Streaming and webcasts are popular in this age of high–speed internet and mobile networks. With the ever increasing amount of audio–visual media data, the ability to index their contents and search for them is becoming more and more important. For data with speech contents, using Automatic Speech Recognition (ASR) to get the transcripts, is an efficient way to search and browse through thousands of hours of recordings. Error rates for the traditional broadcast news programmes could reach below 10% even in 1990s [1, 2, 3]. However broadcast media is not just limited to clean and read studio speech but also includes other types of multi–genre data with diverse speakers, variety of acoustic and recording conditions and diversity of the topics covered resulting in complex acoustic, lexical and linguistic conditions which is not yet well studied .
The wide variety of conditions in complex broadcast media causes mismatch between training and testing data, and therefore degrades the performance of the speech recognition systems 
. Adaptation can compensate for this mismatch. For Gaussian Mixture Model/Hidden Markov Model (GMM/HMM) systems several well established adaptation methods exist. However, adaptation of Deep Neural Networks (DNNs) is still a very active research topic. DNN adaptation methods can be divided into these three main categories:
Subspace methods: a speaker/environment subspace is estimated and then neurons’ weights or transformations are computed, based on the subspace representation of the speaker/environment. Principle Component Analysis (PCA) based adaptation approach
, i-Vector based speaker–aware training or speaker–aware DNNs  can be considered as subspace methods.
Broadcast media is complex in nature. For instance, in a news programme, there are in–studio reporting and live coverage on the scene. Assuming that all content variation from a single show can be described by a single, vaguely–defined domain is unrealistic and also not that helpful to ASR. Nonetheless it is clear that certain show types have very specific characteristics. Being able to assign broadcast media to a mixture of domains can alleviate this problem. Latent Dirichlet Allocation (LDA) is a statistical approach to discover latent variables in a collection of data that is describable with first–order statistic, in an unsupervised manner 
. It is mostly used in Natural Language Processing (NLP) for the categorisation of text documents, but it has been used for audio and image processing as well. In audio tasks, LDA has been used for classifying unstructured audio files into onomatopoeic and semantic descriptions with successful results. We have previously used LDA for domain adaptation of GMM/HMM systems .
This paper builds on this knowledge, and introduces a method on how to use LDA for domain adaptation of hybrid DNNs. Using LDA models, a data class - further referred to as “LDA domain”, is chosen for each utterance. The class information is then provided to the DNN in training. The learning algorithm adjusts the model parameters to exploit these additional information. During testing the same information is supplied. We further refer to this method as Latent–Domain–aware Training (LDaT). Results shown later in this paper indicate significant improvements of LDaT over baseline and input–adapted DNNs.
2 Latent Dirichlet Allocation
Latent Dirichlet Allocation (LDA)  is an unsupervised probabilistic generative model for collections of discrete data. It aims to describe how every item within a collection is generated, assuming that there are a set of latent variables and that each item is modelled as a finite mixture over those latent variables. LDA was originally used for topic modelling of text corpora; however, it is a generic model and can be applied to other tasks, such as object categorisation and localisation in image processing , automatic harmonic analysis in music processing , acoustic information retrieval in unstructured audio analysis  and our previous work for domain adaptation of GMM/HMM systems .
A dataset is defined as a collection of sets where each set is in turn a collection of discrete symbols (in case of topic modelling of text documents, a document is equivalent to a set and words inside a document are equivalent to the discrete symbols). Each set is represented by a -dimensional vector based on the histogram of the symbols’ table which has size of . It is assumed that the sets were generated by the following generative process:
For each set , choose a –dimensional latent variable weight vector from the Dirichlet distribution with scaling parameter :
For each discrete item in set
Draw a latent variable from the multinomial distribution
Given the latent variable, draw a symbol from , where is a matrix and
It is assumed that each set can be represented as a bag–of–symbols - i.e. by first–order statistics, which means any symbol sequence relationship is disregarded. Since speech and text are highly ordered processes this can be an issue. Another assumption is that the dimensionality of the Dirichlet distribution is fixed and known (and thus the dimensionality of the latent variable ).
