We present a bang-bang phase-locked loop (PLL) generator that encapsulates design methodologies for its circuit blocks and the complete PLL system. The generator is fully automated and parameterized, producing the layout and schematic based on process characterization and top-level specifications. Three 14GHz PLLs are instantiated in TSMC 16nm, GF 14nm and Intel 22nm technologies, demonstrating the process portability. The rapid generation time of less than four days enables fast PLL design and technology porting. The PLL design fabricated in TSMC 16nm shows RMS jitter of 565.4fs and power of 6.64mW from a 0.9V supply.

LAYout with Gridded Objects (LAYGO), a Python-based layout-generation engine for enhancing the design productivity of custom circuit layouts in advanced CMOS processes, is presented and verified by implementing a time-interleaved SAR (TI-SAR) ADC instance in a 16 nm CMOS FinFET technology. LAYGO supports rapid generation by placing customized templates on process-specific placement grids, thereby encapsulating the design rules and process-specific structures. The templates can be located based on their relative positional information, which further enhances the description capability and portability. Interconnecting wires are routed on the grids for design rule abstractions, with additional customizations and support for multi-patterning. The functions for the on-grid placement and routing use advanced indexing and slicing with multi-dimensional object containers to improve the description and parameterization capabilities. Multiple TI-SAR ADC layouts are generated using LAYGO in 28-16 nm CMOS technologies. One instance is fabricated in a 16 nm CMOS FinFET process and measured, achieving a 38.2 dB signal-to-noise-and-distortion ratio (SNDR) at 7 GS/s after digital calibration and consuming 45.2 mW. Owing to its high customization capability, the design achieved the highest sampling rate (7 GS/s) among the generated ADCs.

Abstract We focus on the all‐pairs minimum cut (APMC) problem, a graph partitioning problem whose solution requires finding the minimum cut for every pair of nodes in a given graph. While it is solved for undirected graphs, a solution for APMC in directed graphs still requires an O(n2) brute force approach. We show that the empirical number of distinct minimum cuts in randomly generated strongly connected directed graphs is proportional to n rather than the theoretical value of n2, suggesting the possibility of an algorithm which finds all minimum cuts in less than O(n2) time. We also provide an example of the strict upper bound on the number of cuts in graphs with three nodes. We model the distributions with the Generalized extreme value (GEV) distribution and enable the possibility of using a GEV distribution to predict the probability of achieving a certain number of minimum cuts, given the number of nodes and edges. Finally, we contribute to the notion of symmetric cuts by showing that there can be O(n2) symmetric cuts in graphs when node replication is allowed.

In this paper, we forecast the reading of an air quality monitoring station over the next 48 hours, using a data-driven method that considers current meteorological data, weather forecasts, and air quality data of the station and that of other stations within a few hundred kilometers. Our predictive model is comprised of four major components: 1) a linear regression-based temporal predictor to model the local factors of air quality, 2) a neural network-based spatial predictor to model global factors, 3) a dynamic aggregator combining the predictions of the spatial and temporal predictors according to meteorological data, and 4) an inflection predictor to capture sudden changes in air quality. We evaluate our model with data from 43 cities in China, surpassing the results of multiple baseline methods. We have deployed a system with the Chinese Ministry of Environmental Protection, providing 48-hour fine-grained air quality forecasts for four major Chinese cities every hour. The forecast function is also enabled on Microsoft Bing Map and MS cloud platform Azure. Our technology is general and can be applied globally for other cities.

Many developing countries are suffering from air pollution, especially the Particulate Matter with diameter of 2.5 micrometers or less (PM2.5). While quite a few air quality monitoring stations have been built by governments in a city's public areas, the indoor PM2.5 has not yet been monitored and dealt with effectively. Though many office buildings have an HVAC (heating, ventilation, and air conditioning) system, PM2.5 is not considered as a factor when the system circulates fresh air from outdoors. This paper introduces a real system that we have deployed in the offices of four Microsoft campuses in China. This system instantly monitors indoor air quality on different floors of a building (including office areas, gyms, garages, and restaurants), enabling Microsoft employees to enquire the air quality of a place by using a mobile phone or checking a website. The information can guide a user's decision making, e.g., finding the right time to work out in the gym or turn on individual air filters in her own office. Through analyzing the indoor and outdoor air quality data collected over a long period, our system can even offer actionable and energy-efficient suggestion to HVAC systems, e.g., automatically turning on the system only a few hours earlier than usual if it is a heavily polluted day, or identifying the filters in HVAC system that should be renewed.

