The Hear, Speak and Orchestrate layers led the week, with 3 filings each.
Server-side STT endpointing becomes a first-class alternative to pipeline VAD for turn control.
pipecat-ai 0.0.99 adds enable_vad on GladiaSTTService, should_interrupt on Deepgram and Speechmatics STT, Speechmatics split_sentences, and use_ssl on NVIDIA STT/TTS.
pipecat-ai 0.0.99
Moves Deepgram's first_seen earlier than the April Flux filing and records the OfOne restaurant acquihire on the same primary — a platform raise, not a model ship.
Also Speak
Deepgram announced $130 million in Series C funding led by AVP at a $1.3 billion valuation, and the same release says it acquired restaurant voice startup OfOne.
Deepgram platform
Contact-centre Lex bots get a Neural ASR toggle for accented English on 13 January — still Lex-scoped, not a standalone Transcribe SKU.
Also Connect
Amazon Lex launched a neural automatic speech recognition model for English locales that improves recognition of conversational speech, non-native speakers, and regional accents inside Lex voice bots.
Amazon Lex neural ASR (English)
Keeps TTS separate from hear and orchestrate clusters.
livekit-agents 1.3.11 includes Gemini streaming TTS support on the speak path alongside provider websocket model query params.
livekit-agents 1.3.11
Word sync and pronunciation control matter for captioning and branded voice quality in cascaded bots.
pipecat-ai 0.0.99 adds word-level timestamps on AzureTTSService, pronunciation_dict_id for Cartesia TTS, AudioContextTTSService, and Inworld TTS keepalive.
pipecat-ai 0.0.99
Open MIT TTS that clones from ~5 s of audio and stays real-time on an Intel Ultra 7 / Apple M3 laptop — vendor benches versus F5-TTS and Kokoro, English-public-data training only.
Kyutai released Pocket TTS, a 100-million-parameter MIT-licensed text-to-speech model with few-second voice cloning that runs faster than real time on laptop CPUs, trained only on public English speech data.
Pocket TTS
Avatar bots can use LiveAvatar without a parallel custom video path.
pipecat-ai 0.0.99 adds HeyGen LiveAvatar API support on HeyGenTransport for realtime avatar video in Daily rooms.
pipecat-ai 0.0.99
Open full-duplex S2S finally takes text role prompts and voice clones together — FullDuplexBench wins are NVIDIA's own charts.
Also Speak
NVIDIA ADLR released PersonaPlex, a 7B Moshi-based full-duplex speech-to-speech model that takes a voice prompt plus a text role prompt so agents keep a chosen persona while listening and speaking concurrently.
PersonaPlex
Grok Voice joins the Pipecat realtime service set beside OpenAI and Gemini Live paths.
pipecat-ai 0.0.99 adds GrokRealtimeLLMService for xAI's Grok Voice Agent API with realtime audio streaming and function calling.
pipecat-ai 0.0.99
First 2026 Python Agents tag; MCP and EndCallTool are the orchestration headline.
livekit-agents 1.3.11 ships EndCallTool, standardises the Tool interface, and adds allowed_tools and transport_type on MCPServerHTTP.
livekit-agents 1.3.11
Modular ASR→LLM→TTS stacks get a dedicated timing layer — Tavus's 55ms p50 / zero-interruption bench is on a 28-sample internal set.
Also Think
Tavus released Sparrow-1, an audio-native conversational-flow model that predicts floor ownership so agents know when to listen, wait, or speak, instead of relying on silence-based endpoint detection.
Sparrow-1
Turn-taking becomes an ordered strategy list — the foundation later 2026 releases extend with incomplete-turn, eager EOT, and wake-phrase behaviour.
pipecat-ai 0.0.99 adds ordered user turn start/stop strategies, user mute strategies, UserTurnController, and UserTurnProcessor, replacing transport-level turn knobs.
pipecat-ai 0.0.99
Expands Pipecat's telephony connector surface beyond Daily, Twilio, and Telnyx serializers.
pipecat-ai 0.0.99 adds VonageFrameSerializer for the Vonage Video API Audio Connector WebSocket protocol.
pipecat-ai 0.0.99
Consumer voice assistant lands on iOS after web/Chrome — phone-call automation is still roadmap, not this cut.
Also Think
Speechify released its Voice AI Assistant on iOS, extending the Chrome and web assistant so users can ask questions, browse, run multi-turn voice dialogue, and trigger actions such as creating a podcast or opening a book by voice.
Speechify Voice AI Assistant (iOS)