Speaker Identification Implementation

Speaker Identification Implementation Speaker Identification is the process of identifying a speaker from a database of known voices. Unlike diarization ("who spoke when"), it requires answering "who is this person?" It is used in authentication systems, personalized assistants, and broadcast monitoring. ### System Architecture

Audio → VAD → Speaker Encoder → Embedding → Similarity Search → Identity
                  (ECAPA-TDNN)    (d-vector)    (cosine / ANN)
```### Removing speaker embeddings```python
from speechbrain.pretrained import SpeakerRecognition
import torchaudio
import torch

# ECAPA-TDNN — state-of-the-art архитектура
model = SpeakerRecognition.from_hparams(
    source="speechbrain/spkrec-ecapa-voxceleb",
    savedir="tmp_spkrec"
)

def get_embedding(audio_path: str) -> torch.Tensor:
    signal, sr = torchaudio.load(audio_path)
    if sr != 16000:
        signal = torchaudio.functional.resample(signal, sr, 16000)
    embedding = model.encode_batch(signal)
    return embedding.squeeze()

# Регистрация нового говорящего
def register_speaker(name: str, audio_samples: list[str]):
    embeddings = [get_embedding(p) for p in audio_samples]
    mean_embedding = torch.stack(embeddings).mean(0)
    return mean_embedding  # сохраняем в базу
```### Search the voice database```python
import faiss
import numpy as np

# Индекс для быстрого поиска (миллионы голосов)
index = faiss.IndexFlatIP(192)  # cosine similarity через inner product
speaker_names = []

def add_speaker(name: str, embedding: torch.Tensor):
    emb_np = embedding.numpy().reshape(1, -1)
    faiss.normalize_L2(emb_np)
    index.add(emb_np)
    speaker_names.append(name)

def identify_speaker(audio_path: str, threshold: float = 0.75) -> str:
    embedding = get_embedding(audio_path).numpy().reshape(1, -1)
    faiss.normalize_L2(embedding)
    distances, indices = index.search(embedding, k=1)
    score = float(distances[0][0])
    if score >= threshold:
        return speaker_names[indices[0][0]]
    return "UNKNOWN"
```### ECAPA-TDNN EER (Equal Error Rate) performance on VoxCeleb1: 0.87% — industrial-level. When using 10+ seconds of recording for registration: accuracy >95% at a threshold of 0.8. ### Implementation timeline Basic identification system: 1 week. With FAISS index and vote database management: 2 weeks.