Expand description
diaric
The backend-free speaker-diarization core: clustering, PLDA, pipeline assembly, reconstruction, fbank DSP, score normalization and SIMD/mmap numeric ops — with no ONNX/Torch dependency.
§What this is
diaric is the sans-I/O, model-runtime-free half of the
diarization pipeline. It
holds everything that computes over already-produced tensors and audio —
speaker clustering, PLDA projection, the offline diarization assembly,
frame reconstruction / RTTM emission, the kaldi-compatible fbank feature
extractor, adaptive score normalization, and the shared numeric kernels —
but none of the ONNX Runtime (ort) or LibTorch (tch) model
runners.
The segmentation and embedding model runners (SegmentModel,
EmbedModel) and the streaming service layer live in the diarization
crate, which depends on diaric. Consumers with their own inference path
(CoreML, custom CUDA, captured tensors) can depend on diaric directly and
never pull in a native ML runtime.
§Modules
| Module | What it does |
|---|---|
cluster | Offline AHC → VBx clustering, spectral fallback, Hungarian assignment, and the online greedy centroid matcher (FluidAudio SpeakerManager port). |
plda | PLDA projection over WeSpeaker embeddings. Weights are embedded via include_bytes! (models/plda/*.bin). |
pipeline | assign_embeddings — the full AHC + VBx + centroid + constrained-Hungarian assignment stage. |
offline | diarize_offline — batch pyannote-equivalent diarization over pre-computed (segmentation, raw-embedding) tensors. |
aggregate / reconstruct | Speaker counting and frame-level reconstruction → RTTM spans. |
segment | The sans-I/O Segmenter windowing/hysteresis state machine, powerset decoding, and option constants (the ONNX SegmentModel runner stays in diarization). |
embed | The Embedding value types and the bit-exact torchaudio kaldi-fbank DSP (the ONNX EmbedModel runner stays in diarization). |
score_norm | Adaptive score normalization (AS-Norm1). Rescales trial scores against a cohort distribution so one threshold means the same thing for every speaker. Generic over the score source — no embedding dimension in its surface. |
ops / spill | Three-tier (scalar / NEON / AVX2 / AVX-512) numeric primitives and the file-backed mmap spill backend for pathological-size inputs. |
provenance | Model/PLDA identity metadata. |
§Usage
Add diaric to your Cargo.toml:
[dependencies]
diaric = "0.2"Enable serde for Serialize/Deserialize on the public *Options
types.
§Cargo features
| Feature | Default | What it enables |
|---|---|---|
serde | no | Serialize/Deserialize impls for the public *Options types. Duration fields serialize as humantime strings (“250ms”, “1.5s”). |
_bench | no | Internal — exposes the crate-private ops numeric layer (plus ops::pdist_euclidean) to the benches/*.rs harnesses. ops is the only module it affects, and the only one outside the public API; cluster, pipeline, reconstruct and plda are public surface under semver. |
There is deliberately no ort / tch / execution-provider feature: the
model runners that need them are in diarization.
§License
diaric is offered under the composite SPDX expression
(MIT OR Apache-2.0) AND Apache-2.0 AND MIT AND CC-BY-4.0 AND BSD-2-Clause AND BSD-3-ClauseThe (MIT OR Apache-2.0) branch is the original diaric code (caller’s
choice); the remaining terms are mandatory obligations from the third-party
components this core vendors as source ports or embedded data. See the
LICENSE-APACHE, LICENSE-MIT, and NOTICE files for the full third-party
attribution record.
§LICENSE-MAPPING — which term covers what
| SPDX term | Component in diaric | Provenance |
|---|---|---|
| MIT OR Apache-2.0 | The original diaric / diarization Rust code (caller’s choice). | — |
| Apache-2.0 | cluster::online — the greedy online centroid matcher, a source port of FluidAudio’s SpeakerManager. A mandatory obligation, not the OR branch above: choosing MIT for the original code does not discharge it. | FluidAudio (Apache-2.0), FluidInference Team. |
| MIT | The offline cluster flow (AHC/VBx/spectral/centroid), pipeline, reconstruct, aggregate, the segment post-processing, and the PLDA math — algorithm ports of pyannote.audio. | pyannote/pyannote-audio (MIT) |
| CC-BY-4.0 | models/plda/*.bin — PLDA weights embedded into the compiled binary via include_bytes! (src/plda/loader.rs). Attribution is required in any redistributed binary. | pyannote/speaker-diarization-community-1; trained by BUT Speech@FIT. See models/plda/SOURCE.md. |
| BSD-2-Clause | src/embed/fbank.rs — a port of torchaudio.compliance.kaldi.fbank (torchaudio 2.11). torchaudio is BSD-2-Clause. | PyTorch/torchaudio (BSD-2-Clause), © 2017 Facebook Inc. (Soumith Chintala). |
| BSD-3-Clause | src/cluster/hungarian/lsap.rs (a port of SciPy’s rectangular_lsap.cpp) and the scipy-derived PLDA eigenvector blobs (eigenvectors_desc, phi_desc). | SciPy scipy.optimize (BSD-3-Clause). |
Note on the adapted NOTICE. The
NOTICEfile is adapted fromdiarization’s — materially, not carried verbatim: its preamble is rewritten fordiaric’s footprint and it adds the pyannote.audio source-port attribution (section 7). Itspyannote/segmentation-3.0(bundled ONNX) and WeSpeaker entries describe models thediarizationcrate bundles or loads —diaricships no ONNX/Torch model files. The only model artifactdiaricembeds is the CC-BY-4.0 PLDA weight set above; a downstream binary linkingdiaricmust reproduce that PLDA attribution and the source-port notices (pyannote, SciPy, torchaudio, FluidAudio).
Copyright (c) 2026 FinDIT studio authors.
Modules§
- aggregate
- Pyannote-equivalent
Inference.aggregateprimitives. - cluster
- Speaker clustering — generic offline batch
cluster_offline, the pyannotecluster_vbx-pipeline primitives (ahc,vbx,centroid,hungarian), and a streamingonlineengine. - embed
- Speaker-embedding value types + the kaldi-compatible fbank DSP.
- offline
- Offline (non-streaming) diarization.
- pipeline
- Full pyannote-equivalent batch clustering pipeline.
- plda
- PLDA (Probabilistic Linear Discriminant Analysis) transform.
- provenance
- Model-identity accessors so a downstream pipeline reads the loaded model’s family + version instead of hardcoding them, giving each diarization result a reproducible record of which model artifacts (segmentation, embedding, clustering/PLDA) produced it.
- reconstruct
- Pyannote reconstruction stage: hard_clusters + segmentations + count
→ per-output-frame discrete diarization (binary
(frames, clusters)grid). - score_
norm - Adaptive score normalization (AS-Norm) over speaker-trial scores.
- segment
- Speaker segmentation: Sans-I/O state machine + powerset post-processing.
- spill
- Spill-buffer configuration types reachable from public API surfaces
(e.g.
OfflineInput::with_spill_options,AssignEmbeddingsInputandReconstructInput).
Functions§
- axpy_
f32 Non- _bench - f32 fused multiply-add primitive:
y[i] += alpha * x[i]for eachi.