MikkoParkkola/nab

Patch release for the speakerId JSON deserialization fix found during the post-v0.7.0 live smoke test of the nab + hebb composition.

The bug

FluidAudio v0.13+ emits speakerId as a JSON string ("1") instead of an integer (1). This broke nab analyze --diarize on fresh FluidAudio installations.

The fix

Custom serde Visitor that accepts i64, u64, or str/String. Both forms deserialize to the same i32 speaker_id. Regression tests added.

Affected

All v0.7.0 users running FluidAudio v0.13+ (the current version installed by nab models fetch fluidaudio). Upgrade to v0.7.1.

Verification

nab analyze /tmp/test_30s.wav --audio-only --diarize --format json
# → 30s audio, 0.24s wall, 123x realtime, 2 speakers identified

Full smoke test of the nab + hebb composition passed after this fix.

Highlights

nab analyze v2 — multilingual SOTA ASR via FluidAudio (Parakeet TDT v3 on Apple Neural Engine). Live-tested at 131× realtime on a 2h 09m English interview with 97.18% confidence. Optional speaker diarization (PyAnnote community-1), word-level timestamps, and multilingual support across 25 EU languages + opt-in Qwen3-ASR for zh/ja/ko/vi.

URL watch — nab is now RSS for the entire web. Any URL can be a subscribable MCP resource. Background poller with conditional GETs, semantic diff, adaptive backoff. Clients receive `notifications/resources/updated` on change.

Active reading via MCP sampling — during transcription, nab asks the host LLM to identify references and recursively looks them up via `nab fetch`. Inlines citations as footnotes. Novel pattern — no other ASR pipeline does this.

Full MCP 2025-11-25 spec closure — Streamable HTTP transport, structured logging, argument completion, sampling helper, roots helper, `list_changed` notifications, elicitation form + URL mode. Top 1% of MCP servers in the wild.

Apple Vision OCR — 15 languages, ANE-accelerated, zero model download. Integrated into nab fetch (auto-OCRs images with thin/missing alt text, inlines `[Image: ]` annotations in the markdown output).

Cross-platform backends — sherpa-onnx + whisper-rs behind feature flags for Linux/Windows/Intel macOS.

nab + hebb composition — `nab analyze --diarize --include-embeddings` auto-matches speakers against hebb's voiceprint database. `nab fetch` auto-saves URLs to hebb's `kv:urls` namespace. Full sovereign multimodal stack.

nab models fetch fluidaudio — persistent install of the FluidAudio Swift CLI. Survives reboots (no more `/private/tmp` fragility).

Stats

• 13 commits in this release
• ~20,000 lines of Rust added
• 1187 tests passing, 0 errors, 0 deprecation warnings
• 0 commercial API dependencies in the default composition path

Full changelog

See CHANGELOG.md for the complete list.

Installation

```bash
cargo install nab --version 0.7.0
nab models fetch fluidaudio # one-time: clone + build FluidAudio Swift CLI
nab analyze my-video.mp4 --diarize
```

See docs/getting-started.md for full installation options.

Companion repo

Works standalone, but unlocks more when paired with hebb v0.4.0 (voiceprint DB + MCP-as-database) for named speakers and personal web memory.

Changed

• All 8 MCP tools now advertise outputSchema + return structured_content
• Extract truncation constants, consolidate schema helpers
• Benchmark tool reports error count per URL
• README: remove unlisted Windows binary, update test count to 907

Fixed

• Dockerfile: Rust 1.93 + proper build deps for Glama
• CHANGELOG: correct rust-mcp-sdk version reference

Fixed

• Dockerfile: Rust 1.93 (was 1.87), add build-essential/cmake/clang for crypto deps
• Dockerfile: --no-default-features to skip BoringSSL/QUIC for faster Docker builds

Changed

• MCP tools submit, login, validate now return structured_content alongside text
• MCP benchmark tool reports error count per URL instead of silently dropping failures
• CHANGELOG entries for 0.6.1–0.6.4