feat: v2.3.0 — Layer 14 VoiceAnalyzer, [voice] extra, multilingual + learning loop hardening

VoiceAnalyzer (Layer 14) — pip install ethicore-engine-guardian[voice]:
  - Ultrasonic frequency scanning (17-96 kHz, FFT energy ratio)
  - Audio fingerprinting (80-dim MFCC cosine similarity)
  - Transcript cross-verification via faster-whisper (edit-distance divergence)
  - Silence/noise injection detection (frame-level RMS, burst counting)
  - Prosody anomaly scoring (F0 std, speech rate, energy CV)
  - 11-language prosody baselines (EN/ES/FR/DE/PT/RU/ZH/AR/JA/KO/HI)
  - Handles video files — audio extracted automatically via ffmpeg
  - Wired into adversarial learning loop and cross-modal fusion
  - Guardian.analyze_audio() public method added

Also in this release (multilingual + audit hardening):
  - All analyzer pattern groups multilingual (7 languages) across
    output_analyzer, indirect_injection_analyzer, context_tracker,
    response_anomaly_detector, automated_scan_detector
  - OutputAnalyzer: 12 signals (6 new vectors: indirect injection,
    agentic exfiltration, agent propagation, multimodal, credential/PII,
    encoded payload); voice categories in context boost factor
  - AdversarialLearner: category/severity/language now derived from
    actual signal result (was hardcoded)
  - community_guardian.py: PyPI wheel fallback for offline installs
  - threat_detector.py: v5.0.0, Full 14-Layer Defense header
  - transformers pinned <5 (CVE-2026-1839 fix is RC only)
  - 1106 tests pass

pyproject.toml:
  - [voice] extra: librosa, soundfile, pydub, faster-whisper, ffmpeg-python
  - [all] extra: voice deps included

Author: Oracles Technologies LLC

ethicore_guardian/__init__.py

@@ -7,7 +7,7 @@
 """
 
 # Version information
-__version__ = "2.2.0"
+__version__ = "2.3.0"
 __author__ = "Oracles Technologies LLC"
 
 # Core exports — full API-tier guardian preferred; community fallback for wheel installs

ethicore_guardian/versions.py

@@ -2,12 +2,12 @@
 Ethicore Engine™ - Guardian SDK - Version Information
 """
 
-__version__ = "2.2.0"
+__version__ = "2.3.0"
 __version_info__ = tuple(map(int, __version__.split('.')))
 
 # Build information
 __build__ = "stable.1"
-__release_date__ = "2026-05-03"
+__release_date__ = "2026-05-05"
 
 # Feature flags
 FEATURES = {

pyproject.toml

@@ -128,6 +128,22 @@ browser = [
     "lxml>=4.9.0",
 ]
 
+# voice: enables Layer 14 (VoiceAnalyzer — audio/voice threat detection).
+# Covers ultrasonic command injection, adversarial audio fingerprinting,
+# transcript cross-verification via Whisper, silence/noise injection, and
+# prosody anomaly scoring.  Also handles video files — audio is extracted
+# automatically via ffmpeg-python when a video path is supplied.
+# Install: pip install "ethicore-engine-guardian[voice]"
+# Note: faster-whisper requires the ctranslate2 binary (~200 MB on first run).
+#       Override Whisper model size with env var: GUARDIAN_WHISPER_MODEL=tiny
+voice = [
+    "librosa>=0.10.0",
+    "soundfile>=0.12.0",
+    "pydub>=0.25.0",
+    "faster-whisper>=1.0.0",
+    "ffmpeg-python>=0.2.0",  # shared with [video]; also needed for video audio extraction
+]
+
 # ml: enables full transformer-based ML inference in MLInferenceEngine.
 # Without these packages the engine runs its built-in heuristic fallback,
 # which is still effective but less accurate than a fine-tuned classifier.
@@ -156,6 +172,11 @@ all = [
     "opencv-python-headless>=4.8.0",
     "beautifulsoup4>=4.12.0",
     "lxml>=4.9.0",
+    # voice (Layer 14)
+    "librosa>=0.10.0",
+    "soundfile>=0.12.0",
+    "pydub>=0.25.0",
+    "faster-whisper>=1.0.0",
 ]
 
 [project.urls]