README.md

---
license: cc-by-4.0
language:
  - multilingual
tags:
  - coreml
  - speaker-diarization
  - pyannote
  - wespeaker
  - apple-silicon
base_model: pyannote/speaker-diarization-community-1
library_name: coremltools
pipeline_tag: audio-classification
---

# Pyannote Community-1 Core ML

Precompiled Core ML neural stages and native VBx data for the offline
[Pyannote Community-1](https://huggingface.co/pyannote/speaker-diarization-community-1)
speaker-diarization pipeline on Apple platforms.

> Part of the [soniqo.audio](https://soniqo.audio) speech toolkit. See the
> [speaker-diarization guide](https://soniqo.audio/guides/diarize), inspect the
> [native Swift Community-1 runtime](https://github.com/soniqo/speech-swift/tree/a6ed5a5/Sources/SpeechVAD),
> or browse the [Core ML Speech Models collection](https://huggingface.co/collections/aufklarer/coreml-speech-models-69b52fb6987dffb06fd47570).

This is a pipeline bundle, not one end-to-end model. The host must run powerset
decoding, speaker counting, overlap-aware mask selection, VBx clustering, and
timeline reconstruction exactly as described by `config.json`.

## Model

| Component | Parameters | Precision | Format | Size | Context / sample rate |
|---|---:|---|---|---:|---|
| PyanNet segmentation | 1.49M | FP32 | compiled Core ML | 5.7 MiB | 10 s / 16 kHz mono |
| Masked WeSpeaker | 6.86M | FP32 | compiled Core ML | 26.2 MiB | 10 s + three 589-frame masks |
| PLDA transforms | — | FP32 | safetensors | 0.19 MiB | 256 to 128 dimensions |

The WeSpeaker graph includes the exact Kaldi filterbank and weighted statistics
pooling used by Community-1. Its vectors are intentionally not normalized before
PLDA. Both Core ML graphs use fixed batch size one and require iOS 17 or macOS 14
or later.

## Files

| File | Size | Description |
|---|---:|---|
| `segmentation.mlmodelc/` | 5.7 MiB | Precompiled PyanNet segmentation graph |
| `embedding.mlmodelc/` | 26.2 MiB | Precompiled filterbank and masked WeSpeaker graph |
| `plda.safetensors` | 0.19 MiB | Precomputed x-vector and PLDA transforms for VBx |
| `config.json` | — | Tensor interfaces, pipeline constants, hashes, and conversion checks |
| `benchmark.json` | — | Sanitized aggregate and per-recording benchmark results |
| `LICENSE` | — | CC BY 4.0 notice and upstream attribution |

## Performance

Lower diarization error rate (DER) and Jaccard error rate (JER) are better.
Throughput above 1x means faster than realtime.

| Evaluation | DER | JER | Throughput | Interpretation |
|---|---:|---:|---:|---|
| VoxConverse dev, five-file subset, 0.25 s collar, overlap included | 4.65% | 21.42% | 25.0x | Below 10% DER and exactly matches upstream Community-1 |
| speech-swift native runtime, same five files and scorer | 4.66% | 21.43% | 25.2x | 0.02 DER points from the published reference |
| Same five files, strict 0 s collar | 6.99% | 23.07% | 25.0x | Boundary errors are fully counted |
| AMI ES2004a single-meeting diagnostic, known 4 speakers | 23.00% | 23.62% | 24.0x | Exact upstream match; weak absolute accuracy on this meeting |

Tested on Apple M5 Pro with `cpu-and-neural-engine`. Peak process memory
was 863 MiB. The five-file speaker-count estimate
was exact for 3 of 5
recordings, so callers should allow a known or bounded count when available.

The neural stages themselves processed one 10-second window in a median
7.44 ms for segmentation and
33.24 ms for all three masked embeddings.
The benchmark used the official Community-1 host processing and revision
`3533c8cf8e369892e6b79ff1bf80f7b0286a54ee`. Scores and speaker counts matched the upstream PyTorch/MPS
run for every evaluated recording.

The VoxConverse result is a small five-recording release check, not a claim over
the full dataset. The AMI row is one meeting and is shown as a limitation, not a
representative AMI score.

## Swift runtime integration

The matching native runtime is available in
[`soniqo/speech-swift` on `feat/community1-coreml`](https://github.com/soniqo/speech-swift/tree/feat/community1-coreml)
at commit [`a6ed5a5`](https://github.com/soniqo/speech-swift/commit/a6ed5a5).
It runs both Core ML graphs, powerset decoding, speaker counting, PLDA, VBx,
constrained assignment, and timeline reconstruction without Python.

```bash
speech diarize meeting.wav --engine community1
speech diarize meeting.wav --engine community1 --num-speakers 2
speech diarize meeting.wav --engine community1 --min-speakers 2 --max-speakers 6
```

```swift
let diarizer = try await Community1DiarizationPipeline.fromPretrained()
try diarizer.prewarm()

let result = try diarizer.diarize(
    audio: samples,
    sampleRate: 16_000,
    speakerBounds: Community1SpeakerBounds(minimum: 2, maximum: 6)
)
```

The runtime returns diarized segments plus one 256-dimensional centroid for
each detected speaker. Speaker IDs are local to one result; use the centroids
for recording-local or persistent identity matching.

## Download

```bash
hf download aufklarer/Pyannote-Community-1-CoreML --local-dir Pyannote-Community-1-CoreML
```

## Python example

The following runs the segmentation stage. Complete diarization also needs the
host steps and PLDA data described in `config.json`.

```python
import json
from pathlib import Path

import coremltools as ct
import numpy as np

root = Path("Pyannote-Community-1-CoreML")
config = json.loads((root / "config.json").read_text())
model = ct.models.CompiledMLModel(
    str(root / config["segmentation"]["model"]),
    compute_units=ct.ComputeUnit.CPU_AND_NE,
)

# One 10-second, 16 kHz mono window in [-1, 1].
waveform = np.zeros((1, 1, 160_000), dtype=np.float32)
log_probabilities = model.predict({"waveform": waveform})["log_probabilities"]
print(log_probabilities.shape)  # (1, 589, 7)
```

On Apple platforms, load the `.mlmodelc` directories directly. Do not compile an
`.mlpackage` at runtime; compiled artifacts are provided to keep behavior stable
across macOS, iOS, and simulator runtimes.

## Source and license

Derived from
[pyannote/speaker-diarization-community-1](https://huggingface.co/pyannote/speaker-diarization-community-1)
at revision `3533c8cf8e369892e6b79ff1bf80f7b0286a54ee`. Community-1 combines Pyannote segmentation,
WeSpeaker embeddings, and VBx clustering and is distributed under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Preserve this
attribution when redistributing the bundle.

## Links

- [Speaker diarization guide](https://soniqo.audio/guides/diarize) — concepts and public APIs
- [speech-swift](https://github.com/soniqo/speech-swift) — Apple speech SDK
- [Native Community-1 runtime](https://github.com/soniqo/speech-swift/tree/a6ed5a5/Sources/SpeechVAD) — pinned Swift implementation
- [Community-1 runtime branch](https://github.com/soniqo/speech-swift/tree/feat/community1-coreml) — CLI, tests, and documentation
- [Core ML Speech Models](https://huggingface.co/collections/aufklarer/coreml-speech-models-69b52fb6987dffb06fd47570) — related Apple bundles
- [Getting started](https://soniqo.audio/getting-started) — installation and CLI guide
- [soniqo.audio](https://soniqo.audio) — website
- [Blog](https://soniqo.audio/blog) — updates and technical articles