Multi-resolution speech analysis for automatic speech recognition using deep neural networks: Experiments on TIMIT.

Speech Analysis for Automatic Speech Recognition (ASR) systems typically starts with a Short-Time Fourier Transform (STFT) that implies selecting a fixed point in the time-frequency resolution trade-off. This approach, combined with a Mel-frequency scaled filterbank and a Discrete Cosine Transform give rise to the Mel-Frequency Cepstral Coefficients (MFCC), which have been the most common speech features in speech processing for the last decades. These features were particularly well suited for the previous Hidden Markov Models/Gaussian Mixture Models (HMM/GMM) state of the art in ASR. In particular they produced highly uncorrelated features of small dimensionality (typically 13 coefficients plus deltas and double deltas), which was very convenient for diagonal covariance GMMs, for dealing with the curse of dimensionality and for the limited computing resources of a decade ago. Currently most ASR systems use Deep Neural Networks (DNN) instead of the GMMs for modeling the acoustic features, which provides more flexibility regarding the definition of the features. In particular, acoustic features can be highly correlated and can be much larger in size because the DNNs are very powerful at processing high-dimensionality inputs. Also, the computing hardware has reached a level of evolution that makes computational cost in speech processing a less relevant issue. In this context we have decided to revisit the problem of the time-frequency resolution in speech analysis, and in particular to check if multi-resolution speech analysis (both in time and frequency) can be helpful in improving acoustic modeling using DNNs. Our experiments start with several Kaldi baseline system for the well known TIMIT corpus and modify them by adding multi-resolution speech representations by concatenating different spectra computed using different time-frequency resolutions and different post-processed and speaker-adapted features using different time-frequency resolutions. Our experiments show that using a multi-resolution speech representation tends to improve over results using the baseline single resolution speech representation, which seems to confirm our main hypothesis. However, results combining multi-resolution with the highly post-processed and speaker-adapted features, which provide the best results in Kaldi for TIMIT, yield only very modest improvements.

An analysis of the influence of deep neural network (DNN) topology in bottleneck feature based language recognition.

Language recognition systems based on bottleneck features have recently become the state-of-the-art in this research field, showing its success in the last Language Recognition Evaluation (LRE 2015) organized by NIST (U.S. National Institute of Standards and Technology). This type of system is based on a deep neural network (DNN) trained to discriminate between phonetic units, i.e. trained for the task of automatic speech recognition (ASR). This DNN aims to compress information in one of its layers, known as bottleneck (BN) layer, which is used to obtain a new frame representation of the audio signal. This representation has been proven to be useful for the task of language identification (LID). Thus, bottleneck features are used as input to the language recognition system, instead of a classical parameterization of the signal based on cepstral feature vectors such as MFCCs (Mel Frequency Cepstral Coefficients). Despite the success of this approach in language recognition, there is a lack of studies analyzing in a systematic way how the topology of the DNN influences the performance of bottleneck feature-based language recognition systems. In this work, we try to fill-in this gap, analyzing language recognition results with different topologies for the DNN used to extract the bottleneck features, comparing them and against a reference system based on a more classical cepstral representation of the input signal with a total variability model. This way, we obtain useful knowledge about how the DNN configuration influences bottleneck feature-based language recognition systems performance.

Reviewing the connection between speech and obstructive sleep apnea.

BACKGROUND: Sleep apnea (OSA) is a common sleep disorder characterized by recurring breathing pauses during sleep caused by a blockage of the upper airway (UA). The altered UA structure or function in OSA speakers has led to hypothesize the automatic analysis of speech for OSA assessment. In this paper we critically review several approaches using speech analysis and machine learning techniques for OSA detection, and discuss the limitations that can arise when using machine learning techniques for diagnostic applications. METHODS: A large speech database including 426 male Spanish speakers suspected to suffer OSA and derived to a sleep disorders unit was used to study the clinical validity of several proposals using machine learning techniques to predict the apnea-hypopnea index (AHI) or classify individuals according to their OSA severity. AHI describes the severity of patients' condition. We first evaluate AHI prediction using state-of-the-art speaker recognition technologies: speech spectral information is modelled using supervectors or i-vectors techniques, and AHI is predicted through support vector regression (SVR). Using the same database we then critically review several OSA classification approaches previously proposed. The influence and possible interference of other clinical variables or characteristics available for our OSA population: age, height, weight, body mass index, and cervical perimeter, are also studied. RESULTS: The poor results obtained when estimating AHI using supervectors or i-vectors followed by SVR contrast with the positive results reported by previous research. This fact prompted us to a careful review of these approaches, also testing some reported results over our database. Several methodological limitations and deficiencies were detected that may have led to overoptimistic results. CONCLUSION: The methodological deficiencies observed after critically reviewing previous research can be relevant examples of potential pitfalls when using machine learning techniques for diagnostic applications. We have found two common limitations that can explain the likelihood of false discovery in previous research: (1) the use of prediction models derived from sources, such as speech, which are also correlated with other patient characteristics (age, height, sex,…) that act as confounding factors; and (2) overfitting of feature selection and validation methods when working with a high number of variables compared to the number of cases. We hope this study could not only be a useful example of relevant issues when using machine learning for medical diagnosis, but it will also help in guiding further research on the connection between speech and OSA.

Language Identification in Short Utterances Using Long Short-Term Memory (LSTM) Recurrent Neural Networks.

Long Short Term Memory (LSTM) Recurrent Neural Networks (RNNs) have recently outperformed other state-of-the-art approaches, such as i-vector and Deep Neural Networks (DNNs), in automatic Language Identification (LID), particularly when dealing with very short utterances (∼3s). In this contribution we present an open-source, end-to-end, LSTM RNN system running on limited computational resources (a single GPU) that outperforms a reference i-vector system on a subset of the NIST Language Recognition Evaluation (8 target languages, 3s task) by up to a 26%. This result is in line with previously published research using proprietary LSTM implementations and huge computational resources, which made these former results hardly reproducible. Further, we extend those previous experiments modeling unseen languages (out of set, OOS, modeling), which is crucial in real applications. Results show that a LSTM RNN with OOS modeling is able to detect these languages and generalizes robustly to unseen OOS languages. Finally, we also analyze the effect of even more limited test data (from 2.25s to 0.1s) proving that with as little as 0.5s an accuracy of over 50% can be achieved.

Speech Signal and Facial Image Processing for Obstructive Sleep Apnea Assessment.

Obstructive sleep apnea (OSA) is a common sleep disorder characterized by recurring breathing pauses during sleep caused by a blockage of the upper airway (UA). OSA is generally diagnosed through a costly procedure requiring an overnight stay of the patient at the hospital. This has led to proposing less costly procedures based on the analysis of patients' facial images and voice recordings to help in OSA detection and severity assessment. In this paper we investigate the use of both image and speech processing to estimate the apnea-hypopnea index, AHI (which describes the severity of the condition), over a population of 285 male Spanish subjects suspected to suffer from OSA and referred to a Sleep Disorders Unit. Photographs and voice recordings were collected in a supervised but not highly controlled way trying to test a scenario close to an OSA assessment application running on a mobile device (i.e., smartphones or tablets). Spectral information in speech utterances is modeled by a state-of-the-art low-dimensional acoustic representation, called i-vector. A set of local craniofacial features related to OSA are extracted from images after detecting facial landmarks using Active Appearance Models (AAMs). Support vector regression (SVR) is applied on facial features and i-vectors to estimate the AHI.