Many cities suffer from noise pollution, which compromises people's working efficiency and even mental health. New York City (NYC) has opened a platform, entitled 311, to allow people to complain about the city's issues by using a mobile app or making a phone call; noise is the third largest category of complaints in the 311 data. As each complaint about noises is associated with a location, a time stamp, and a fine-grained noise category, such as "Loud Music" or "Construction", the data is actually a result of "human as a sensor" and "crowd sensing", containing rich human intelligence that can help diagnose urban noises. In this paper we infer the fine-grained noise situation (consisting of a noise pollution indicator and the composition of noises) of different times of day for each region of NYC, by using the 311 complaint data together with social media, road network data, and Points of Interests (POIs). We model the noise situation of NYC with a three dimension tensor, where the three dimensions stand for regions, noise categories, and time slots, respectively. Supplementing the missing entries of the tensor through a context-aware tensor decomposition approach, we recover the noise situation throughout NYC. The information can inform people and officials' decision making. We evaluate our method with four real datasets, verifying the advantages of our method beyond four baselines, such as the interpolation-based approach.

This paper instantly infers the gas consumption and pollution emission of vehicles traveling on a city's road network in a current time slot, using GPS trajectories from a sample of vehicles (e.g., taxicabs). The knowledge can be used to suggest cost-efficient driving routes as well as identifying road segments where gas has been wasted significantly. The instant estimation of the emissions from vehicles can enable pollution alerts and help diagnose the root cause of air pollution in the long run. In our method, we first compute the travel speed of each road segment using the GPS trajectories received recently. As many road segments are not traversed by trajectories (i.e., data sparsity), we propose a Travel Speed Estimation (TSE) model based on a context-aware matrix factorization approach. TSE leverages features learned from other data sources, e.g., map data and historical trajectories, to deal with the data sparsity problem. We then propose a Traffic Volume Inference (TVI) model to infer the number of vehicles passing each road segment per minute. TVI is an unsupervised Bayesian Network that incorporates multiple factors, such as travel speed, weather conditions and geographical features of a road. Given the travel speed and traffic volume of a road segment, gas consumption and emissions can be calculated based on existing environmental theories. We evaluate our method based on extensive experiments using GPS trajectories generated by over 32,000 taxis in Beijing over a period of two months. The results demonstrate the advantages of our method over baselines, validating the contribution of its components and finding interesting discoveries for the benefit of society.

Online shoppers are generally highly task-driven: they have a certain goal in mind, and they are looking for a product with features that are consistent with that goal. Unfortunately, finding a product with specific features is extremely time-consuming using the search functionality provided by existing web sites.In this paper, we present a new search system called Red Opal that enables users to locate products rapidly based on features. Our fully automatic system examines prior customer reviews, identifies product features, and scores each product on each feature. Red Opal uses these scores to determine which products to show when a user specifies a desired product feature. We evaluate our system on four dimensions: precision of feature extraction, efficiency of feature extraction, precision of product scores, and estimated time savings to customers. On each dimension, Red Opal performs better than a comparison system.

Modeling units play a very important role in state-of-art speech recognition systems. The design and selection of them will directly impact the performance of the final speech recognition engine. As a tonal language, Mandarin's modeling units are more special for the tonal processing. In this paper, after fully investigating several dominant modeling strategies, we propose a new phone set design strategy for Mandarin, called segmental tonal modeling. Instead of modeling tone types directly, we realize them implicitly and jointly by two segments, which both carry tonal information. Both HTK and SAPI based experiments confirmed that such a method is very efficient. In addition to improving the accuracy by 9-23%, it greatly reduces the decoding time by 30-45%. Given the similar decoding speed, the new phone set configuration can reduce the error rate by relatively 35%.

