Refine
Year of publication
Document Type
- Conference Proceeding (14)
- Book (2)
- Contribution to a Periodical (2)
Language
- English (18) (remove)
Has Fulltext
- no (18)
Is part of the Bibliography
- no (18)
Keywords
- Speech Recognition (18) (remove)
Institute
We present a novel lecture browser that utilizes ranked key phrases displayed on a stream graph to overcome the shortcomings of traditional extractive (query-based) summaries. The system extracts key phrases from the ASR transcripts, performs an unsupervised ranking, and displays an initial number of phrases on the stream graph. This graph gives an intuition of when which key phrase is spoken, and how dominant it is throughout the lecture. The user can select the phrases to be displayed and furthermore adjust the ranking of the all phrases. All user interactions are logged to a server to improve the ranking algorithms and provide user specific rankings.
Voice scrambling is widely used to add privacy to the radio communication of various authorities - but is also used by criminals to evade prosecution. In this article, we consider various analog voice scrambling techniques such as fixed frequency inversion, splitband inversion and rolling code scramblers. We explain how to break them using automatically extracted measures and scoring algorithms, and evaluate the proposed system using simulated data. While the simple inversion can be easily broken, the more advanced techniques require additional work prior to unsupervised automatization; the presented user interface allows the user to refine the automatic results to obtain a high quality solution.
We describe a state-of-the-art large vocabulary continuous speech recognition (LVCSR) and keyword search (KWS) system trained on roughly 70 hours of conversational telephone speech. Using the Kaldi speech recognition toolkit, we investigate several aspects: for the acoustic front-end, we analyze the use of mel-frequency cepstral coefficients (MFCC), pitch and probability-of-voicing (PoV), and deep neural network (DNN) bottleneck (BN) features, as well as their feature-level combination ("tandem"). For the acousticphonetic decision tree, we explore different hidden Markov model (HMM) topologies for the glottalization phoneme /?/ to model its typically short duration. For the acoustic model, we compare regular continuous HMM with a sort of multi-codebook subspace Gaussian mixture model (SGMM) that lead to an overall best word error rate (WER) of 58.7% and 56.3%, respectively. The KWS is implemented as a word lattice search, and is augmented by a syllable lattice back-up search to capture out-of-vocabulary keywords as well as misrecognized lexical surface forms due to ambiguous prefix and hyphenation rules.
The tracheoesophageal (TE) substitute voice is currently state–of–the–art treatment to restore the ability to speak after laryngectomy. The intelligibility while talking over a telephone is an important clinical factor, as it is a crucial part of the patients’ social life. An objective way to rate the intelligibility of substitute voices when talking over a telephone is desirable to improve the post–laryngectomy speech therapy. An automatic speech recognition (ASR) system was applied to 41 high quality recordings of post–laryngectomy patients. The ASR system was trained with normal, non–pathologic speech. It yielded a word accuracy (WA) of 36.9%±18.0%; compared to the intelligibility rating of a group of human experts the ASR system had a correlation coefficient of -.88. After downsampling the 41 recordings to telephone quality, the ASR system reached a WA of 26.4%±13.9% leading to a correlation coefficient of -.80. These results confirm that an ASR system can be used for objective intelligibility rating over the telephone.
One aspect of voice and speech evaluation after laryngeal cancer is acoustic analysis. Perceptual evaluation by expert raters is a standard in the clinical environment for global criteria such as overall quality or intelligibility. So far, automatic approaches evaluate acoustic properties of pathologic voices based on voiced/unvoiced distinction and fundamental frequency analysis of sustained vowels. Because of the high amount of noisy components and the increasing aperiodicity of highly pathologic voices, a fully automatic analysis of fundamental frequency is difficult. We introduce a purely data-driven system for the acoustic analysis of pathologic voices based on recordings of a standard text.
Reverberation effects as observed by room microphones severely degrade the performance of automatic speech recognition systems. We investigate the use of dereverberation by spectral subtraction as proposed by Lebart and Boucher and introduce a simple approach to estimate the required decay parameter by clapping hands. Experiments on small vocabulary continuous speech recognition task on read speech show that using the calibrated dereverberation improves WER from 73.2 to 54.7 for the best microphone. In combination with system adaptation, the WER could be reduced to 28.2, which is only a 16% relative loss of performance comparison to using a headset instead of a room microphone.
We describe a lattice generation method that is exact, i.e. it satisfies all the natural properties we would want from a lattice of alternative transcriptions of an utterance. This method does not introduce substantial overhead above one-best decoding. Our method is most directly applicable when using WFST decoders where the WFST is “fully expanded”, i.e. where the arcs correspond to HMM transitions. It outputs lattices that include HMM-state-level alignments as well as word labels. The general idea is to create a state-level lattice during decoding, and to do a special form of determinization that retains only the best-scoring path for each word sequence. This special determinization algorithm is a solution to the following problem: Given a WFST A, compute a WFST B that, for each input-symbol-sequence of A, contains just the lowest-cost path through A.
Folks that have been here last winter prior to ASRU might be familiar with the title of that talk. But don't be misled, I'll have something new for you. In this talk, I will give an overview over the FAU Lecture Browser which I developed in the context of my thesis. I will start out with the description of a novel data set: The LME Lectures are a corpus of two series of graduate level computer science lectures with 18 recordings each. The courses cover topics in medical image processing and pattern analysis/machine learning. The roughly 40 hours of speech were manually transcribed, and one particular lecture was annotated with key phrases by five human raters. Using this data set, I trained three different speech recognizers using regular continuous, multi-codebook semi-continuous and subspace Gaussian mixture models, that show an error rate of about 10% WER. I will then briefly describe the key phrase extraction and automatic ranking, which was then compared against five raters on one lecture recording. Finally, I will talk about a little usability study where 10 students were asked to perform a certain task-- with and without the proposed lecture browser. Although the number of contestants is limited, the numbers are interesting: the users that had the interface could complete the tasks about 30% faster than the control group, while maintaining about the same accuracy.