Increase In Word Error Rate Research Articles

Identification of related languages from spoken data: Moving from off-line to on-line scenario

The accelerating flow of information we encounter around the world today makes many companies deploy speech recognition systems that, to an ever-growing extent, process data on-line rather than off-line. These systems, e.g., for real-time 24/7 broadcast transcription, often work with input-stream data containing utterances in more than one language. This multilingual data can correctly be transcribed in real-time only if the language used is identified with just a small latency for each input frame. For this purpose, a novel approach to on-line spoken language identification is proposed in this work. Its development is documented within a series of consecutive experiments starting in the off-line mode for 11 Slavic languages, going through artificially prepared multilingual data for the on-line scenario, and ending with real bilingual TV programs containing utterances in mutually similar Czech and Slovak. The resulting scheme that we propose operates frame-by-frame; it takes in a multilingual stream of speech frames and outputs a stream of the corresponding language labels. It utilizes a weighted finite-state transducer as a decoder, which smooths the output from a language classifier fed by multilingual and augmented bottleneck features. An essential factor from the accuracy point of view is that these features, as well as the classifier itself, are based on deep neural network architectures that allow the modeling of long-term time dependencies. The obtained results show that our scheme allows us to determine the language spoken in real-world bilingual TV shows with an average latency of around 2.5 seconds and with an increase in word error rate by a mere 2.9% over the reference 18.1% value yielded by using manually prepared language labels.

Fast Likelihood Computation in Speech Recognition using Matrices

Acoustic modeling using mixtures of multivariate Gaussians is the prevalent approach for many speech processing problems. Computing likelihoods against a large set of Gaussians is required as a part of many speech processing systems and it is the computationally dominant phase for Large Vocabulary Continuous Speech Recognition (LVCSR) systems. We express the likelihood computation as a multiplication of matrices representing augmented feature vectors and Gaussian parameters. The computational gain of this approach over traditional methods is by exploiting the structure of these matrices and efficient implementation of their multiplication. In particular, we explore direct low-rank approximation of the Gaussian parameter matrix and indirect derivation of low-rank factors of the Gaussian parameter matrix by optimum approximation of the likelihood matrix. We show that both the methods lead to similar speedups but the latter leads to far lesser impact on the recognition accuracy. Experiments on 1,138 work vocabulary RM1 task and 6,224 word vocabulary TIMIT task using Sphinx 3.7 system show that, for a typical case the matrix multiplication based approach leads to overall speedup of 46 % on RM1 task and 115 % for TIMIT task. Our low-rank approximation methods provide a way for trading off recognition accuracy for a further increase in computational performance extending overall speedups up to 61 % for RM1 and 119 % for TIMIT for an increase of word error rate (WER) from 3.2 to 3.5 % for RM1 and for no increase in WER for TIMIT. We also express pairwise Euclidean distance computation phase in Dynamic Time Warping (DTW) in terms of matrix multiplication leading to saving of approximately ${1} \over {3}$ of computational operations. In our experiments using efficient implementation of matrix multiplication, this leads to a speedup of 5.6 in computing the pairwise Euclidean distances and overall speedup up to 3.25 for DTW.

Hierarchical search for large-vocabulary conversational speech recognition: working toward a solution to the decoding problem

Large vocabulary continuous speech recognition (LVCSR) systems have advanced significantly due to the ability to handle extremely large problem spaces in fairly small amounts of memory. The article introduces the search problem, discusses in detail a typical implementation of a search engine, and demonstrates the efficacy of this approach on a range of problems. The approach presented is scalable across a wide range of applications. It is designed to address research needs, where a premium is placed on the flexibility of the system architecture, and the needs of application prototypes, which require near-real-time speed without a great sacrifice in word error rate (WER). One major area of focus for researchers is the development of real-time systems. With only minor degradations in performance (typically, no more than a 25% increase in WER), the systems described in this article can be transformed into systems that operate at 10/spl times/RT or less. There are four active areas of research related to this problem. First, more intelligent pruning algorithms that prune the search space more heavily are required. Look-ahead and N-best strategies at all levels of the system are key to achieving such large reductions in the search space. Second, multi-pass systems that perform a quick search using a simple system, and then rescore only the N-best resulting hypotheses using better models are very popular for real-time implementation. Third, since much of the computation in these systems is devoted to acoustic model processing, fast-matching strategies within the acoustic model are important. Finally, since Gaussian evaluation at each state in the system is a major consumer of CPU time, vector quantization-like approaches that enable one to compute only a small number of Gaussians per frame are proven to be successful. In some sense, the Viterbi (1967) based system presented represents only one path through this continuum of recognition search strategies.