v0.14.0

What's Changed

• fix: honor CLI flags on the default path and repair MCP/CI surfaces by @ekhodzitsky in #77
• refactor!: consolidate the public API surface and remove parallel worlds (0.13.0) by @ekhodzitsky in #78
• ci: gate standalone lockfiles and align companion-crate hygiene by @ekhodzitsky in #79
• docs: sync documentation with the 0.13.0 API and scrub tracker refs by @ekhodzitsky in #80
• perf: give embedder pool sessions a fair share of CPU cores (~3.6× on CPU) by @ekhodzitsky in #81
• perf: parallelize v2 segmentation + embedding (~3× RTF), release 0.13.1 by @ekhodzitsky in #82
• perf: widen powerset window hop to 2.0s (~2.5× faster segmentation) by @ekhodzitsky in #83
• perf: recalibrated INT8 models as opt-in fast profile (~2.8x e2e) by @ekhodzitsky in #84
• chore(release): 0.14.0 by @ekhodzitsky in #85

Full Changelog: v0.12.0...v0.14.0

INT8 model artifacts v2 (recalibrated)

Recalibrated INT8 models: powerset_int8 (weights-only dynamic QInt8) and resnet34_int8 (static QDQ calibrated on 1819 real fbank+CMVN VoxConverse-dev windows). DER at/below FP32 baseline on AMI; embedder stage ~7x faster. Supersedes the int8 assets attached to v0.6.0-alpha.2, which were calibrated incorrectly (raw audio fed as fbank).

v0.12.0

What's Changed

• fix(ci): dual CLI help snapshots and docker-smoke PLDA by @ekhodzitsky in #70
• feat: ort-free library path as first-class CI-gated surface by @ekhodzitsky in #71
• feat: BYO Embedder on Pipeline and StreamingPipeline by @ekhodzitsky in #72
• feat: resolve VBx PLDA through the model registry by @ekhodzitsky in #73
• refactor: align production surfaces and reduce dual-path debt by @ekhodzitsky in #74
• feat: library-mode STT consumer surface (BYO, labeling, errors) by @ekhodzitsky in #75
• feat: library mode VBx BYO path and surface contract by @ekhodzitsky in #76

Full Changelog: v0.11.0...v0.12.0

v0.11.0

What's Changed

• feat(asr): CJK/multilingual companion crate polyvoice-asr-sherpa by @ekhodzitsky in #47
• feat(release): packaging phase 2 — Docker image, Linux ARM64 binary, install docs by @ekhodzitsky in #28
• ci: fix asr-sherpa by forcing sherpa-rs-sys build-script rerun by @ekhodzitsky in #48
• chore: drop crossbeam-queue and fastrand; document ort EP migration by @ekhodzitsky in #49
• refactor(onnx): decouple neural stages from ort via InferenceRuntime by @ekhodzitsky in #51
• feat(cluster): reduce speaker-count errors with cVBx-style clustering by @ekhodzitsky in #50
• docs: refresh production readiness for 0.10 and define 1.0 GO gate by @ekhodzitsky in #53
• feat: ship exclusive diarization, segment confidence, and WDER by @ekhodzitsky in #54
• feat(attribution): speed up speaker join and harden word coverage by @ekhodzitsky in #56
• feat: make model adapters selectable and self-describing by @ekhodzitsky in #57
• feat: plug optional Sortformer streaming diarizer for four speakers by @ekhodzitsky in #58
• feat: accept any common audio format and sample rate on the CLI by @ekhodzitsky in #60
• feat: offer an optional pure-Rust tract path for ONNX inference by @ekhodzitsky in #61
• feat(streaming): stabilize online labels with cache and latency presets by @ekhodzitsky in #62
• feat: offer optional earshot VAD without changing the default by @ekhodzitsky in #63
• feat: add optional ERes2NetV2 and CAM++-zh speaker embedders by @ekhodzitsky in #64
• docs: publish measured streaming, VAD, and embedder parity numbers by @ekhodzitsky in #65
• Promote pipeline v2 + VBx to default after full DER gate by @ekhodzitsky in #66
• fix: release-check under all-features and clippy 1.96 by @ekhodzitsky in #67
• fix: VBx PLDA fixtures for release-check default path by @ekhodzitsky in #68
• style: rustfmt CLI DER PLDA fixture args by @ekhodzitsky in #69