A graphical representation of the LDA model is shown at Figure 1 as a three–level hierarchical Bayesian model. In this model, the only observed variable is and the rest are all latent. and are dataset level parameters, is a set level variable and , are symbol level variables. The generative process is described formally as:
The posterior distribution of the latent variables given the symbols and and parameters is:
Computing requires some intractable integrals. A reasonable approximate can be acquired using variational approximation, which is shown to work reasonably well in various applications . The approximated posterior distribution is:
where is the Dirichlet parameter that determines and is the parameter for the multinomial that generates the latent variables.
3 Acoustic LDA
As outlined above, LDA is a model to describe latent factors in sets of discrete symbols  which are here interpreted as “domains”. In order to fit into that concept speech signals need to be converted into such a form. Typically speech is represented using continuous features (e.g. with Mel frequency cepstral coefficients), and has variable length. In our previous work  we used Linde–Buzo–Gray vector quantization algorithm  to represent each speech frame with a discrete symbol, equivalent to an acoustic word or phone label.
In this paper an approach similar to that used in  was implemented. A GMM model with
components is trained using all of the training data. The model is then used to get the posterior probabilities of the Gaussian components to represent each frame with index of the Gaussian component with the highest posterior probability. Frames of every speech segment of length, are represented as:
where is the Gaussian component. After applying this process to each utterance, each speech segment is represented as where is index of the Gaussian component and thus a natural number . Here we refer to each speech utterance as an acoustic document. With this information, a fixed length vector of size were constructed to represent the count of every Gaussian component in an acoustic document. This leads to a type of bag–of–sounds representation. The sounds would normally be expected to relate to phones, however given the acoustic diversity of background conditions many other factors may play a role. Once these bag–of–sounds representations of acoustic documents are derived, LDA models can be trained.
4 LDA–DNN Adaptation
After acoustic symbols are established and speech segments are represented as bag–of–sounds, LDA models with designated latent domain sizes are trained using the variational EM algorithm . Hence, the posterior distribution of latent domains () for each utterance is computed. Since there can be many utterances in the training set, to effectively incorporate domain information in the vast amount of data, each utterance is assigned to only one domain. The assignment is made according to the maximum posterior estimate of domains .
The maximum posterior assumption requires high domain homogeneity for each acoustic document. This can to some degree be controlled by the size of domains. With a large number of domains, the resolution may be too high and the domain homogeneity within one acoustic document may be therefore lowered. On the other hand it is desirable to have a sufficient number of domains such that the variability in shows and between different types of shows are sufficiently covered.
Finally, domain information derived from the LDA model with domains is encoded with a -dimensional one–hot vector called Unique Binary Index Code (UBIC) . UBIC indicates the most likely domain of the utterance using the posterior domain probability. UBIC is then used to augment the input feature vectors. Apart from the extra nodes and connections in the input layer, the DNN architecture is identical to other baseline DNNs which are not domain–aware.
With the baseline DNNs, activation of the first layer is:
where superscripts denote the layer index, is the activation vector of the first layer, and
are the weight matrix and bias vector associated with layerand is the input features. With augmented UBIC in LDaT training this becomes:
where is the –dimensional domain assignment vector from the LDA model, is the weigh matrix for the acoustic features and it is initialised from of equation 6. is the weigh matrix for the augmented LDA domain assignment input. Comparing equations 6 and 7 the only difference is in the bias vector where there was a fixed bias before () and now with the augmented LDA domain information, there is a new adapted bias for each of the LDA domains. This type of adaptation is efficient, since it is implicit in the training process and does not require further adaptation steps . Figure 2 illustrates the DNN architecture with the augmented UBIC code.
