Why it matters: Real-time translation is moving from segmented transcription pipelines toward a single conversational audio system that preserves timing and interaction.
AWS released Nova Sonic in Amazon Bedrock with a bidirectional streaming API, English speech input and output, function calling, retrieval grounding, content moderation and watermarking. The initial service was available only in US East (N. Virginia).
Why it matters: A single speech-to-speech model can preserve timing, tone and interruptions better than a serial speech-to-text, language-model and text-to-speech pipeline.
OpenAI released gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-mini-tts to developers worldwide. The transcription models targeted lower word-error rates than Whisper, while the text-to-speech model accepted style instructions such as tone and delivery.
Why it matters: Voice agents could combine more robust transcription and controllable speech without assembling separate third-party speech models.