The Speak, Think and Orchestrate layers led the week, with 3 filings each.
Self-hosted Flux STT and medical domain recognition extend regulated speech deployments.
pipecat-ai 0.0.108 adds DeepgramFluxSageMakerSTTService and AssemblyAISTTSettings.domain for modes such as medical-v1, plus on_end_of_turn on AssemblyAI STT.
pipecat-ai 0.0.108
Adds Indian-language-friendly Smallest Waves and xAI HTTP TTS to the vendor matrix.
pipecat-ai 0.0.108 adds SmallestTTSService for Waves Lightning and XAIHttpTTSService for xAI's HTTP TTS API, plus TTSService.on_turn_context_created.
pipecat-ai 0.0.108
Dated fal primary for a production TTS host ship — fills the prior pass skip that treated Inworld-on-fal as an undated third-party listing.
fal hosted Inworld TTS-1.5 Max with sub-~250ms P90 time-to-first-audio, 15-language support, and pricing around $0.01 per minute.
Inworld TTS-1.5 Max on fal · vendor Inworld
API TTS at $0.016/1k characters with a claimed 70ms model latency and 3s zero-shot voice adaptation — open-weight reference voices are CC BY-NC, so commercial use still needs the API.
Mistral AI released Voxtral TTS, its first text-to-speech model: a 4B-parameter multilingual system covering nine languages with claimed 70ms model latency, API pricing from $0.016 per 1k characters, and open weights for reference voices under CC BY-NC 4.0.
Voxtral TTS
Indian-market Sarvam chat models become selectable beside OpenAI-compatible endpoints.
pipecat-ai 0.0.108 adds SarvamLLMService (sarvam-30b/105b variants) and NovitaLLMService, plus developer-role messages across LLM adapters.
pipecat-ai 0.0.108
Realtime S2S with a claimed 90.8% ComplexFuncBench Audio score and SynthID on every clip — LiveKit is named among early adopters for agent stacks.
Also Use
Google released Gemini 3.1 Flash Live, its highest-quality realtime audio and voice model, in preview via the Gemini Live API in Google AI Studio, in Gemini Enterprise for Customer Experience, and under Gemini Live and Search Live for consumers.
Gemini 3.1 Flash Live
Responses API becomes a first-class Pipecat LLM path ahead of the 1.0 WebSocket default.
pipecat-ai 0.0.107 adds OpenAIResponsesLLMService for the OpenAI Responses API with streaming text, function calling, and usage metrics.
pipecat-ai 0.0.107
watsonx Orchestrate builders get ElevenLabs TTS/STT with PCI and Zero Retention options across 70 languages — a managed-voice path for IBM's agent platform rather than a new foundation model.
Also SpeakHear
IBM and ElevenLabs announced an integration that brings ElevenLabs text-to-speech and speech-to-text into IBM watsonx Orchestrate so builders can add multilingual voice to agentic workflows with enterprise security controls.
watsonx Orchestrate × ElevenLabs · vendor ElevenLabs
Parallel branches can emit in deterministic order for eval and multi-modal sync.
pipecat-ai 0.0.107 adds frame_order on SyncParallelPipeline so synchronized outputs can follow pipeline order.
pipecat-ai 0.0.107
Toolset and observability orchestration after 1.5.0 defaults. Class-2: folded quiet patch(es) 1.5.7 into this row.
livekit-agents 1.5.1 adds MCPToolset, LIVEKIT_OBSERVABILITY_URL for custom observability endpoints, and enables AGC by default on RoomInput audio.
livekit-agents 1.5.1
Avatar and image frames stay lip-syncable with TTS audio on the wire.
pipecat-ai 0.0.107 adds sync_with_audio on OutputImageRawFrame and audio_out_auto_silence on TransportParams.
pipecat-ai 0.0.107