5 Experimental Setup
TV broadcasts from the BBC were selected for the experiments. The data is identical to the one defined and provided for the 2015 Multi–Genre Broadcast (MGB) Challenge [23, 24, 25]. The shows were chosen to cover the full range of broadcast show types in current TV, and categorised in terms of 8 genres: advice, children’s, comedy, competition, documentary, drama, events and news. Acoustic Model training data was fixed and limited to more than 2,000 shows, broadcast by the BBC during 6 weeks in April and May of 2008. The development data for the task was 47 shows broadcast by the BBC during a week in mid–May 2008. The amount of shows and broadcast time for training and development data is shown in Table 1.
For the training data high quality transcription was not available. Instead only the subtitle text broadcast with each show plus an aligned version of the subtitles were available where the time stamps of the subtitles had been corrected in a lightly supervised manner . After this process, the new transcripts for the training shows had two potential problems: first, the subtitle text might not always match the actual spoken words and second, the time boundaries given might have errors arising from the lightly supervised alignment. To alleviate these two problems, only segments with Word Matching Error Rate (WMER) of lower than were used, which yielded around 500h of data. The WMER was a by–product of the semi–supervised alignment process that measures how similar the text in the subtitle matched the output of a lightly supervised ASR system for that segment .
For the Language Model (LM) subtitles from shows broadcast from 1979 to March 2008, with a total of 650 million words were used to train statistical language models.
Initial models were GMM/HMM systems with 13 dimensional PLP  features where four neighbouring frames on each side were spliced together to form a 117–dimensional feature vector. Using Linear Discriminant Analysis 
this feature vector was projected down to 40–dimensional vector and a global Constrained Maximum Likelihood Linear Regression transformation was applied to de–correlate the features. Speaker Adaptive Training (SAT)  was performed and then models were discriminatively trained using the Boosted Maximum Mutual Information criterion  and used to get the state level alignments for the DNN training. The input to the DNN was 440 dimensional PLP features that were
5 frames to the left/right of the current frame. The network had 6 hidden layers of size 2048 and an output layer of size 6478. The network was initialised using Deep Belief Network
pre–training and then trained to optimise per frame Cross Entropy objective function with Stochastic Gradient Descent. A speaker adapted DNN was also trained as the second base–line system using SAT style training. Speaker–based CMLLR transformations were applied to the features to make the inputs of the DNN closer to an average speaker. The Kaldi open–source speech recognition toolkit was used to train the acoustic models.
For decoding a 50k lexicon with a highly pruned 3–gram language model was used to generate lattices and then those lattices were re–scored using a 4–gram language model. Both of the language models were trained on the 650M words of the subtitles data using the SRILM toolkit
Table 2 presents the Word Error Rate (WER) of the development set with baseline models. There is a 19% relative WER reduction from GMM/HMM models to the baseline DNN models as usually expected. Speaker–adapted DNN also yields a further 6% relative WER reduction compared to the un–adapted DNN.
5.3 LDA–DNN Experiments
A GMM with 4k Gaussian mixtures is constructed using the mix–up procedure. Using this GMM, the audio frames are mapped to discrete symbols to train the LDA models . With LDA models, we experimented with different number of latent domains, namely , , , , , , , and .
For each of the domain sizes mentioned above, we computed the average domain entropy over all acoustic documents. Entropy increases from to when domain size increases from to . In this experiment, domain sizes and were used. This leveraged the considerations about the homogeneity and sparsity of the discovered domains discussed in section 4.
Apart from selecting an appropriate size of domain, cross-agreement data filtering was performed to ensure high domain homogeneity for each acoustic document. A domain-tuple with items was established. These items come from the Cartesian product of the domain mappings from the two corresponding LDA models. It is assumed that the two LDA models share a significant portion of the domains. If there is a high heterogeneity within an acoustic document, maximum-a-posteriori domain assignment from either or both LDA models will not be accurate, and they would appear in the rare classes in the domain–tuple items. Histogram pruning based on normalised pairs counts was performed to remove those rare items. The pruning cut–off was determined to result in a target training set size of around 500h, which was comparable to the data amount in our previous baseline experiments. Figure 3 shows the amount of data (in hours) for each of the 64 LDA domains.
