Clawd BotandClaude Opus 4.6
ca9b510922
- Reset master to upstream/main (16,697 commits) - Overlay 2,271 local-only files (skills, tools, workspace, configs, apps) - Restore IDENTITY.md and USER.md templates - Build verified, gateway running, Discord working Co-Authored-By: Claude Opus 4.6 <[email protected]>
116 lines
4.2 KiB
Python
116 lines
4.2 KiB
Python
#!/usr/bin/env python3
|
|
"""
|
|
openWakeWord sidecar for voice-surface server.
|
|
|
|
Protocol (stdio):
|
|
stdin: raw signed 16-bit PCM at 16kHz, little-endian, mono
|
|
fed in 1280-sample chunks (80ms) by the Node.js parent
|
|
stdout: line-based JSON {"detected": true, "score": 0.92} or {"detected": false, "score": 0.1}
|
|
|
|
Wake word priority:
|
|
1. Custom sklearn model at models/james_classifier.pkl (trained with train_james.py)
|
|
2. Fallback: hey_jarvis_v0.1.onnx bundled with openWakeWord
|
|
"""
|
|
import os, sys, json, pathlib, warnings
|
|
warnings.filterwarnings("ignore")
|
|
os.environ["ORT_LOGGING_LEVEL"] = "3"
|
|
|
|
import numpy as np
|
|
import openwakeword
|
|
from openwakeword.model import Model
|
|
from openwakeword.utils import AudioFeatures
|
|
|
|
SAMPLE_RATE = 16000
|
|
CHUNK_SAMPLES = 1280 # 80 ms per chunk
|
|
CHUNK_BYTES = CHUNK_SAMPLES * 2 # int16 = 2 bytes
|
|
|
|
THRESHOLD = float(os.environ.get("OWW_THRESHOLD", "0.5"))
|
|
MODELS_DIR = pathlib.Path(__file__).parent / "models"
|
|
JAMES_CLF = MODELS_DIR / "james_classifier.pkl"
|
|
|
|
_oww_base = pathlib.Path(openwakeword.__file__).parent
|
|
JARVIS_MODEL = str(_oww_base / "resources" / "models" / "hey_jarvis_v0.1.onnx")
|
|
|
|
# ---- Load models ----
|
|
use_custom = JAMES_CLF.exists()
|
|
|
|
if use_custom:
|
|
import joblib
|
|
clf = joblib.load(JAMES_CLF)
|
|
af = AudioFeatures()
|
|
sys.stdout.write(json.dumps({"ready": True, "model": "james-custom"}) + "\n")
|
|
else:
|
|
oww = Model(wakeword_model_paths=[JARVIS_MODEL])
|
|
model_key = list(oww.models.keys())[0]
|
|
sys.stdout.write(json.dumps({"ready": True, "model": "hey_jarvis_fallback"}) + "\n")
|
|
|
|
sys.stdout.flush()
|
|
|
|
# Accumulator: build up enough audio to run embed_clips (needs >1280 samples)
|
|
# For the custom model we collect 5 chunks (400ms) before scoring to get stable embeddings
|
|
# Compute the required detection window size so the embedding model has enough frames.
|
|
# openWakeWord's embedding model expects at least ~76 frames. Each frame ≈160 samples.
|
|
MIN_EMB_FRAMES = 76
|
|
MIN_EMB_SAMPLES = (MIN_EMB_FRAMES + 3) * 160 # mirrors openwakeword utils calculation
|
|
CUSTOM_WINDOW = int(np.ceil(MIN_EMB_SAMPLES / CHUNK_SAMPLES))
|
|
# Ensure a sensible lower bound (keep previous behaviour if model changes)
|
|
CUSTOM_WINDOW = max(5, CUSTOM_WINDOW)
|
|
custom_buf = [] # list of int16 arrays
|
|
|
|
buf = b""
|
|
|
|
try:
|
|
stdin_raw = sys.stdin.buffer
|
|
while True:
|
|
needed = CHUNK_BYTES - len(buf)
|
|
chunk = stdin_raw.read(needed)
|
|
if not chunk:
|
|
break
|
|
buf += chunk
|
|
if len(buf) < CHUNK_BYTES:
|
|
continue
|
|
|
|
audio_int16 = np.frombuffer(buf[:CHUNK_BYTES], dtype=np.int16)
|
|
buf = buf[CHUNK_BYTES:]
|
|
|
|
if use_custom:
|
|
custom_buf.append(audio_int16)
|
|
if len(custom_buf) < CUSTOM_WINDOW:
|
|
# Not enough audio yet — emit low score placeholder
|
|
sys.stdout.write(json.dumps({"detected": False, "score": 0.0}) + "\n")
|
|
sys.stdout.flush()
|
|
continue
|
|
|
|
# Got CUSTOM_WINDOW chunks — extract embedding and classify
|
|
audio_window = np.concatenate(custom_buf)
|
|
custom_buf = custom_buf[1:] # slide window by 1 chunk
|
|
|
|
# embed_clips expects a numpy array with shape (N, samples).
|
|
# passing a Python list caused an AttributeError in openwakeword.utils (list has no .shape).
|
|
embs = af.embed_clips(np.stack([audio_window])) # shape (N, 96)
|
|
if embs is None or len(embs) == 0:
|
|
sys.stdout.write(json.dumps({"detected": False, "score": 0.0}) + "\n")
|
|
sys.stdout.flush()
|
|
continue
|
|
|
|
feat = embs.mean(axis=0).reshape(1, -1)
|
|
score = float(clf.predict_proba(feat)[0, 1])
|
|
detected = score >= THRESHOLD
|
|
|
|
if detected:
|
|
custom_buf = [] # reset window after detection
|
|
|
|
else:
|
|
# hey_jarvis ONNX model — feed 1 chunk at a time
|
|
scores = oww.predict(audio_int16)
|
|
score = float(scores.get(model_key, 0.0))
|
|
detected = score >= THRESHOLD
|
|
if detected:
|
|
oww.reset()
|
|
|
|
sys.stdout.write(json.dumps({"detected": detected, "score": round(score, 3)}) + "\n")
|
|
sys.stdout.flush()
|
|
|
|
except (BrokenPipeError, KeyboardInterrupt):
|
|
pass
|