Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

fix(diarization): scale clustering with known speaker counts

tandede9dbf3e1c0858
diff --git a/funasr/models/campplus/cluster_backend.py b/funasr/models/campplus/cluster_backend.py
index 998ccd679..a369d626d 100644
--- a/funasr/models/campplus/cluster_backend.py
+++ b/funasr/models/campplus/cluster_backend.py
@@ -11,6 +11,7 @@ import numpy as np
from sklearn.cluster._kmeans import k_means
from sklearn.cluster import HDBSCAN
+from sklearn.preprocessing import normalize
class SpectralCluster:
@@ -191,6 +192,21 @@ class UmapHdbscan:
return labels
+class KMeansCluster:
+ r"""Cluster a large set of speaker embeddings into a known number of groups."""
+
+ def __call__(self, X, num_clusters):
+ """Cluster L2-normalized embeddings with bounded memory usage."""
+ normalized_X = normalize(X)
+ _, labels, _ = k_means(
+ normalized_X,
+ num_clusters,
+ random_state=0,
+ n_init=10,
+ )
+ return labels
+
+
class ClusterBackend(torch.nn.Module):
r"""Perfom clustering for input embeddings and output the labels.
Args:
@@ -210,6 +226,7 @@ class ClusterBackend(torch.nn.Module):
self.spectral_cluster = SpectralCluster()
self.umap_hdbscan_cluster = UmapHdbscan()
+ self.kmeans_cluster = KMeansCluster()
def forward(self, X, **params):
# clustering and return the labels
@@ -223,9 +240,10 @@ class ClusterBackend(torch.nn.Module):
assert len(X.shape) == 2, "modelscope error: the shape of input should be [N, C]"
if X.shape[0] < 20:
return np.zeros(X.shape[0], dtype="int")
- if X.shape[0] < 2048 or k is not None:
- # unexpected corner case
+ if X.shape[0] < 2048:
labels = self.spectral_cluster(X, k)
+ elif k is not None:
+ labels = self.kmeans_cluster(X, k)
else:
labels = self.umap_hdbscan_cluster(X)
diff --git a/tests/test_cluster_backend.py b/tests/test_cluster_backend.py
new file mode 100644
index 000000000..c2a6daeec
--- /dev/null
+++ b/tests/test_cluster_backend.py
@@ -0,0 +1,44 @@
+import numpy as np
+import pytest
+import torch
+
+from funasr.models.campplus.cluster_backend import ClusterBackend
+
+
+def test_large_known_speaker_count_uses_fixed_k_clustering(monkeypatch):
+ backend = ClusterBackend()
+ embeddings = torch.ones((2048, 2))
+ expected = np.arange(embeddings.shape[0]) % 2
+
+ def fail_spectral(*args, **kwargs):
+ pytest.fail("large inputs must not use dense spectral clustering")
+
+ def fail_umap(*args, **kwargs):
+ pytest.fail("a known speaker count should use fixed-K clustering")
+
+ def fixed_k_cluster(actual_embeddings, num_clusters):
+ assert actual_embeddings is embeddings
+ assert num_clusters == 2
+ return expected
+
+ monkeypatch.setattr(backend, "spectral_cluster", fail_spectral)
+ monkeypatch.setattr(backend, "umap_hdbscan_cluster", fail_umap)
+ monkeypatch.setattr(backend, "kmeans_cluster", fixed_k_cluster, raising=False)
+
+ labels = backend(embeddings, oracle_num=2)
+
+ np.testing.assert_array_equal(labels, expected)
+
+
+def test_large_fixed_k_clustering_separates_cosine_clusters():
+ backend = ClusterBackend()
+ embeddings = torch.zeros((2048, 2))
+ embeddings[:1024, 0] = 1
+ embeddings[1024:, 1] = 1
+
+ labels = backend(embeddings, oracle_num=2)
+
+ assert np.unique(labels).size == 2
+ assert np.unique(labels[:1024]).size == 1
+ assert np.unique(labels[1024:]).size == 1
+ assert labels[0] != labels[-1]