The Speak, Orchestrate and Use layers led the week, with 2 filings each.
Domain vocabulary steering reduces proper-noun errors on Scribe pipelines.
pipecat-ai 1.2.0 improves ElevenLabs STT services so Scribe callers can bias transcription with keyterms.
pipecat-ai 1.2.0
Latency versus prosody trade-offs become per-bot settings rather than provider defaults only.
pipecat-ai 1.2.0 adds max_buffer_delay_ms on CartesiaTTSService and mip_opt_out on Deepgram TTS services.
pipecat-ai 1.2.0
Screenplay-style direction plus default watermarking targets game/audiobook/agent delivery that flat TTS still misses — English-only by design, not a multilingual race.
Resemble released DramaBox, an English-only 3.3B diffusion-transformer TTS directed by plain-language screenplay prompts, with optional 10-second voice cloning, ~2.5 s generation on a warm H100, and Resemble Watermarker embedded on every output.
DramaBox
Adds a face layer to Boson's own STT/TTS stack for production voice agents — private preview only, eight sessions per H100 is the cost claim to watch.
Also Speak
Boson AI introduced Higgs Avatar, a real-time avatar foundation model that generates an expressive speaking face from a single still image, co-designed with Higgs TTS, streaming frames at ~16 ms on one H100 hosting up to eight concurrent conversations.
Higgs Avatar
Incomplete-turn handling moves from an LLM mixin into composable turn strategies. Class-2: folded quiet patch(es) 1.2.1 into this row.
pipecat-ai 1.2.0 adds FilterIncompleteUserTurnStrategies, DeferredUserTurnStopStrategy, ExternalUserTurnCompletionStopStrategy, and LLMTurnCompletionUserTurnStopStrategy.
pipecat-ai 1.2.0
Outbound call classification is a new orchestration capability.
livekit-agents 1.5.9 introduces Answering Machine Detection so outbound calls can classify person, voicemail, IVR, or unreachable numbers.
livekit-agents 1.5.9
Voice bots can observe and cancel UI jobs over the same RTVI channel as audio.
pipecat-ai 1.2.0 adds first-class RTVI ui-event, ui-snapshot, and ui-cancel-task client-to-server messages for UI-driven agents.
pipecat-ai 1.2.0
EU self-serve voice orchestration with compliance as the opener — lowers the bar for agencies and SMBs that refuse US-only stacks, while keeping an enterprise services lane.
Also OrchestrateConnect
Kolsetu opened Elba Self-Serve on 13 May for European builders: GDPR-oriented voice-agent platform with provider-flexible models, MCP server connections, no credit card to start, and an enterprise path for mission-critical deployments.
Elba Self-Serve
Enterprise CX platform ships voice as a native runtime layer rather than another bolted STT/TTS pair — 500+ language claim and sub-400 ms e2e are vendor figures to pressure-test.
Also ThinkSpeakConnect
Yellow.ai launched Nexus Vox inside its Nexus agentic platform: a single-runtime voice stack claiming sub-400 ms end-to-end latency, 500+ languages/dialects, 10-second voice cloning, real-time sentiment, and direct CRM/ticketing orchestration under one SLA.
Nexus Vox