The baseline DNN systems had an input layer of size
. That input was expanded by augmenting the LDA inferred domain with one–hot encoding. The new input had the size of(). The new LDA–DNN was trained similarly to the base line DNNs. Table 3 shows the frame classification accuracy of DNNs on a held–out cross–validation set with and without augmenting UBIC vectors.
|Model||CV Set Frame Accuracy (%)|
|Without LDA UBIC||LDA UBIC|
|Speaker Adapted DNN||48||46|
Table 4 presents the WER of baseline and adapted models for all of the eight genres. LDaT training reduces the WER from 33.3% to 30.6%, which is even better than speaker adapted DNN (31.4%). Combining speaker adaptation and domain adaptation (SAT+LDaT, linear input transformation for the speaker and bias adaptation for the latent domain) yields 28.9%, which is 13% relative WER reduction compared to the baseline DNN model and 8% relative improvement over the speaker adapted DNN. This also suggests that LDA inferred domains were not speaker clusters (since combining two adaptations still improves the performance). Because of the diverse nature of the data used, WER differs a lot across genres. Namely comedy and drama had the highest errors (43.8% and 45.0% respectively with LDaT+SAT models) showing the difficult nature of these genres. On the other hand, news had the lowest WER (14.3%). The WER diversity across the genres was consistent between all of the four models presented in table 4.
6 LDA Domain Analysis
It is of interest to understand how the data is structured by the LDA model. Unfortunately ground truth labelling is only available for words and 8 different genres, and for both labels the quality is highly variable. One would suspect that the words themselves are less important, however acoustic attributes such as the presence of music or laughter may be very informative. Unfortunately such labels are not available. However one can still get some impression on the differences by looking at the raw relationship to genres and differences between individual shows.
The domain assignment with the procedure outlined above, is visualised for the training data. The amount of data (time) assigned to an LDA domain is accumulated, for each of the 8 genres. Figure 4 shows the distribution of data for the most important 16 LDA domains (based on duration), across genres. All remaining domains have been subsumed into one group at the top of the figure, for better illustration. In this and the following figure LDA domains are sorted by the amount of data overall. From the graph it is clear that different genres exhibit significantly different LDA domain composition, with significant fine structure. Therefore such domain classification is very useful for genre classification.
One can also investigate how the LDA domain assignment varies within a genre, and between genres. In particular multiple episodes of shows are interesting in such analysis as one should expect high similarity due to similar programme structure. We obtained distributions for two sample programmes from two different genres. Figure 5 shows the LDA domain distribution for 8 episodes of Bargain Hunt (competition), followed by a further 8 from Waking the Dead (drama). Again 16+1 domains are displayed. One can observe that the distribution shows similarity within a genre (e.g., similarities of the red region on the lower left corner or the green area on the lower right corner). However between the two genres clear and systematic differences can be observed. One can further observe that more than 50% of each show is typically described by the top 2 or 3 LDA domains, and these differ in case of different genres but agree for the same programme within the genre. This indicates that individual shows are far more consistently described than the accumulated statistic allows to observe.
This paper introduced a new method, latent-domain-aware training, to adapt Deep neural networks to new domains. The method employs acoustic Latent Dirichlet Allocation to identify acoustically distinctive data clusters. These so-called LDA domains are then encoded using one–hot encoding, and used to augment standard input features for DNNs in training and testing. We further introduced coherence data selection to improve classification quality, and presented results on a diverse set of BBC TV broadcasts, with 500h of training and 28h of testing data. Word Error Rate reduction of 13% relative was achieved using the proposed adaptation method, compared to the baseline hybrid DNNs.
