EchoForge-ASR v0.1.2 hardens the real-model evaluation path while preserving a fail-closed publication boundary.
Highlights
- Added resumable AISHELL-1 acquisition with persistent receipts, legacy-manifest migration, safe nested-tar extraction, atomic staging, and extraction inventories.
- Bound the complete marker -> prepare -> runner -> evaluator evidence chain, including split, package, device, timing, decoder, and final-stage provenance.
- Rejects duplicate JSON keys, non-finite numbers, malformed UTF-8/surrogates, truncated PCM WAV files, unsafe model paths/symlinks, and model artifacts that change during inference.
- Separates static preflight readiness from verified streaming-model load readiness.
- Adds a Windows verifier contract pinned to ctranslate2==4.5.0 and CI coverage for Python 3.10, 3.11, and 3.12.
- Keeps evaluator output sanitized: no per-utterance transcript text or audio paths are published.
Validation
- 264 passed, 4 skipped locally on Python 3.10 and 3.11.
- Linux CI passed on Python 3.10/3.11/3.12; the Windows verifier contract also passed.
- The wheel was installed and imported in a fresh virtual environment.
Publication boundary
No CER, WER, RTF, endpoint-latency, or production-SLA claim is published in this release. The current real-model compatibility smoke is explicitly evaluation_authorized=false and frozen=false because the local sherpa model package does not include a separately confirmable weight license. Authorized metrics require a complete single AISHELL dev/test split, CPU execution, an independent warm-up, frozen evidence, and confirmed model-weight licensing.
SHA-256
- echoforge_asr-0.1.2-py3-none-any.whl: 49b2868d3fd9c49d3eb5ab6e91eb393a0f221dc66f1cc188412484470cc0ebcf
- echoforge_asr-0.1.2.tar.gz: 8bc6566d4f31e63316f1c9629a3d6d49ed1b9ebe7d393f1838e36093a8731d8f