Compute next randomized delay (seconds).

Common wake-up sequence: animation + restart subsystems + go interactive.

Return duration for a WAV byte string. Note: OpenAI's TTS WAV output currently uses placeholder chunk sizes (0xFFFFFFFF) for streaming.

Decode WAV and resample to 16kHz mono int16 PCM.

Thread-safe: push raw PCM int16 audio into the consumer queue.

Get a tool handler by name.

Sleep animation. Uses robot's built-in goto_sleep() method.

Event loop: listen for Gateway messages and dispatch invoke requests. Run as ``asyncio.create_task(client.run())``.

Create a sentence chunk (complete sentence).

Set the identified speaker for the next chat.send call.

Monitor for wake word while robot sleeps. Transitions to INTERACTIVE when wake word detected. If no wake word model availabl

Signal the manager to stop.

Synthesize text to PCM int16 audio bytes (24kHz mono).

Base64-URL encode without padding (RFC 4648 §5).

Check if a user face has been detected recently.

Continuously listen for speech during AWAKE state. Yields speech audio bytes via queue when pause detected. Runs throughout

Load the SNAC neural audio codec (24kHz).

Lazy-load the resemblyzer VoiceEncoder (downloads model on first use).

Sleep in small increments so stop() takes effect quickly.

Load the APM shim shared library from the same directory as this module.

Play audio data using pygame mixer (interruptible) or robot's media_manager. Args: audio_data: WAV audio data bytes

Convert captured samples to WAV bytes. Returns: WAV file as bytes (16kHz, mono, int16)

Send a node.invoke.result RPC back to the Gateway.

Split text into smaller pieces to improve perceived responsiveness and barge-in. We prefer to split on punctuation, otherwise on whitespace.

Evaluate breathing move at time t.

Feed speaker WAV (any rate/format) — resampled to 16kHz internally.

Generate a sine wave that simulates speech loudness for testing.

Load configuration from YAML file. Args: config_path: Path to config file Returns: Configuration dictionary Raises

Trigger a one-shot face orientation snapshot (for orient_on_speech mode). Thread-safe: can be called from any thread. The next control loop

Send a message and print all streaming events.

Enable/disable head tracking.

Create a text chunk.

Continuous antenna wiggle loop (runs in background). Uses set_target() with explicit head pose to prevent head wilting. Smoo

Apply the most recent speech/face offset updates.

Await a task completion in the background to prevent "Task exception was never retrieved". This is used for interruption paths where

Blend target antennas with listening freeze state and update blending.

Call the TTS trigger callback in a non-blocking way. Args: text: Text to send to TTS

Call TTS callback that returns a thread for tracking. Args: text: Text to send to TTS

Call TTS service via HTTP asynchronously. Args: text: Text to synthesize and play

Compose primary and secondary poses into a single command pose.

Decode Orpheus token IDs into PCM int16 audio bytes via SNAC.

Large-file-friendly download with resume support. Falls back to `_download` if `curl` isn't available.

Lazy-initialize the SCRFD model on first use.

`openWakeWord/openwakeword/utils.AudioFeatures` expects these by default: - resources/models/melspectrogram.onnx - resources/models/e

Convert raw LLM token IDs to SNAC codec indices.

Fetch audio from TTS service and add to playback queue. Args: text: Text to synthesize

Find Reachy's microphone device index. Returns: Device index for Reachy's microphone Raises: Runtim

Format prompt with voice and special tokens.

Get the primary full body pose from current move or neutral.

Get the secondary full body pose from speech and face tracking offsets.

Return the list of available actions.

Perform a gentle physical action on the robot.

Return the list of available cameras (single front-facing camera).

Capture a frame from CameraWorker and return as base64 JPEG.

Handle a single cross-thread command.

Check if the SoundDevice output buffer has drained.

Send the fused pose to the robot with throttled error logging.

Lazy-load the recorded emotions library.

Load or generate a persistent Ed25519 identity for this node. Returns dict with keys: device_id, public_key_b64url, private_key_pem, public_key_p

Load gateway token from env or ~/.openclaw/openclaw.json.

Load the shared library that exports WebRtcAecm_* symbols. Debian trixie package installs: /usr/lib/x86_64-linux-gnu/libwebrtc-audio-p

Normalize dB into [0,1] with gamma; clipped to [0,1].

Encode int16 PCM as WAV bytes.

Manage automatic breathing when idle.

Manage the primary move queue (sequential execution).

Emit frequency telemetry when enough loops have elapsed.

Background worker that plays audio from the queue sequentially. Ensures only one audio plays at a time. Singleton worker shar