State of the art speech recognizers use spectral features for the recognition. For improving of performance, especially in cars, the use of syllable durations is investigated: in experiments a 13.3% reduction of word error rate (WER) was achieved on a Chinese Mandarin continuous digits recognition in car environment. The explored approach was realized in two stages. First, the system performed standard speech recognition using acoustic spectral features. As a result, an n-best list of hypotheses was generated. In the second stage the hypothesis probabilities were re-estimated using syllable duration information, thus, the hypotheses were reordered such that the correct ones were pushed to the top of the nbest list. In such a way the word error rate (WER) was reduced. The syllable durations were normalized in order to eliminate the influence of speech rate variations, two different speech rate computation approaches were investigated. Introduction Reliable speech recognition in cars is a challenging task. Modern speech controlled systems perform well in quiet environment, but motor and street noise strongly reduce their accuracy. Noise reduction algorithms solve this problem, yet, the recognition quality does not satisfy user demands for car applications. Higher speech recognition quality may be achieved in another way: some other information sources such as several audio channels from microphone arrays, video stream used for lip reading or implementation of additional features may supplement commonly used audio spectral features. Typically, suprasegmental features (durations, pitch contours, energy) with lengths from one half to several seconds are ignored in common recognition systems: the speech signal is considered as a sequence of short term frames of 10-100 ms. This paper considers the application of duration models for improved continuous digit recognition in Chinese Mandarin. The normalization of durations to the syllable rate by means of the syllable center detector is another main point. In experiments, samples from Mandarin SPEECON database were recognized using hidden Markov models (HMMs) designed for embedded devices, thus, the recognition quality in cars was evaluated. Duration models and speech rate The speech recognition result is an n-best list of hypotheses where each entry is supplied with respective word segmentation and hypothesis probabilities. From the segmentation the duration of syllables are obtained. Afterwards, the combination of the probabilities from the duration models and HMMs results in a new order of best hypotheses (rescoring). Because of the different speech rates of spoken utterances the desired precision of word duration models may not be reached. Normalization of the word durations to the speech rate may increase model accuracy. But first, the confusing term “speech rate” has to be clarified, as it has several interpretations in literature. The next two sections deal with two most often observed definitions of speech rate. Relative speech rate The relative speech rate reflects how fast the utterance is pronounced in comparison to the “average” speech rate of the “average” person [4]. The equation below shows the estimation of the relative speech rate in a sentence:

This paper describes the ECESS evaluation campaign of voice activity and voicing detection. Standard VAD classifies signal into speech and non-speech, we extend it to VAD+ so that it classifies a signal as a sequence of non-speech, voiced and unvoiced segments. The evaluation is performed on a portion of the Spanish SPEECON database with manually labeled segmentation. To avoid errors caused by the limited precision of manual labeling we introduce “dead zones” –tolerance intervals ±5 ms around label changes in the data set. In these tolerance intervals we don't evaluate the signal.

This paper describes practical approaches for improving Mandarin digit recognition accuracy, especially in cars. We consider syllable and subword unit durations as additional source of information. The explored approach was realized in two stages. First, the system performs standard speech recognition using acoustic spectral features. As a result, an n-best list of hypotheses is generated. In the second stage the hypothesis probabilities are re-estimated using duration models, thus, the hypotheses are reordered such that the correct ones are pushed to the top of the n-best list. In such a way the word error rate (WER) is reduced.We explore state of the art approach of duration n-grams. In order to eliminate the influence of speech rate variations, the durations are normalized to a relative speech rate, a 10% relative reduction of WER was achieved. A novel approach led to 13.3% WER reduction: the durations were normalized to a syllable rate obtained from the syllable center detector.

Speech recognition on embedded systems requires components of low memory footprint and low computational complexity. In this paper a POS-based (part of speech based) language modeling approach is presented which decreases the number of language model parameters combined with a method for reducing memory consumptions via quantization of language model penalties. For the application of short message dictation a language model with about 10,000 words of vocabulary is generated. Using the POS-based language modeling approach the number of parameters comprises 70,058 penalties. The memory consumptions for storing those penalties are reduced about 50% using the presented coding method. Experiments show that the POS-based language model is able to reduce the WER up to 65% for n-best isolated word recognition in comparison to the case without language model. Moreover the increase of WER caused by coding of the language model penalties is not significant.

This paper presents the Siemens speech recognizer for mobile phones, VSR. VSR employs HMM technology and uses general-purpose phoneme-based acoustic models which make it speaker and vocabulary independent. The system can be easily reconfigured to work with arbitrary vocabularies. This provides full flexibility for the design of the user interface which contrasts with the capabilities of other low-resource recognizers. The system requirements of VSR are very low. The emission probability calculation and the Viterbi search with a vocabulary of 30 words need only 16 MHz for real-time operation on an ARM microcontroller. The HMM acoustic models take up about 12 kilobytes of permanent storage. The most significant algorithmic improvement is the newly developed 3-D stream-based coding of the HMMs. Despite low requirements in terms of system resources VSR achieves an outstanding recognition performance. The word error rate (WER) for a recognition task with 62 German isolated words including highly confusable digits is 7.0%.

In this paper the implementation of a word-stem based tree search for large vocabulary speaker independent isolated word recognition for embedded systems is presented. Two fast search algorithms combine the effectiveness of the tree structure for large vocabularies and the fast Viterbi search within the regular structures of word-stems. The algorithms are proved to be very effective for workstation and embedded platform realizations. In order to decrease the processing power the word-stem based tree search with frame dropping approach is used. The recognition speed was increased by a factor of 5 without frame dropping and by a factor of 10 with frame dropping in comparison to linear Viterbi search for isolated word recognition task with a vocabulary of 20102 words. Thus, the large vocabulary isolated word recognition becomes possible for embedded systems.

In order to make hidden Markov model (HMM) speech recognition suitable for mobile phone applications, Siemens developed a recognizer, Very Smart Recognizer (VSR), for deployment in future mobile phone generations. Typical applications will be name dialling, command and control operations suited for different environments, for example in cars. The paper describes research and development issues of a speech recognizer in mobile devices focusing on noise robustness, memory efficiency and integer implementation. The VSR is shown to reach a word error rate as low as 4.1% on continuous digits recorded in a car environment. Furthermore by means of discriminative training and HMM-parameter coding, the memory requirements of the VSR HMMs are smaller than 64 kBytes.

Low-cost recognition systems based on hidden Markov models (HMM) for mobile speech recognizers (mobile phones, PDAs) have a limited quantity of memory and processing power. Furthermore, the resources have to be shared between several applications. In this paper memory efficient HMMs were investigated for low-cost recognition platforms. The feature parameter tying HMM and subspace distribution clustering HMM (SDCHMM) were explored. In order to achieve less memory requirements, a shared codebook approach for feature parameter tying HMM and SDCHMM was developed and its effectiveness was experimentally proved. It was shown that this approach leads to a relative increase of word error rate of less than 10% for 50% of memory reduction.