EchoForge-ASR v0.1.1

EchoForge-ASR v0.1.1

This maintenance release hardens the operator-controlled real-model path while
preserving the deterministic Pages demo and evidence boundary.

Highlights

• Adds echoforge preflight with JSON output and stable exit codes (0 pass,
  1 prerequisite failure).
• Performs real runtime imports/API checks without loading weights, validates
  sherpa ONNX/token files, and requires a complete local faster-whisper
  CTranslate2 directory including tokenizer.json.
• Supports common official sherpa filenames such as
  encoder-epoch-*-avg-*.onnx with deterministic ambiguity rejection.
• Propagates --model-type through serve, factory construction, and the
  sherpa adapter.
• Separates service readiness, static preflight status, and verified streaming
  model load state in the API and VoiceLab UI.
• Guarantees that session.stopped is emitted only after registry cleanup;
  the race-sensitive lifecycle test passed 20 consecutive local runs and the
  Python 3.11/3.12 CI matrix.
• Updates package metadata to 0.1.1; full verification is 119 passed plus
  Ruff, mypy, compileall, Node syntax, Pages build, sdist/wheel build, and an
  isolated installed-wheel smoke test.

Evidence boundary

Static preflight does not load model weights and does not establish CER, WER,
RTF, CUDA compatibility, production latency, or any speech-quality claim.
Those values remain unpublished until a frozen model/data manifest and an
end-to-end benchmark report exist.

Artifacts

• echoforge_asr-0.1.1-py3-none-any.whl
  • SHA-256: c4de25d3a27679766a7dbb8575bbff8138e49fbf63572fcf26910359ae156cc5
• echoforge_asr-0.1.1.tar.gz
  • SHA-256: 36e790bf929aec679f8994f78a271fb75cd97ed2f089e85a6be1a713eabbc48e

EchoForge-ASR v0.1.0

EchoForge-ASR v0.1.0

Evidence-first Chinese streaming ASR laboratory with:

• EFA1 ordered PCM16LE WebSocket transport with bounded-session backpressure.
• Deterministic hysteresis VAD and monotonic self-healing transcript revisions.
• Optional sherpa-onnx streaming and faster-whisper endpoint adapters.
• Local Mic/File transport, deterministic Pages replay, timeline exports, and
  fail-closed evaluation tooling.

Verification for this release: 116 tests passed; Ruff, format, mypy,
compileall, JavaScript syntax checks, wheel smoke, and Pages deployment passed.

The Pages artifact is an interactive deterministic fixture. It does not publish
CER, WER, RTF, production SLA, accent coverage, or business-impact claims.
No model weights or evaluation recordings are included.