For efficient organization of speech recordings - meetings, interviews, voice mails, lectures - the ability to search for spoken keywords is an essential capability. Today, most spoken-document retrieval systems use large-vocabulary recognition. For the above scenarios, such systems suffer from both the unpredictable vocabulary/domain and generally high word-error rates (WER). We present a vocabulary-independent system to index and to search rapidly spontaneous speech. A speech recognizer generates lattices of phonetic word fragments, against which keywords are matched phonetically. We first show the need to use recognition alternatives (lattices) in a high-WER context, on a word-based baseline. Then we introduce our new method of phonetic word-fragment lattice generation, which uses longer-span language knowledge than a phoneme recognizer. Last we introduce heuristics to compact the lattices to feasible sizes that can be searched efficiently. On the LDC voice mail corpus, we show that vocabulary/domain-independent phonetic search is as accurate as a vocabulary/domain-dependent word-lattice based baseline system for in-vocabulary keywords (FOMs of 74-75%), but nearly maintains this accuracy also for out-of-vocabulary keywords.

Over the past several years, the primary focus for the speech-recognition research community has been speaker-independent speech recognition, with the emphasis of working on databases with larger and larger numbers of speakers. For example, the most recent EARS program, which is sponsored by DARPA, calls for recordings of thousands of speakers. However, we are interested in making a speech interface work well for one particular individual, and we propose using massive amounts of speaker-specific training data recorded in daily life. We call this massively speaker-specific recognition (MSSR). As a pre-research, we leverage the large corpus we have available from speech-synthesis work to study the benefit of MSSR only from the acoustic-modeling aspect. Initial results show that, by changing the focus to MSSR, word error rates can drop very significantly. In comparison with speaker-adaptive speech recognition systems, MSSR also performs better since model parameters can be tuned to be suitable to one particular individual.

Different feature extraction and tone modeling schemes are investigated on both speaker-dependent and speaker-independent continuous speech databases. Tone recognition features can be classified as detailed features which use the entire F0 curve, and outlined features which capture the main structure of the F0 curve. Tone models of different size, ranging from very simple one-tone-one-model tone models to complex phoneme-dependent tone models, have different abilities to characterize tone. Our experiments show two conclusions. First, the detailed information of the F0 curve is not necessary for tone recognition. The outlined features can, not only reduce the number of parameters, but also improve the accuracy of tone recognition. The proposed subsection average F0 and /spl Delta/F0 are shown to be effective outlined features. The second conclusion is that the one-tone-one-model scheme is not sufficient. Building phoneme-dependent tone models can highly improve the recognition accuracy, especially for speaker-independent data. Thus we suggest using fractionized models, trained with the outlined features, for tone recognition.

Tone modeling is an unavoidable problem in Mandarin speech recognition. In continuous speech, the pitch contour exhibits variable patterns, and it is strongly influenced by its tone context. Although several effective methods have been proposed to improve the accuracy for tonal syllables in Mandarin continuous speech recognition, many recognition errors are caused by poor tone discrimination capability of the acoustic model. Furthermore, the case becomes worse for the recognition of spontaneous speech. In this paper, we report our work on tone articulation modeling. Tone context dependent models are used to model unstable pitch patterns caused by co-articulation in continuous speech. Corresponding acoustic features are investigated as well. Our methods are evaluated on two test sets: one is reading-style speech data, the other is spontaneous. The experimental results show that for the test set of casual speech, the proposed method turns out to be more effective than tone context independent model, while they are comparable for the test set of reading-style speech. Several factors which have potential to improve the proposed method are discussed in the final part in this paper.

This paper describes a bilingual text-to-speech (TTS) system, Microsoft Mulan, which switches between Mandarin and English smoothly and which maintains the sentence level intonation even for mixed-lingual texts. Mulan is constructed on the basis of the Soft Prediction Only prosodic strategy and the Prosodic-Constraint Orient unit-selection strategy. The unit-selection module of Mulan is shared across languages. It is insensitive to language identity, even though the syllable is used as the smallest unit in Mandarin, and the phoneme in English. Mulan has a unique module, the language-dispatching module, which dispatches texts to the language-specific front-ends and merges the outputs of the two front-ends together. The mixed texts are "uttered" out with the same voice. According to our informal listening test, the speech synthesized with Mulan sounds quite natural. Sample waves can be heard at: http://research.microsoft.com/-echang/proiects/tts/mulan.htm.

