Muxiva built-in Nodes

Built-ins are vendor-neutral factories compiled into Muxiva. They are classified by capability even though they share the Rust runtime binary.

Node type	Layer	Capability	Contract
builtin.audio_resampler	Media	audio.resample	PCM S16LE Audio in and configured-rate Audio out
builtin.audio_vad	Algorithm	speech.vad	PCM Audio in and speech activity Event out
builtin.speech_formatter	Algorithm	text.speech_format	Streaming Markdown Text and cancellation Signal in; TTS-safe plain Text out
builtin.voice_turn_context	Control	conversation.turn_context	Transcript plus speech Event in and turn context Text out
builtin.voice_turn_controller	Control	conversation.turn_control	Filters meaningless transcripts, admits turns, and exclusively emits canonical cancellation
builtin.interval_tick	Control	clock.interval	Periodic Event out
builtin.text_source	Utility	text.source	Configured UTF-8 Text out
builtin.uppercase	Utility	text.uppercase	UTF-8 Text in and uppercase Text out
builtin.text_sink	Utility	text.collect	UTF-8 Text in
builtin.stdout_text_sink	Utility	observability.stdout	UTF-8 Text in and branded stdout logging

builtin.demo.* factories are deterministic architecture previews used by tests and are marked utility / demo.voice. They are not production microphone, ASR, LLM, TTS, or speaker Nodes.

Select any built-in in Studio to inspect its configuration and Port schemas. Media conversion is explicit: incompatible sample rates should be connected through builtin.audio_resampler, not silently converted inside an Edge. The generic Node declares sample_format, sample_rate_hz, and channels in explicit input and output configuration objects, so 16 kHz input and 48 kHz output are not separate Node types.

builtin.speech_formatter lets an Agent's original Markdown branch directly to a rich chat client while a derived plain-text branch feeds TTS. It strips emphasis markers and bare URLs, keeps link labels, and replaces streamed fenced-code blocks and Markdown tables with configurable spoken messages. It also owns sentence batching for TTS through minimum_chunk_characters and maximum_chunk_characters; the Agent itself emits semantic text deltas without voice policy. A terminal Agent Event flushes the final partial sentence. suppressed_parenthetical_terms configures product-specific parenthetical presentation that must not be spoken. The Builtin contains no character or device vocabulary; Xiaozhi terms live only in the Xiaozhi Graph. The streaming parser retains fence and table state across Text Frames; an interruption Signal or a new sequence resets that state, so stale unfinished Markdown cannot silence a later turn.

builtin.voice_turn_controller is the sole interruption decision point in a cascade voice graph. VAD speech.started/stopped values are observations. Only an admitted final transcript produces muxiva.turn.cancelled and a same-sequence Prompt. Fillers, coughs, minimum length, and short command allowlists are configurable. An authoritative device stop first emits muxiva.turn.interrupt.requested and still passes through the controller before reaching Agent, TTS, and playback Nodes.