The Speak layer led the week with 4 of 12 filings, from LiveKit, Inworld, Daily / Pipecat and CAMB.AI.
Material STT knobs before the 1.4.0 line. Class-2: folded quiet patch(es) 1.3.11 into this row.
livekit-agents 1.3.12 adds AssemblyAI language detection and a Google STT profanity filter.
livekit-agents 1.3.12
CPU-first ASR for Hinglish-style mid-sentence mixes targets Indian enterprise voice where English-centric STT still fails — OpenASR 3.10% WER is Shunya's claim.
Shunya Labs released Zero Codeswitch, a speech recognition foundation model trained for Hinglish and other Indian mixed-language utterances that emits mixed-language tokens natively without an English translation layer.
Zero Codeswitch
New TTS engines as one speak cluster.
livekit-agents 1.3.12 integrates AsyncAI TTS and adds lexicon_uri on Azure TTS.
livekit-agents 1.3.12
Sub-130ms Mini P90 and AA leaderboard #1 give agent builders a priced TTS rung before Inworld's later realtime stack — 75% cheaper claim is Inworld's.
Inworld released TTS-1.5 Mini and Max for realtime voice apps, claiming P90 time-to-first-audio under 130ms (Mini) and 250ms (Max), 30% more expressiveness, a 40% lower word error rate, and 15-language coverage including Hindi.
Inworld TTS-1.5
Adds another streaming-quality TTS vendor option beside Cartesia, ElevenLabs, and Inworld.
pipecat-ai 0.0.100 adds CambTTSService for Camb.ai MARS models (mars-flash, mars-pro, mars-instruct).
pipecat-ai 0.0.100
Splits TTS into Flash/Pro/Instruct/Nano so call-centre agents and film dubbing stop sharing one API SKU — Pipecat lands CambTTSService the same day.
CAMB.AI released MARS8 as four TTS architectures — Flash for sub-150ms agents, Pro for media dubbing, Instruct for director-level prosody control, and Nano (50M) for on-device — with compute-first pricing on customer infrastructure across 25+ platforms.
MARS8
Open-weight native S2S with claimed ~135ms TTFT gives dual-use agent builders a cascade-free path — SIM 0.817 is FlashLabs' own bench.
Also Speak
FlashLabs released Chroma 1.0, an open-source ~4B end-to-end speech-to-speech model with few-second voice cloning, claiming end-to-end TTFT under 150ms (~135ms with Day-0 SGLang) without an ASR→LLM→TTS cascade.
Chroma 1.0
LLM-side knob distinct from hear/speak.
livekit-agents 1.3.12 exposes a reasoning parameter on the OpenAI Responses LLM integration.
livekit-agents 1.3.12
Dates the infra platform that later 2026 LiveKit Agents ships sit on — a $1B Series C, not a product drop.
LiveKit announced a $100 million Series C led by Index Ventures, joined by Salesforce Ventures, Hanabi Capital, Altimeter and Redpoint Ventures, at a $1 billion valuation.
LiveKit Agents
Idle prompts and re-engagement hooks become a shared controller instead of ad-hoc timers in each bot.
pipecat-ai 0.0.100 adds UserIdleController on LLMUserAggregator and UserTurnProcessor, plus mute start/stop events on the user aggregator.
pipecat-ai 0.0.100
Removes a common fork point when bots sit behind authenticated WebSocket proxies.
pipecat-ai 0.0.100 adds additional_headers on WebsocketClientParams for authenticated or proxied WebSocket connects.
pipecat-ai 0.0.100
Puts a vertical voice capture loop on farm ops (speak → structured field records) rather than another form-heavy FMS — early public SKU before the Apr iOS expansion.
Numanac publicly released its voice-first farm management app so growers can speak field observations (or upload files) and have them structured into searchable operational records.
Numanac (public app)