The maximum likelihood estimation (MLE) and Bayesian maximum a-posteriori (MAP) adaptation methods for Gaussian mixture models (GMM) have proven to be effective and efficient for speaker verification, even though each speaker model is trained using only his own training utterances. Discriminative criteria aim at increasing discriminability by using out-of-class data. In this paper, we consider the speaker verification task using three discriminative training methods to compare performance. Comparisons are discussed for the maximum mutual information (MMI), minimum classification error (MCE) and figure of merit (FOM) criteria. Experiments on the 1996 NIST speaker recognition evaluation data set show that FOM training method outperforms the other two methods for speaker verification in terms of system performance. Meanwhile, logistic regression is investigated and successfully employed as a discriminative score-normalization technique.

The paper presents a particle-filtering method for estimating formant frequencies of speech signals from spectrograms. First, frequency bands corresponding to the analyzed formants are extracted via a two-step dynamic programming based algorithm. A particle-filtering method is then used to locate accurately formants in every formant area based on the posterior PDF described by a set of support points with associated weights. Formant trajectories of voiced frames of a group of 81 utterances were manually tracked and labeled, partly for model training and partly for algorithm evaluation. In the experiments, the proposed method obtains average estimation errors of 72, 115, and 113 Hz for the first three formants, respectively, whereas the LPC based method induces 118, 172, and 250 Hz deviations. The experimental results show that the formants estimated by the proposed method are quite reliable and the trajectories are more accurate than LPC.

The Gaussian mixture models (GMM) has proved to be an effective probabilistic model for speaker verification, and has been widely used in most of state-of-the-art systems. In this paper, we introduce a new method for the task: that using AdaBoost learning based on the GMM. The motivation is the following: While a GMM linearly combines a number of Gaussian models according to a set of mixing weights, we believe that there exists a better means of combining individual Gaussian mixture models. The proposed AdaBoost-GMM method is non-parametric in which a selected set of weak classifiers, each constructed based on a single Gaussian model, is optimally combined to form a strong classifier, the optimality being in the sense of maximum margin. Experiments show that the boosted GMM classifier yields 10.81% relative reduction in equal error rate for the same handsets and 11.24% for different handsets, a significant improvement over the baseline adapted GMM system.

This paper puts forward a domain adaptation problem that has not been studied well. For corpus-driven TTS systems, domain adaptation is realized by adding a small amount of domain-specific speech that will provide the maximum increase in average length of units that are used for synthesizing speech in that domain. An approach for generating optimized script for adaptation is proposed, the core of which is a dynamic programming based algorithm that segments domain-specific corpus into minimum number of segments that appear in the unit inventory. Increase in MOS after adaptation can be estimated from the generated script without recording speech from it. The results show that the amount of MOS increase depends not only on the size of the training set and the size of the script for adaptation, but also on the broadness of the domain. Narrower domains have larger increase in MOS.

Acoustic variability across speakers is one of the challenges of speaker independent (SI) speech recognition systems. As a powerful solution, dominant speaker adaptation technologies such as MLLR and MAP may become inefficient because of the lack of enough enrollment data. In this paper, we propose an adaptation method based on speaker selection training, which makes full use of statistics of training corpus. Relative error rate reduction of 5.31 % is achieved when only one utterance is available. We compare different speaker selection strategies, namely. PCA, HMM and GMM based methods. In addition, impacts of number of selected cohort speakers and number of utterances from target speaker are investigated. Furthermore, comparison and integration with MLLR adaptation are also shown. Finally, some ongoing work such as dynamicalJy varying number of selected speakers, measuring the relative contribution among the selected speakers and speeding up the computationally expensive procedure of re-estimation with model synthesis are also discussed.

A novel discriminative training method of Gaussian mixture model for text-independent speaker verification, Figure of Merit (FOM) training, is proposed in this paper. FOM training aims at maximizing the FOM of a ROC curve by adjusting the model parameters, rather than only approximating the underlying distribution of acoustic observations of each speaker that Maximum Likelihood Estimation does. The text-independent speaker verification experiments were conducted on the 1996 NIST Speaker Recognition Evaluation corpus. Compared with standard EM training method, FOM training provides significantly improved performance, e.g. the detection cost function (DCF) was reduced to 0.0286 from 0.0369 and to 0.0537 from 0.0826 in matched and mismatched conditions respectively.

