Peter Beyerlein

Towards language independent acoustic modeling

We describe procedures and experimental results using speech from diverse source languages to build an ASR system for a single target language. This work is intended to improve ASR in languages for which large amounts of training data are not available. We have developed both knowledge-based and automatic methods to map phonetic units from the source languages to the target language. We employed HMM adaptation techniques and discriminative model combination to combine acoustic models from the individual source languages for recognition of speech in the target language. Experiments are described in which Czech Broadcast News is transcribed using acoustic models trained from small amounts of Czech read speech augmented by English, Spanish, Russian, and Mandarin acoustic models.

mRNA-seq with agnostic splice site discovery for nervous system transcriptomics tested in chronic pain.

mRNA-seq is a paradigm-shifting technology because of its superior sensitivity and dynamic range and its potential to capture transcriptomes in an agnostic fashion, i.e., independently of existing genome annotations. Implementation of the agnostic approach, however, has not yet been fully achieved. In particular, agnostic mapping of pre-mRNA splice sites has not been demonstrated. The present study pursued dual goals: (1) to advance mRNA-seq bioinformatics toward unbiased transcriptome capture and (2) to demonstrate its potential for discovery in neuroscience by applying the approach to an in vivo model of neurological disease. We have performed mRNA-seq on the L4 dorsal root ganglion (DRG) of rats with chronic neuropathic pain induced by spinal nerve ligation (SNL) of the neighboring (L5) spinal nerve. We found that 12.4% of known genes were induced and 7% were suppressed in the dysfunctional (but anatomically intact) L4 DRG 2 wk after SNL. These alterations persisted chronically (2 mo). Using a read cluster classifier with strong test characteristics (ROC area 97%), we discovered 10,464 novel exons. A new algorithm for agnostic mapping of pre-mRNA splice junctions (SJs) achieved a precision of 97%. Integration of information from all mRNA-seq read classes including SJs led to genome reannotations specifically relevant for the species used (rat), the anatomical site studied (DRG), and the neurological disease considered (pain); for example, a 64-exon coreceptor for the nociceptive transmitter substance P was identified, and 21.9% of newly discovered exons were shown to be dysregulated. Thus, mRNA-seq with agnostic analysis methods appears to provide a highly productive approach for in vivo transcriptomics in the nervous system.

Automatic transcription of unknown words in a speech recognition system

We address the problem of automatically finding an acoustic representation (i.e. a transcription) of unknown words as a sequence of subword units, given a few sample utterances of the unknown words, and an inventory of speaker-independent subword units. The problem arises if a user wants to add his own vocabulary to a speaker-independent recognition system simply by speaking the words a few times. Two methods are investigated which are both based on a maximum-likelihood formulation of the problem. The experimental results show that both automatic transcription methods provide a good estimate of the acoustic models of unknown words. The recognition error rates obtained with such models in a speaker-independent recognition task are clearly better than those resulting from separate whole-word models. They are comparable with the performance of transcriptions drawn from a dictionary.

Large vocabulary continuous speech recognition of Broadcast News - The Philips/RWTH approach

Abstract Automatic speech recognition of real-live broadcast news (BN) data (Hub-4) has become a challenging research topic in recent years. This paper summarizes our key efforts to build a large vocabulary continuous speech recognition system for the heterogenous BN task without inducing undesired complexity and computational resources. These key efforts included: • automatic segmentation of the audio signal into speech utterances; • efficient one-pass trigram decoding using look-ahead techniques; • optimal log-linear interpolation of a variety of acoustic and language models using discriminative model combination (DMC); • handling short-range and weak longer-range correlations in natural speech and language by the use of phrases and of distance-language models; • improving the acoustic modeling by a robust feature extraction, channel normalization, adaptation techniques as well as automatic script selection and verification. The starting point of the system development was the Philips 64k-NAB word-internal triphone trigram system. On the speaker-independent but microphone-dependent NAB-task (transcription of read newspaper texts) we obtained a word error rate of about 10%. Now, at the conclusion of the system development, we have arrived at Philips at an DMC-interpolated phrase-based crossword-pentaphone 4-gram system. This system transcribes BN data with an overall word error rate of about 17%.

