[Python] PyAnnote Speaker Verification: All Speakers Getting Perfect 1.000 Similarity Scores

I'm experiencing an issue with PyAnnote's speaker verification where all speakers are getting perfect similarity scores (1.000), even when they are clearly different voices.

Environment

• 
  pyannote.audio: 3.1.1
  
  
• 
  torch: 2.5.1+cu124
  
  
• 
  Platform: Google Colab (Ubuntu Linux)
  
  
• 
  CUDA: Yes
  
  
• 
  GPU: Tesla T4
  
  
• 
  Python: 3.11
  
  
• 
  torchaudio: 2.5.1+cu124
  

Issue

Using pyannote/embedding for speaker verification, all speakers are getting perfect similarity scores (1.000) when compared to a reference sample, even between obviously different speakers in a professional audiobook.

Code

python

import torch

import torchaudio

from pyannote.audio import Model

import torch.nn.functional as F

Setup

device = torch.device("cuda")

embedding_model = Model.from_pretrained("pyannote/embedding",

use_auth_token='xxx').to(device)

Load and process reference audio

reference_path = "reference.flac" # 10-minute reference file

reference_waveform, sample_rate = torchaudio.load(reference_path)

reference_waveform = reference_waveform.mean(dim=0, keepdim=True).to(device)

reference_features = embedding_model(reference_waveform.unsqueeze(0))

reference_features = F.normalize(reference_features, p=2, dim=1)

Load test audio segment

test_path = "test.flac" # Different speaker segment

test_waveform, = torchaudio.load(test_path)

test_waveform = test_waveform.mean(dim=0, keepdim=True).to(device)

test_embedding = embedding_model(test_waveform.unsqueeze(0))

test_embedding = F.normalize(test_embedding, p=2, dim=1)

Calculate similarity

similarity = F.cosine_similarity(reference_features, test_embedding, dim=1).mean()

print(f"Similarity score: {similarity.item():.6f}")

Current Behavior

• 
  Every speaker gets similarity scores of 0.999+ to 1.000
  
  
• 
  This happens consistently across different speakers
  
  
• 
  Reference and speaker embeddings both have shape [1, 512]
  
  
• 
  Even clearly different voices (male/female) get perfect matches
  

Debug Information

Example similarity scores between different speakers:

Speaker A vs Reference: 1.000000

Speaker B vs Reference: 0.999998

Speaker C vs Reference: 1.000000

Tensor Shapes:

Reference waveform shape: [1, 31246073]

Reference embedding shape: [1, 512]

Test embedding shape: [1, 512]

What I've Tried

• 
  Testing with different audio segments
  
  
• 
  Testing with obviously different speakers
  
  
• 
  Using high-quality audio sources
  

Question

Why are all speakers getting perfect similarity scores (1.000) regardless of how different their voices are? Is this a known issue with the current version of PyAnnote?

Continue reading...