X-VECTOR:ROBUST DNN EMBEDDING FOR SPEAKER RECOGNITION

2019/04/02

X-VECTOR:robust dnn embedding for speaker recognition

1. abstract

DNN + data augmentation -> better performance of speaker recognition

1. 1. DNN

• • • i. variable-length utterances –(mapping)-> fixed-dimensional embeddings(x-vectors) 
    • ii. + embeddings leverage large-scale training datasets better than i-vectors
    • iii. - challenging to collect substantial quantities of labeled data for training.
  • Solution: data augmentation (noise and reverberation)
    • i. An inexpensive method to multiply the amount of training data and improve robustness

 

1. 1. Result
      • i. Before augmentation: The x-vectors are compared with i-vector baselines on Speakers in the Wild and NIST SRE 2016 Cantonese.
      • ii. After augmentation:
        1. + PLDA beneficial
        2. – i-vector extractor
        3. + x-vector DNN: due to supervised training
        4. x-vector achieves  superior performance on the evaluation datasets.
2. Introduction
   1. I-vectors
      • i. Universal background model (high-dimensional statistics)–-(large projection matrix T (learned in an unsupervised way to maximize the data likelihood)) --> i-vector (a low-dimensional representation) + probabilistic linear discriminant analysis (PLDA) classifier
   2. DNN used to enhance phonetic modeling in the i-vector UBM
      • i. DNN: acoustic models for ASR
      • ii. Posteriors from DNN replace those from GMM
      • iii. Bottleneck features are extracted from the DNN and combined with acoustic features
      • iv. For in-domain data, the improvement of DNN is substantial
        1. - Transcribing data
        2. – Computational complexity
   3. DNN directly discriminate between speakers
      • i. Frame-level representations extracted from DNN + Gaussian speaker models
      • ii. End-to-end system: jointly embedding + similarity matric
      • iii. Ii->text-independent + a temporal pooling layer into the network to handle variable-length segments
      • iv. DNN embedding + separate classifier: accumulated backend technology developed over the years for i-vectors, such as length-normalization, PLDA scoring, and domain adaptation techniques.

 

1. SPEAKER RECOGNITION SYSTEMS
   1. Acoustic-vector: GMM-UBM
      • i. Features are 20 MFCCs with a frame-length of 25ms that are mean-normalized over a sliding window of up to 3 seconds.
      • ii. Delta and acceleration are appended to create 60 dimension feature vectors
      • iii. An energy-based speech activity detection (SAD) system selects features corresponding to speech frames.
      • iv. The UBM is a 2048 component full-covariance GMM.
      • v. The system uses a 600 dimensional i-vector extractor and PLDA for scoring.
   2. Phonetic bottleneck i-vector
      • i. DNN: time-delay acoustic model with p-norm nonlinearities
      • ii. Penultimate layer is replaced with a 60 dimensional linear bottleneck layer
      • iii. BNFs are concatenated with the same 20 dimensional MFCCs described in Section 2.1 plus deltas to create 100 dimensional features
      • iv. Feature processing, UBM, i-vector extractor, and PLDA classifier are identical to the previous acoustic system
   3. The x-vector system
      • i. The embedding DNN architecture: x-vectors are extracted at layer segment6, before the nonlinearity. The N in the softmax layer corresponds to the number of training speakers

1. • • ii. 24 dimensional filterbanks with a frame-length of25ms, mean-normalized over a sliding window of up to 3 seconds + The same energy SAD
     • iii. This pooling layer aggregates information across the time dimension so that subsequent layers operate on the entire segment.
     • iv. A training example consists of a chunk of speech features (about 3 seconds average), and the corresponding speaker label
     • v. 4.2 million Parameters.
   • PLDA classifier
     • i. Centered, and projected using LDA
     • ii. LDA dimension was tuned on the SITW development set to 200 for i-vectors and 150 for x-vectors.
     • iii. Length-normalized and modeled by PLDA
     • iv. Scores are normalized using adaptive s-norm
2. EXPERIMENTAL SETUP
   1. Training data

dataset	contents		size	Number of participants	sample rate	note	transcriptions
sre	04		unknown		8k	Telephone speech
	05		525h			English/Chinese
	06		437h
	08		942h			Multilingual
	10		2255h
Swbd	Phase 2	Part1	303h (3638*5min)	657		Conversation
		Part2	373h (4472*5min)	679		Telephone conversation
		Part3	222h (2728*5min)	640		Telephone speech, gender, environment	Provide transcription information
	Cellular	Part1	109h (1309*5min)	254		gender, environment	Provide transcription information
		Part2	200
sre16_eval_enroll	Call My Net speech collection					Channel mismatch, language mismatch
sre16_eval_test						From both same and different telephone numbers as the enrollment
sre16_major	SRE16 unlabeled data

 

1. results