VoiceShield v0.1.1 — Advisory Engineering Patch Release

VoiceShield v0.1.1

Release date: 2026-10-25

Supported scope

• Local development and demonstration on an Apple Silicon Mac.
• Authorized local microphone capture and WAV/FLAC file analysis.
• Advisory, uncalibrated acoustic anomaly scoring with explicit abstention and mixed-evidence reporting.
• Fail-closed prerequisite validation for an official ASVspoof 2019 LA evaluation using the separately acquired AASIST-L checkpoint and dataset.

Changes

• Offline file analysis now preserves meaningful raw threshold crossings and uncertain windows. A terminal LOW window no longer presents contradictory history as uniformly low.
• HIGH still requires sustained temporal-policy evidence.
• The ASVspoof reproduction runner validates official eval cardinality, protocol-to-audio coverage, unique IDs, deterministic two-class smoke allocation, canonical eval-only EER comparison, provenance, and complete report rendering.
• Documentation now distinguishes raw logits, uncalibrated softmax scores, the local diagnostic normalized minDCF, and official min t-DCF.

Validation evidence

• Official eval prerequisite validation passed with 71,237 protocol trials: 7,355 bona fide and 63,882 spoof.
• Coverage validation found 71,933 unique FLAC stems, zero missing protocol files, and 696 permitted extras.
• The complete local make check-all gate passed.
• Final CI run 40141988422 passed on main.
• Final Security run 40141988371 passed on main.
• Release commit: fb61fd2f8a74fdfc5ed6d4e73771443a4fb8352c.
• One 23.7-second ElevenLabs example produced 20 windows with four raw threshold crossings and no sustained-high decision. v0.1.1 reports that history as MIXED/UNCERTAIN. This verifies evidence preservation, not detector accuracy or modern-TTS reliability.
• The unchanged real-checkpoint M3 engineering benchmark is available in the canonical Markdown report and canonical JSON report.

Explicit limitations

• The full ASVspoof scoring attempt on an 8 GB M3 Mac was interrupted because of severe memory pressure. Its incomplete output is invalid and is not release evidence.
• No locally reproduced EER, official min t-DCF, calibrated probability, modern-generator accuracy, or broad scientific-validity claim is made.
• Thresholds remain unvalidated operating heuristics.
• VoiceShield does not authenticate callers, prove identity or fraud, automatically block calls, or intercept cellular calls.
• The AASIST source, checkpoint, and ASVspoof dataset have separate rights. The checkpoint and dataset are not redistributed by VoiceShield.
• This is not a production deployment or security certification.

VoiceShield v0.1.0 — Advisory Engineering Release

VoiceShield v0.1.0

VoiceShield v0.1.0 is a portfolio-grade advisory acoustic anomaly analysis engineering release for authorized local microphone streams and WAV/FLAC files.

VoiceShield does not authenticate speakers, prove identity, determine fraud, intercept cellular calls, or automatically block communications.

Supported scope

• Local development and demonstration on Apple Silicon Macs
• Authorized local microphone capture
• WAV and FLAC file analysis
• Real-time advisory acoustic anomaly scoring
• CPU and Apple MPS execution
• Official AASIST-L checkpoint loaded separately from its pinned upstream source
• Explicit insufficient-speech abstention and audio-quality diagnostics
• Volatile audio processing without intentional persistent audio storage

Validation evidence

• 494 Python tests passed
• 88.60% Python coverage
• 27 Vitest tests passed
• Mocked Playwright suite passed
• 3 live frontend/backend Playwright tests passed
• Linux CPU-only CI passed
• Security workflow passed
• Chrome and Safari manual smoke testing completed
• Official AASIST-L checkpoint integrity and provenance verified
• Real-checkpoint CPU and MPS benchmark completed on Apple M3

Verified GitHub Actions

• CI run 40128389057
• Security run 40128388932

Apple M3 real-checkpoint benchmark

The benchmark used the verified official AASIST-L checkpoint, 10 warm-up iterations, and 100 steady-state iterations per measured stage.

Device	Model inference p50	Model inference p95	End-to-end p50	End-to-end p95	End-to-end RTF
CPU	112.999 ms	129.814 ms	111.981 ms	117.614 ms	0.02800
MPS	30.187 ms	30.761 ms	33.754 ms	34.546 ms	0.00844

MPS end-to-end execution was 3.32× faster than CPU in this measured run.

A 600-second accelerated streaming stress soak completed with:

• 18,386 evaluated windows
• 0 reported dropped windows
• Maximum backlog of 1
• 0.00 MB measured RSS growth
• Bounded memory and backlog verdict: passed

Accelerated ingestion is not real-time-paced. These results are engineering latency and stability measurements, not detection-accuracy evidence.

Benchmark artifacts

• Markdown report
• JSON report

Artifact SHA-256:

• Markdown: 5fc5432b220e4c5b9ed2a980948a6fa8c7de32808c1a8d8e88e8873049531438
• JSON: 73b548187230662ef487b4829cf36b3e0e6f7ee494d9937956353a7fdbd74405

Official checkpoint SHA-256:

814331d088032bb4c3fa61cc014789eadeed464209dd094ab3a2dd6ffbdce27a

Important scientific limitations

• Local ASVspoof scientific reproduction has not been completed.
• Local EER, min-tDCF, ROC-AUC, balanced accuracy, FPR/FNR, and confidence intervals remain unmeasured.
• Scores are uncalibrated acoustic anomaly scores, not probabilities.
• Operating thresholds remain unvalidated heuristics.
• No accuracy claim is made for modern generators, Indic languages, replay attacks, codecs, telephony, or noisy channels.
• VoiceShield cannot intercept cellular calls.
• VoiceShield does not automatically block calls.
• VoiceShield is not identity proof or fraud proof.
• This release is not a production-deployment or security certification.

Checkpoint and licensing boundary

Original VoiceShield code is released under Apache-2.0. Vendored AASIST source retains its MIT terms.

The AASIST-L checkpoint and evaluation datasets have separate rights and are not redistributed with this release. Obtain the checkpoint directly from the pinned upstream source and verify its SHA-256 before use.

See the repository README, model card, evaluation card, benchmark evidence, and release notes for complete setup instructions and limitations.