Load local Orukeet-compatible ASR bundles without repository fallback

[Nathan-Roll1]

Why is this change needed?

AsrModels.load(from:) resolves the selected NVIDIA repository even when the caller supplies a directory. This adds loadLocal(from:) and --local-model-dir so compatible Orukeet weights load from the exact compiled bundle, with missing components failing locally. The existing default download path and dependencies are preserved.

The Orukeet example consumes the pinned Hugging Face JSON manifest and verifies the archive's SHA-256 and byte count before compilation. These are normal Hugging Face model downloads; the local loader makes no network requests.

Initial validation at 0e4dda3 on an M5 Max: the library and CLI compiled, and the CLI transcribed a real recording through the new option. English, German and French recordings each produced identical text across three fresh decoder states (nine speech decodes); one-second silence returned empty text on three runs. The missing-directory check failed at the requested path without repository fallback. The changed Swift files pass formatting and whitespace checks.

The complete default package build, including NeMo text processing, also passed at 0e4dda3 on Xcode 26.6 CI before the vocabulary follow-up. All test sources compiled; the focused AsrModelsLocalTests|AsrModelsTests run had 32 passes, one opt-in real-model test skipped, and no failures. The real-weight transcription checks above ran locally.

The vocabulary follow-up at 0cbec7d accepts the official v2 vocabulary's extra entries while still requiring every token below the blank ID. Its full default package build and focused tests passed on Xcode: 34 passed, one optional real-model test skipped, zero failures. The added regressions use the published 1,031-entry v2 vocabulary and a temporary copy with one required token removed. Formatting and whitespace checks pass. The earlier real-weight checks above were not rerun for this vocabulary-validation change.

[Copilot]

🟢 Approval recommended

No unresolved issues were identified that would block approval.

Pull request overview

Adds explicit local loading for compiled Orukeet-compatible ASR bundles without repository resolution or downloads.

Changes:

• Added AsrModels.loadLocal(from:) with local validation.
• Added CLI --local-model-dir support.
• Added Orukeet documentation and opt-in regression tests.

File summaries

File	Description
Tests/FluidAudioTests/ASR/Parakeet/SlidingWindow/TDT/AsrModelsLocalTests.swift	Updated as part of this pull request.
Sources/FluidAudioCLI/Commands/ASR/Parakeet/SlidingWindow/TranscribeCommand.swift	Updated as part of this pull request.
Sources/FluidAudio/ASR/Parakeet/SlidingWindow/TDT/AsrModels.swift	Updated as part of this pull request.
README.md	Updated as part of this pull request.
Documentation/Orukeet.md	Updated as part of this pull request.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Alex-Wengg]

[P2] The vocabulary count check rejects supported v2 bundles —AsrModels.swift:235–238. The official v2 vocabulary contains 1,031 entries, including every required ID below its blank ID of 1,024. Requiring vocabulary.count == version.blankId causes --local-model-dir … --model- version v2 to fail before loading any models. I reproduced this with the compiled CLI and the published vocabulary. Check required token coverage while permitting extra entries, and add regression coverage.