Full Changelog: v0.10.0...v0.11.0

v0.10.0

What's Changed

• ci: pin stable Xcode on macOS jobs (fix clang_rt.osx link flake) by @ekhodzitsky in #27
• feat(bindings): who-said-what format parity for Python and C FFI by @ekhodzitsky in #29
• test(der): gate the headline no-collar DER and activate the long-form floor by @ekhodzitsky in #30
• feat(models): require manifest signatures for profile-resolved models in release builds by @ekhodzitsky in #31
• feat(onnx): single EP-aware session builder for embedding and segmentation by @ekhodzitsky in #32
• feat(pipeline_v2): thread execution_provider from config to sessions; add CLI selector by @ekhodzitsky in #33
• feat(bench): per-backend RTFx with per-stage runtime breakdown by @ekhodzitsky in #34
• feat(vad): route the Silero VAD session through the EP-aware builder; CoreML full-coverage smoke by @ekhodzitsky in #35
• feat(onnx): wire the XNNPACK execution provider by @ekhodzitsky in #36
• feat(segmentation): calibrated binarization of posteriors (hysteresis + min-duration) by @ekhodzitsky in #37
• docs(benchmarks): add pyannote community-1 and speakrs; provenance comparison artifact by @ekhodzitsky in #38
• ci: make the miri lane blocking for its covered targets by @ekhodzitsky in #39
• test(bench): lock the DER aggregation arithmetic in a pure, unit-tested helper by @ekhodzitsky in #40
• test(streaming): integration test for the cumulative feed()/flush() contract by @ekhodzitsky in #41
• security(docs): ort native-binary provenance documented; pinned binary cached in CI by @ekhodzitsky in #42
• docs(streaming): differentiation thesis, cannot-link and latency-mode specs by @ekhodzitsky in #43
• fix(spectral): correct BIC perfect-fit under-selection in the legacy clusterer by @ekhodzitsky in #44
• chore(release): 0.10.0 by @ekhodzitsky in #45
• ci: fix workflow YAML broken by the ort-cache insertion (duplicate keys) by @ekhodzitsky in #46

Full Changelog: v0.9.0...v0.10.0

v0.9.0

What's Changed

• feat(pipeline_v2): segmentation-derived overlap reconstruction by @ekhodzitsky in #23
• feat(vbx): first-class VBx clusterer for pipeline v2 (accuracy/meeting path) by @ekhodzitsky in #24
• feat(pipeline_v2): optional dense embedding (sub-window per segment) by @ekhodzitsky in #25
• chore(release): 0.9.0 by @ekhodzitsky in #26

Full Changelog: v0.8.0...v0.9.0

v0.8.0

What's Changed

• docs: streamline README, move benchmarks out, drop agent-scaffolding docs by @ekhodzitsky in #20
• release: 0.8.0 — who-said-what, MCP, schema, VBx, v2 segmentation fix by @ekhodzitsky in #21
• fix: pipeline v2 default min_cluster_size 12→1 (unblocks 0.8.0 release gate) by @ekhodzitsky in #22

Full Changelog: v0.7.0...v0.8.0

v0.7.0

What's Changed

• release: v0.7.0 — DER correctness, packaging, CI hardening by @ekhodzitsky in #19

Full Changelog: v0.6.9...v0.7.0

v0.6.9

Fixed

• segmentation/aggregator: cumulative speaker permutation was double-applied across window boundaries in stitch. Added a 3-window regression test.
• pipeline: run_from_wav now validates WAV sample rate and returns PipelineError::UnsupportedSampleRate on mismatch.
• embedder: EmbedderPool::new returns Result and rejects embedders with differing dimensions.
• types: SpeakerIdRemap::from_mapping runtime duplicate old SpeakerId check returns Option<Self>.
• clusterer: uniform embedding dimension checks in AhcClusterer, KMeansClusterer, and NmeScClusterer.

v0.6.8

Full Changelog: v0.6.7...v0.6.8