The proposed method lends itself to several future investigations. In the current LDA domain representation, each domain is described as a point on one of the axes of a high–dimensional space, where all have same distance from each other. Representing these points differently so that similar domains became closer in that space and verifying how that improves the performance can be an interesting problem to verify as a future work. Newer sets of features, better targeted to describe background acoustic characteristics , could also provide an improvement.
8 Data Access Statement
-  P. C. Woodland, M. J. F. Gales, D. Pye, and S. J. Young, “Broadcast news transcription using HTK,” in Proc. of ICASSP, Munich, Germany, 1997.
-  J. L. Gauvain, L. Lamel, and G. Adda, “The LIMSI broadcast news transcription system,” Speech Communication, vol. 37, no. 1–2, pp. 89–108, 2002.
-  M. J. F. Gales, D. Y. Kim, P. C. Woodland, H. Y. Chan, D. Mrva, R. Sinha, and S. E. Tranter, “Progress in the CU-HTK broadcast news transcription system,” IEEE Trans. on Audio, Speech and Language Processing, vol. 14, no. 5, pp. 1513–1525, 2006.
-  P. Lanchantin, P. Bell, M. Gales, T. Hain, X. Liu, Y. Long, J. Quinnell, S. Renals, O. Saz, and M. Seigel, “Automatic transcription of multi–genre media archives,” in Proc. of SLAM, Marseille, France, 2013.
M. Doulaty, O. Saz, and T. Hain,
“Data-selective transfer learning for multi-domain speech recognition,”in Proc. of Interspeech, Dresden, Germany, 2015.
D. Yu and L. Deng,
Automatic Speech Recognition: A Deep Learning Approach, Springer-Verlag, London, UK, 2015.
-  V. Abrash, H. Franco, A. Sankar, and M. Cohen, “Connectionist speaker normalization and adaptation,” in Proc. of EuroSpeech, Madrid, Spain, 1995.
-  B. Li and K. C. Sim, “Comparison of discriminative input and output transformations for speaker adaptation in the hybrid nn/hmm systems,” in Proc. of Interspeech, Makuhari, Japan, 2010.
-  R. Gemello, F. Mana, S. Scanzio, P. Laface, and R. De Mori, “Linear hidden transformations for adaptation of hybrid ANN/HMM models,” Speech Communication, vol. 49, no. 10, pp. 827–835, 2007.
-  J. Stadermann and G. Rigoll, “Two-stage speaker adaptation of hybrid tied-posterior acoustic models.,” in Proc. of ICASSP, Philadelphia, USA, 2005.
-  R. Doddipatla, M. Hasan, and T. Hain, “Speaker dependent bottleneck layer training for speaker adaptation in automatic speech recognition,” in Proc. of Interspeech, Singapore, 2014.
-  S. Dupont and L. Cheboub, “Fast speaker adaptation of artificial neural networks for automatic speech recognition,” in Proc. of ICASSP, Istanbul, Turkey, 2000.
-  G. Saon, H. Soltau, D. Nahamoo, and M. Picheny, “Speaker adaptation of neural network acoustic models using i-Vectors,” in Proc. of ASRU, Olomouc, Czech Republic, 2013.
-  Y. Liu, P. Karanasou, and T. Hain, “An investigation into speaker informed DNN front-end for LVCSR,” in Proc. of ICASSP, Brisbane, Australia, 2015.
D. M. Blei, A. Y. Ng, and M. I. Jordan,
“Latent Dirichlet Allocation,”
Journal of Machine Learning Research, vol. 3, pp. 993–1022, 2003.
-  S. Kim, S. Narayanan, and S. Sundaram, “Acoustic topic model for audio information retrieval,” in Proc. of WASPAA, New Paltz NY, USA, 2009, pp. 37–40.
-  M. Doulaty, O. Saz, and T. Hain, “Unsupervised domain discovery using latent dirichlet allocation for acoustic modelling in speech recognition,” in Proc. of Interspeech, Dresden, Germany, 2015.
