The Speak layer led the week with 4 of 14 filings, from LiveKit, Microsoft, Daily / Pipecat and Alibaba / Qwen.
Speechify streaming TTS addition.
livekit-agents 1.6.7 adds Speechify streaming TTS with word-level timestamps and Fish Audio loudness/generation-tuning options.
livekit-agents 1.6.7
In-house TTS at $15/1M characters for Dynamics Contact Center and Azure Voice Live. Vendor also claims up to 89% GPU-cost reduction in Contact Center.
Microsoft put MAI-Voice-2-Flash in public preview: 2× faster and 32% cheaper than MAI-Voice-2, at $15 per 1M characters, for Dynamics 365 Contact Center and Azure Voice Live.
MAI-Voice-2-Flash
Builders can pair Flux STT and Flux TTS inside one Pipecat cascade.
pipecat-ai 1.6.0 adds DeepgramFluxTTSService for Deepgram's Flux TTS early access over WebSocket.
pipecat-ai 1.6.0
Q3 TTS comparisons that stop at Cartesia, ElevenLabs, and Inworld miss a multilingual, promptable option already on Alibaba Cloud.
Alibaba Cloud shipped Qwen-Audio-3.0-TTS in Flash (about 300ms first packet) and Plus variants on Model Studio, covering 16 languages with natural-language style control and inline non-verbal tags.
Qwen-Audio-3.0-TTS
Avatar gateway provisioning lands with Spatius.
livekit-agents 1.6.7 adds a Spatius avatar plugin and inference.AvatarSession for avatar provisioning via the gateway.
livekit-agents 1.6.7
Silent tool outcomes and Responses reasoning controls land for agentic voice flows.
pipecat-ai 1.6.0 adds reasoning support on OpenAIResponsesLLMService and NO_RESPONSE returns from Pipecat Flows consolidated functions.
pipecat-ai 1.6.0
Realtime interruption parity with pipeline agents. Class-2: folded quiet patch(es) 1.6.9 into this row.
livekit-agents 1.6.7 supports adaptive interruption for realtime models.
livekit-agents 1.6.7
Turns Vapi's model-agnostic catalog into guided defaults, so builders pick a goal (latency, cost, intelligence) instead of hand-picking every STT, LLM, and TTS pair.
Vapi shipped Model Intelligence: curated Model Presets for Balanced, High Intelligence, Ultra Fast, and Cost Saver stacks, plus dashboard cost, latency, and quality metrics drawn from live Vapi traffic.
Vapi Model Intelligence
Negative assertions catch bots that speak when they should stay silent.
pipecat-ai 1.6.0 lets eval scenario expectations set absent: true so a turn passes only when no matching event arrives within the window.
pipecat-ai 1.6.0
Telephony noise isolation mode.
livekit-agents 1.6.7 adds Krisp voice isolation telephony mode on the noise-cancellation plugin.
livekit-agents 1.6.7
MoQ is a new transport class beside WebRTC and WebSockets for fan-out media.
pipecat-ai 1.6.0 adds MOQTransport for bidirectional low-latency Media over QUIC between bots and clients.
pipecat-ai 1.6.0
Turns restaurant voice AI from a standalone answering SKU into kitchen-connected fulfilment via an existing digital-ordering backbone.
SoundHound and Deliverect connected Smart Ordering voice agents to Deliverect menus and POS routing so drive-thru, phone, kiosk, and in-car orders reach the kitchen without re-keying.
Smart Ordering × Deliverect
Opus and Sonnet on voice without barge-in. Date from @claudeai coverage; tweet URL not independently verified.
Anthropic expanded the Claude voice mode beta to Opus 5 and Sonnet 5 (plus Haiku 4.5) on mobile, desktop, and web. It remains turn-based, not full-duplex.
Claude Voice Mode
Moves OpenAI from API voice models into a managed contact-centre-style SKU, competing with Sierra and other agent platforms on governed production deployments rather than only on gpt-realtime and GPT-Live.
Also ThinkOrchestrate
OpenAI released Presence, a managed enterprise product for deploying governed AI agents on voice and chat that answer questions, take approved actions in company systems, and escalate to people under customer-set policies.
OpenAI Presence