Method for constructing a model of a new word for addition to a word model database of a speech recognition system

For speech recognition a new word is represented as based on a stored inventory of models of sub-word units. First a plurality of utterances are presented that all should conform to the word. For building a word model from the utterances, these are represented by a sequence of feature vectors. First, the utterances are used to train a whole-word model that is independent of the models of the sub-word units. The length of the whole-word model equals the average length of the utterances. Next, a sequence of Markov states and associated probability densities of acoustic events of the whole-word model is interpreted as a reference template represented by a string of averaged feature vectors. Finally, the string is recognized by matching to models in the inventory and storing a recognition result as a model of the utterances.

Targeted or whole genome sequencing of formalin fixed tissue samples: potential applications in cancer genomics.

Current genomic studies are limited by the poor availability of fresh-frozen tissue samples. Although formalin-fixed diagnostic samples are in abundance, they are seldom used in current genomic studies because of the concern of formalin-fixation artifacts. Better characterization of these artifacts will allow the use of archived clinical specimens in translational and clinical research studies. To provide a systematic analysis of formalin-fixation artifacts on Illumina sequencing, we generated 26 DNA sequencing data sets from 13 pairs of matched formalin-fixed paraffin-embedded (FFPE) and fresh-frozen (FF) tissue samples. The results indicate high rate of concordant calls between matched FF/FFPE pairs at reference and variant positions in three commonly used sequencing approaches (whole genome, whole exome, and targeted exon sequencing). Global mismatch rates and C·G > T·A substitutions were comparable between matched FF/FFPE samples, and discordant rates were low (<0.26%) in all samples. Finally, low-pass whole genome sequencing produces similar pattern of copy number alterations between FF/FFPE pairs. The results from our studies suggest the potential use of diagnostic FFPE samples for cancer genomic studies to characterize and catalog variations in cancer genomes.

Discriminative Generalized Hough transform for localization of joints in the lower extremities

A fully automatic iterative training approach for the generation of discriminative shape models for usage in the Generalized Hough Transform (GHT) is presented. The method aims at capturing the shape variability of the target object contained in the training data as well as identifying confusable structures (anti-shapes) and integrating this information into one model. To distinguish shape and anti-shape points and to determine their importance, an individual positive or negative weight is estimated for each model point by means of a discriminative training technique. The model is built from edge points surrounding the target point and the most confusable structure as identified by the GHT. Through an iterative approach, the performance of the model is gradually improved by extending the training dataset with images, where the current model failed to localize the target point. The proposed method is successfully tested on a set of 670 long-leg radiographs, where it achieves a localization rate of 74–97% for the respective tasks.

Application of clustering techniques to mixture density modelling for continuous-speech recognition

Clustering techniques have been integrated at different levels into the training procedure of a continuous-density hidden Markov model (HMM) speech recognizer. These clustering techniques can be used in two ways. First acoustically similar states are tied together. It will help to reduce the number of parameters but also allow to train otherwise rarely seen states together with more robust ones (state-tying). Secondly densities are clustered across states, this reduces the number of densities while at the same time keeping the best performances of our recognizer (density-clustering). We have applied these techniques both to word-based small-vocabulary and phoneme-based large-vocabulary recognition tasks. On the WSJ task, we could achieve a reduction of the word error rate by 7%. On the TI/NIST-connected digit task, the number of parameters was reduced by a factor 2-3 while keeping the same string error rate.

Robust gene expression and mutation analyses of RNA-sequencing of formalin-fixed diagnostic tumor samples

Current genomic studies are limited by the availability of fresh tissue samples. Here, we show that Illumina RNA sequencing of formalin-fixed diagnostic tumor samples produces gene expression that is strongly correlated with matched frozen tumor samples (r > 0.89). In addition, sequence variations identified from FFPE RNA show 99.67% concordance with that from exome sequencing of matched frozen tumor samples. Because FFPE is a routine diagnostic sample preparation, the feasibility results reported here will facilitate the setup of large-scale research and clinical studies in medical genomics that are currently limited by the availability of fresh frozen samples.

Speaker adaptation in the Philips system for large vocabulary continuous speech recognition

The combination of maximum likelihood linear regression (MLLR) with maximum a posteriori (MAP) adaptation has been investigated for both the enrollment of a new speaker as well as for the asymptotic recognition rate after several hours of dictation. We show that a least mean square approach to MLLR is quite effective in conjunction with phonetically derived regression classes. Results are presented for both ARPA read-speech test sets and real-life dictation. Significant improvements are reported. While MLLR achieves a faster adaptation rate when only few data is available, MAP has desirable asymptotic properties and the combination of both methods provides the best results. Both incremental and iterative batch modes are studied and compared to the performance of speaker-dependent training.