Manages periodic spontaneous speech with randomized timing and presence gating. Usage:: manager = SpontaneousSpeechManager(

Persistent async WS client for OpenClaw Gateway in **node** role. Declares ``camera.snap``, ``camera.list``, ``action.perform``, and ``a

Simulates a WebSocket connection with a sequence of frames.

Thread-safe AEC using NLMS adaptive filter. The speaker reference is buffered and consumed frame-by-frame as the microphone signal is pr

Persistent async WS client for the OpenClaw Gateway. Usage:: client = OpenClawGatewayClient(gateway_url, token) await clien

Captures speech using VAD with circular buffer pattern from GLaDOS. Uses Silero VAD to detect voice activity, buffers pre-activation audio,

Simulates a WebSocket connection with a sequence of frames.

Thread-safe camera worker with frame buffering and face tracking.

Main orchestrator for Reachy MVP voice assistant. Manages the full lifecycle: startup → sleep → wake → listen → process → respond

Continuous thinking animation that escalates over time.

Local speech-to-text engine with noise detection. Supports ``faster_whisper`` and ``parakeet`` backends.

Face detection head tracker using SCRFD via insightface. Same interface as the YOLO HeadTracker from the conversation app: get_head_posit

Periodically queues ambient recorded emotions on the MovementManager. Usage:: manager = SpontaneousGestureManager( queue_mov

Sentence-level TTS with direct SoundDevice playback. Class-level (singleton) queue so multiple pipeline instances don't overlap audio.

Returns a synthetic test frame.

Coordinate sequential moves, additive offsets, and robot output at 100 Hz. Responsibilities: - Own a real-time loop that samples the current

Identifies speakers from short audio clips using enrolled embeddings. Args: enrollments_path: Path to JSON file mapping speaker names to

Voice Activity Detection using Silero VAD v5 ONNX model. Adapted from GLaDOS implementation.

Audio stream wrapper that applies AEC to every read.

Track rolling loop frequency statistics.

Middleware that streams LLM responses from the OpenClaw Gateway. Like ProviderMiddleware, this is a *source* middleware: it ignores the inpu

Result of processing audio input through STT. Contains transcribed text along with metadata about whether the input should be processed

Curious look around — head rotates slowly to one side, pauses, returns.

Slow head stretch — tilt to one side, then the other, return to center.

Yawn-like motion: slow head tilt back, antennas droop, then recover.

Wrap a PyAudio input stream and expose a 16kHz-style .read() API. The returned bytes are little-endian PCM int16 mono at `output_rate_hz`.

Thread-safe wrapper for WebRTC APM with full AEC3. APM operates on 10ms frames: - 8kHz: 80 samples - 16kHz: 160 samples - 32kHz

Filter chunks based on a predicate function. Useful for removing empty chunks, filtering specific patterns, or conditionally passing thr

Quick antenna wiggle — like perking up or reacting to something.

Gentle nod — as if agreeing or acknowledging something.

Gentle head shake — as if saying 'no' or musing.

Converts audio deltas into head movement offsets.

Synchronous Orpheus TTS engine using vLLM LLM class.

Manages Reachy Mini animations for different events. Wraps blocking robot animations in async executors to prevent blocking the main asy

Buffer text chunks until complete sentences are formed. Yields sentence chunks when sentence-ending punctuation is detected (. ! ?). Thi

Pipeline for processing streams through multiple middlewares.

Thread-safe wrapper for WebRTC AECM. AECM operates on 10ms frames: - 8kHz: 80 samples - 16kHz: 160 samples

Processes audio input through STT pipeline with noise detection. Responsibilities: - Transcribe audio using local STT engine - Detec

Breathing move with interpolation to neutral and then continuous breathing patterns.

Returns a synthetic test frame (colored gradient) for testing without robot.

State tracking for the movement system.

Result of speech-to-text transcription

Filter stderr output to suppress noisy async cleanup errors. Filters out "Exception ignored" messages from httpx/openai async generator clea

Feed audio chunks -> per-hop sway outputs. Usage: rt = SwayRollRT() rt.feed(pcm_int16_or_float, sr) -> List[dict]

Middleware that sends sentence chunks to TTS service and plays them. Only processes sentence chunks (complete sentences), not individual tex

Sentence-level TTS -> testbench playback. Class-level (singleton) queue so multiple pipeline instances don't overlap audio.

The APM instance wrapper — holds the APM pointer and the StreamConfig so we don't need to reconstruct them on every call.

TTS middleware that can be interrupted by external signals. Only processes sentence chunks (complete sentences), not individual text chunks.

States for Reachy's operation.

Represents a single chunk from the stream. Chunk types: - "text": Raw text content from the model (character/word level) - "sentence

Base class for stream middleware.

Test CameraWorker thread safety and interface.

Tests that expose confirmed bugs. Each test documents a real bug found during code review. The test SHOULD pass when the bug is fixed. Until

Tests that expose confirmed bugs in capture_speech. Each test documents a real bug found during code review. Marked xfail — they should pass