-  J. Sivic, B. C. Russell, A. A. Efros, A. Zisserman, and W. T. Freeman, “Discovering objects and their location in images,” in Proc. of ICCV, Beijing, China, 2005.
D. Hu and L. K. Saul,
“A probabilistic topic model for unsupervised learning of musical key-profiles.,”in Proc. of ISMIR, Kobe, Japan, 2009.
-  T. L. Griffiths and M. Steyvers, “Finding scientific topics,” Proc. of National Academy of Sciences of the United States of America, vol. 101, pp. 5228–5235, 2004.
-  A. Gersho and R. M. Gray, Vector quantization and signal compression, Springer Science & Business Media, Berlin, Germany, 1992.
-  C. Ni, C. C. Leung, L. Wang, N. F. Chen, and B. Ma, “Unsupervised data selection and word–morph mixed language model for tamil low-resource keyword search,” in Proc. of ICASSP, Brisbane, Australia, 2015.
-  P. Bell, M. J. F. Gales, T. Hain, J. Kilgour, P. Lanchantin, X. Liu, A. McParland, S. Renals, O. Saz, M. Webster, and P. Woodland, “The MGB Challenge: Evaluating multi-genre broadcast media recognition,” in Proc. of ASRU, Arizona, USA, 2015.
-  O. Saz, M. Doulaty, S. Deena, R. Milner, R. W. M. Ng, M. Hasan, Y. Liu, and T. Hain, “The 2015 Sheffield system for transcription of multi–genre broadcast media,” in Proc. of ASRU, Arizona, USA, 2015.
-  R. W. M. Ng, M. Doulaty, R. Doddipatla, O. Saz, M. Hasan, T. Hain, W. Aziz, K. Shaf, and L. Specia, “The USFD spoken language translation system for IWSLT 2014,” in Proc. of IWSLT, Lake Tahoe NV, USA, 2014.
-  Y. Long, M. J. F. Gales, P. Lanchantin, X. Liu, M. S. Seigel, and P. C. Woodland, “Improving lightly supervised training for broadcast transcriptions,” in Proc. of Interspeech, Lyon, France, 2013.
-  H. Hermansky, “Perceptual linear predictive (PLP) analysis of speech,” the Journal of the Acoustical Society of America, vol. 87, no. 4, pp. 1738–1752, 1990.
-  R. Haeb-Umbach and H. Ney, “Linear discriminant analysis for improved large vocabulary continuous speech recognition,” in Proc. of ICASSP, San Francisco, USA, 1992.
-  M. Gales, “Maximum likelihood linear transformations for HMM-based speech recognition,” Computer Speech & Language, vol. 12, no. 2, pp. 75 – 98, 1998.
-  T. Anastasakos, J. McDonough, R. Schwartz, and J. Makhoul, “A compact model for speaker-adaptive training,” in Proc. of ICSLP, Philadelphia, USA, 1996.
-  D. Povey, D. Kanevsky, B. Kingsbury, B. Ramabhadran, G. Saon, and K. Visweswariah, “Boosted MMI for model and feature-space discriminative training,” in Proc. of ICASSP, Las Vegas, USA, 2008.
-  G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for deep belief nets,” Neural Computation, vol. 18, no. 7, pp. 1527–1554, 2006.
-  D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motlicek, Y. Qian, P. Schwarz, J. Silovsky, G. Stemmer, and K. Vesely, “The Kaldi speech recognition toolkit,” in Proc. of ASRU, Hawaii, USA, 2011.
-  A. Stolcke, “Srilm-an extensible language modeling toolkit.,” in Proc. of Interspeech, Denver, US, 2002.
-  R. Rehurek and P. Sojka, “Software framework for topic modelling with large corpora,” in Proc. of LREC, Valletta, Malta, 2010.
-  O. Saz, M. Doulaty, and T. Hain, “Background–tracking acoustic features for genre identification of broadcast shows,” in Proc. of SLT, Lake Tahoe NV, USA, 2014.