2017年10月31日 (二) 10:53的版本

Project name

Phonetic Temporal Neural Model for Language Identification

Project members

Dong Wang, Zhiyuan Tang, Lantian Li, Yixiang Chen, Ying Shi

Introduction

Deep neural models, particularly the LSTM-RNN model, have shown great potential for language identification (LID). However, the use of phonetic information has been largely overlooked by most existing neural LID methods, although this information has been used very successfully in conventional phonetic LID systems. We present a phonetic temporal neural model for LID, which is an LSTM-RNN LID system that accepts phonetic features produced by a phone-discriminative DNN as the input, rather than raw acoustic features. This new model is similar to traditional phonetic LID methods, but the phonetic knowledge here is much richer: it is at the frame level and involves compacted information of all phones. Our experiments conducted on the Babel database and the AP16-OLR database demonstrate that the temporal phonetic neural approach is very effective, and significantly outperforms existing acoustic neural models. It also outperforms the conventional i-vector approach on short utterances and in noisy conditions.

Speaker feature learning

The discovery of the short-time property of speaker traits is the key step towards speech signal factorization, as the speaker trait is one of the two main factors: the other is linguistic content that we have known for a long time being short-time patterns.

The key idea of speaker feature learning is simply based on the idea of discriminating training speakers based on short-time frames by deep neural networks (DNN), date back to 2014 by Ehsan et al.[2]. As shown below, the output of the DNN involves the training speakers, and the frame-level speaker features are read from the last hidden layer. The basic assumption here is: if the output of the last hidden layer can be used as the input feature of the last hidden layer (a software regression classifier), these features should be speaker discriminative.

However, the vanilla structure of Ehsan et al. performs rather poor compared to the i-vector counterpart. One reason is that the simple back-end scoring is based on average to derive the utterance-based representations (called d-vectors) , but another reason is the vanilla DNN structure that does not consider much of the context and pattern learning. We therefore proposed a CT-DNN model that can learn stronger speaker features. The structure is shown below[1]:

Recently, we found that an 'all-info' training is effective for learning features. Looking back to DNN and CT-DNN, although the features read from last hidden layer are discriminative, but not 'all discriminative', because some discriminant info can be also impelemented in the last affine layer. A better strategy is let the feature generation net (feature net) learns all the things of discrimination. To achieve this, we discarded the parametric classifier (the last affine layer) and use the simple cosine distance to conduct the classification. An iterative training scheme can be used to implement this idea, that is, after each epoch, averaging the speaker features to derive speaker vectors, and then use the speaker vectors to replace the last hidden layer. The training will be then taken as usual. The new structure is as follows[4]:

Speech factorization

The short-time property is a very nice thing, which tells us it is possible to factorize speech signals. By factorization, we can achieve significant benefits:

A. Individual tasks can be largely improved, as unrelated factors have been removed.

B. Factors that are disturbs now becomes valuables things, leading to conditional training and collaborative training [5].

C. Once the factors have been separated, single factors can be manipulated, and reassemble these factors can change the signal according to the need.

D. It is a new speech coding scheme that leverage knowledge learned from large data.

Traditional factorization methods are based on probabilistic models and maximum likelihood learning. For example, in JFA, a linear Gaussian is assumed for speaker and channel, and then a ML estimation is applied to estimate the loading matrices of each factor, based on a long duration of speech. Almost all these factorizations share these features: shallow, linear, Gaussian, long-term segments.

We are more interested in factorization on frame-level, and plays not much assumption on how the factors are mixed. A cascaded factorization approach has been proposed[6]. The basic idea is to factorize significant factors first,and then conditioned on the factors that have been derived. The architecture is as follows, where we factorized speech signals into three factors: linguistic content, speaker trait, emotion. When factorizing each factor, supervised learning is used. Note by this architecture, databases with different target labels can be used in a complementary way, which is different from previously joint training approach that needs full-labelled data.

Speech reconstruction

To verify the factorization, we can reconstruct the speech signal from the factors. The reconstruction is simply based on a DNN, as shown below. Each factor passes a unique deep neural net, the output of the three DNNs are added together, and compared with the target, which is the logarithm of the spectrum of the original signal. This means that the output of the DNNs of the three factors are assumed to be convolved together to produce the original speech.

Note that the factors are learned from Fbanks, by which some speech information has been lost, however the recovery is rather successfull.

View the reconstruction

More recovery examples can be found here.

Listen to the reconstruction

We can listen to the wave for each factor, by using the original phase.

Original speech: [1]

Linguistic factor: [2]

Speaker factor: [3]

Emotion factor: [4]

Liguistic+ Speaker + Emotion: [5]

Research directions

• Adversarial factor learning
• Phone-aware multiple d-vector back-end for speaker recognition
• TTS adaptation based on speaker factors

Reference

[1] Lantian Li, Yixiang Chen, Ying Shi, Zhiyuan Tang, and Dong Wang, “Deep speaker feature learning for text-independent speaker verification,”, Interspeech 2017.

[2] Ehsan Variani, Xin Lei, Erik McDermott, Ignacio Lopez Moreno, and Javier Gonzalez-Dominguez, “Deep neural networks for small footprint text-dependent speaker verification,”, ICASSP 2014.

[3] Lantian Li, Dong Wang, Yixiang Chen, Ying Shing, Zhiyuan Tang, http://wangd.cslt.org/public/pdf/spkfact.pdf

[4] Lantian Li, Zhiyuan Tang, Dong Wang, FULL-INFO TRAINING FOR DEEP SPEAKER FEATURE LEARNING, http://wangd.cslt.org/public/pdf/mlspk.pdf

[5] Zhiyuan Thang, Lantian Li, Dong Wang, Ravi Vipperla "Collaborative Joint Training with Multi-task Recurrent Model for Speech and Speaker Recognition", IEEE Trans. on Audio, Speech and Language Processing, vol. 25, no.3, March 2017.

[6] Dong Wang,Lantian Li,Ying Shi,Yixiang Chen,Zhiyuan Tang., "Deep Factorization for Speech Signal", https://arxiv.org/abs/1706.01777