Neural Sync Active

Learning Objectives

Registry Synced

Learning Objectives

428 words

2 min read

Reading compass

Now · 1. Speaker Recognition Tasks

Learning Objectives

Understand speaker verification and identification
Apply i-vectors and x-vectors
Evaluate speaker recognition systems
Weeks 3-5: Features, GMMs
Understanding of binary classification

1. Speaker Recognition Tasks

Speaker Identification: Who is speaking? (N-class classification) Speaker Verification: Is this the claimed speaker? (Binary: Accept/Reject)

2. Feature Extraction

MFCCs: Capture vocal tract shape (speaker-dependent) F0 statistics: Pitch range, mean, variation Spectral features: Formants, spectral tilt Prosodic features: Speaking rate, rhythm patterns Key requirement: Features should capture speaker characteristics but NOT linguistic content.

3. i-Vectors (Total Variability)

Map variable-length utterance to fixed-length vector (i-vector):

Universal Background Model (UBM): GMM trained on many speakers
Baum-Welch statistics: Collect sufficient statistics per utterance
Total variability matrix T projects statistics to low-dimensional i-vector Scoring: Cosine distance or PLDA (Probabilistic Linear Discriminant Analysis)

4. x-Vectors (DNN Embeddings)

Deep neural network extracts speaker embeddings:

Input: MFCCs with context window
Frame-level layers (TDNN): Process temporal context
Statistics pooling: Aggregate frame-level into utterance-level
Embedding layer (x-vector): Speaker representation
Output layer: Speaker classification during training PLDA Scoring: Compares two x-vectors for verification.

Q1: What is the difference between speaker identification and verification?

Identification: who is speaking? 1:N matching. Verification: is this who they claim to be? 1:1 matching. Verification has threshold (accept/reject), identification returns ranked list. Q2: What are i-vectors and how are they extracted?

Fixed-length speaker vectors from variable-length utterances. Steps: UBM (universal GMM), compute sufficient statistics, project using total variability matrix T. I-vectors capture both speaker and channel variability but can be compensated. Q3: How do x-vectors differ from i-vectors?

x-vectors: DNN-based embeddings. Use time-delay neural network (TDNN) to process temporal context. Statistics pooling converts frame-level to utterance-level. Generally better than i-vectors with sufficient data. Q4: What is PLDA scoring?

Probabilistic Linear Discriminant Analysis. Computes likelihood ratio: same speaker vs different speakers given two embeddings. Accounts for within-speaker and between-speaker variability. Standard scoring method for both i-vectors and x-vectors. Q5: How is speaker verification evaluated?

Equal Error Rate (EER): where false acceptance = false rejection. Lower is better. Detection Cost Function (DCF): weighted sum of errors (different costs for FA and FR). MinDCF = minimum achievable at optimal threshold. Q6: X-vector extraction example:
python
import speechbrain as sb
# Pretrained x-vector model
classifier = sb.inference.SpeakerRecognition.from_hparams(
    source="speechbrain/spkrec-ecapa-voxceleb")
embedding = classifier.encode_file("speech.wav")
# embedding is a 192-dimensional x-vector

Join Discord PreviousText-to-Speech Synthesis (TTS)NextSpeech Enhancement & Noise Robustness