Speaker Clustering Methods

1# Speaker Clustering Methods
2 
3## Overview
4 
5After extracting speaker embeddings from audio segments, you need to cluster them to identify unique speakers. Different clustering methods have different strengths.
6 
7## When to Use
8 
9- After extracting speaker embeddings from VAD segments
10- Need to group similar speakers together
11- Determining number of speakers automatically or manually
12 
13## Available Clustering Methods
14 
15### 1. Hierarchical Clustering (Recommended for Auto-tuning)
16 
17**Best for:** Automatically determining number of speakers, flexible threshold tuning
18 
19```python
20from scipy.cluster.hierarchy import linkage, fcluster
21from scipy.spatial.distance import pdist
22import numpy as np
23 
24# Prepare embeddings
25embeddings_array = np.array(embeddings_list)
26n_segments = len(embeddings_array)
27 
28# Compute distance matrix
29distances = pdist(embeddings_array, metric='cosine')
30 
31# Create linkage matrix
32linkage_matrix = linkage(distances, method='average')
33 
34# Auto-tune threshold to get reasonable speaker count
35min_speakers = 2
36max_speakers = max(2, min(10, n_segments // 2))
37 
38threshold = 0.7
39labels = fcluster(linkage_matrix, t=threshold, criterion='distance')
40n_speakers = len(set(labels))
41 
42# Adjust threshold if needed
43if n_speakers > max_speakers:
44    for t in [0.8, 0.9, 1.0, 1.1, 1.2]:
45        labels = fcluster(linkage_matrix, t=t, criterion='distance')
46        n_speakers = len(set(labels))
47        if n_speakers <= max_speakers:
48            threshold = t
49            break
50…