v0.5.0 — Browser personal-model infrastructure

v0.5.0 — Browser personal-model infrastructure

This release tags the local-first browser Personal Voice Model infrastructure milestone. It is an infrastructure, portability, hardening, and release-gate milestone; it does not publish production Personal Voice Model accuracy or reference-hardware performance claims.

Included

• Personal Voice Model schemas and contracts: SpeechModelManifestV3, BrowserTrainingContractV1, SpeechProfileManifestV2, BrowserTrainingJobV1, TrainingFeatureShardV1, and PortableSpeechModelManifestV1.
• Browser-training infrastructure behind @speech/browser-training: fixed adapter-math backend, residual bottleneck/LHUC reference math, balanced sampling, AdamW, validation/early stopping, deterministic checkpoints/resume, Web Locks coordination, ASR-priority pause, and accessible progress/recovery UI.
• Local preparation/evaluation flow: training-readiness reports, frozen enrollment/vocabulary revisions, prompt-identity splits, FP16 feature shards, CTC frame labels, selected-vocabulary metadata, personal/anchor evaluation, activation decisions, rollback, and aggregate release gates.
• Personal Models UI: profile cards, capability preflight, readiness tasks, activation comparison/override/rollback, multi-profile dedupe/import-as-new/replace/rename/delete, and worker-owned lifecycle operations.
• Portable .speechmodel lifecycle: deterministic inner bundles, default Web Crypto encrypted export, hostile importer validation, exact base-model compatibility, worker-owned smoke vectors, atomic staging, CLI residual-adapter compatibility, and V1-to-V2 speech-profile migration/recovery.
• Hardening/release plumbing: cross-browser/fault-injection smoke coverage, parser/import-security corpus, synthetic fixture licensing, privacy/security/licensing review, bilingual cohort gate contract, and reference-hardware benchmark gate contract.

Release limitations

• ADR 0004: no user-approved 30-speaker bilingual quality cohort evidence is available in this repository. Do not claim production Personal Voice Model accuracy or quality gates pass.
• ADR 0005: no declared reference-hardware Personal Voice Model benchmark evidence is available in this repository. Do not claim production memory, storage, latency, RTF, export/import, offline, or zero-network performance gates pass.
• ADR 0006/0007: release notes must preserve local-first privacy boundaries. Raw audio, transcript text, feature tensors, checkpoints, adapter weights, profile JSON, private vocabulary, prompt/case/profile IDs, storage paths, and network payloads are not release artifacts.
• Synthetic fixtures, CI smoke tests, local diagnostics, and contract tests are regression evidence only; they are not substitutes for cohort or reference-hardware evidence.
• Production ASR/model weights remain out of Git. The current external VietASR catalog entry remains metadata-only and is not advertised as a low-latency bilingual streaming model.

Validation

Required CI passed on the tagged commit 8e72dd120e41e69cc52458804fa8b8804e74b9bc before release publication:

• lint, typecheck, test, build
• chromium smoke

Local release asset generation ran from tag v0.5.0:

• source archive generation from tag v0.5.0
• pnpm --filter @speech/web build
• web build archive generation from apps/web/dist
• SHA-256 checksum generation for attached archives

Deployment verification

Public production alias verified over HTTPS:

• https://speech-amber-beta.vercel.app/ loads the app shell.
• Required COOP/COEP/CSP/Permissions-Policy headers are present.
• manifest.webmanifest and sw.js are reachable.
• Service-worker-backed offline reload renders the app shell after first load.

The Vercel unique deployment URL for the merge commit is SSO-protected; public verification used the production alias above.

Assets

• speech-v0.5.0-source.tar.gz — source archive generated from tag v0.5.0
• speech-v0.5.0-web-dist.tar.gz — generated web build archive
• SHA256SUMS.txt — SHA-256 checksums for both archives