Taiwan Semiconductor Manufacturing Company launched the first 300 mm fully automation foundry in 1999. The newly introduced technology will automate the 300 mm fab for better manufacturing control and less human labor. Based on the accumulated TSMC data from year 1999 to year 2000, it is found that to maintain is more difficult than to build up the system. It is a great challenge to integrate so many hardware and software components closely. In fact, few 300 mm foundries are successful from the maintenance point of view. TSMC is still continuing to improve the full fab automation by strict integration test methodology and refer to daily problem analysis system. TSMC enlarges the acceptance test scope to cover the process equipment, transportation, material control and manufacturing system. In order to make sure the integration is smooth and on the right track, TSMC has developed a methodology to execute hardware and software integration. Here TSMC exposes some of the critical issues and shares the experience for unmanned factory. TSMC believes these can help the 300 mm fab to speed up the system construction task and find out their own problems by TSMC's methodology and experience.

This paper proposes a channel equalization algorithm for a large speech database with application in concatenative TTS systems. The convolutional channel distortion is equalized by comparing the power spectral densities (PSDs) of utterances of different recording sessions. Autoregressive linear filters are designed on a corpus level and are used offline to filter the corresponding sentences to compensate for the relative distortions caused by the channel effects. Two experiments are carried out to evaluate the benefit of the channel equalization approach. First, this method is used to reduce the distance of their PSDs between two recording sessions to verify the effectiveness of the method. Secondly, it is applied practically in the TTS system. The whole TTS speech database is processed to reduce the PSDs variance over all sessions. Moreover, a subjective listening test is carried out to obtain human evaluation of the new TTS system. Almost all listeners prefer the synthetic speech generated by the new TTS system. Furthermore, an analysis of variance (ANOVA) on this subjective listening test demonstrates that the channel equalization process has significant effect on increasing the perceived voice-quality consistency of the TTS system.

With the proliferation of handheld devices, information access on mobile devices is a topic of growing relevance. This paper presents a system that allows the user to search for information on mobile devices using spoken natural-language queries. We explore several issues related to the creation of this system, which combines state-of-the-art speech-recognition and information-retrieval technologies. This is the first work that we are aware of which evaluates spoken query based information retrieval on a commonly available and well researched text database, the Chinese news corpus used in the National Institute of Standards and Technology (NIST)s TREC-5 and TREC-6 benchmarks. To compare spoken-query retrieval performance for different relevant scenarios and recognition accuracies, the benchmark queries-read verbatim by 20 speakers-were recorded simultaneously through three channels: headset microphone, PDA microphone, and cellular phone. Our results show that for mobile devices with high-quality microphones, spoken-query retrieval based on existing technologies yields retrieval precisions that come close to that for perfect text input (mean average precision 0.459 and 0.489, respectively, on TREC-6).

This paper proposes a two-module text to speech system (TTS) structure, which bypasses the prosody model that predicts numerical prosodic parameters for synthetic speech. Instead, many instances of each basic unit from a large speech corpus are classified into categories by a classification and regression tree (CART), in which the expectation of the weighted sum of square regression error of prosodic features is used as splitting criterion. Better prosody is achieved by keeping slender diversity in prosodic features of instances belonging to the same class. A multi-tier non-uniform unit selection method is presented. It makes the best decision on unit selection by minimizing the concatenated cost of a whole utterance. Since the largest available and suitable units are selected for concatenating, distortion caused by mismatches at concatenated points is minimized. Very natural and fluent speech is synthesized, according to informal listening test.

The necessity of gathering data has been an impediment for researchers and students who are interested in getting started in the fields related to speech recognition. We are proposing a new approach of distributing data that is designed to quickly help researchers and students achieve a set of baseline results to build upon. Furthermore, by leveraging publicly available programs, all researchers will be able to exactly reproduce results that are described in this paper. We also aim to facilitate comparison of recognition results in the field of Mandarin speech recognition by including a testing set in the toolbox. We describe a toolbox that includes Mandarin speech data from 125 speakers, suitable language model, scripts and data files required for recreating a set of baseline experiments, and a copy of Microsoft SAPI 5.0 SDK that can help professors and students who wish to jumpstart research programs in speech technologies. By lowering the barrier of entry to the field, we hope to encourage more participation in the study of Mandarin speech recognition.

In this paper, we describe a new method to perform speech recognition based on Dynamic HMM architecture. Pitch values are treated as hidden layer and used to modify the parameters of observation probability functions. The results show that the new model achieves approximately 10 percent relative error reductions both in base-syllable recognition task and tonal syllable recognition task. The new method can be used compatibly with conventional HMM based EM training algorithm and Viterbi decoding algorithm.

Analysis and modeling of speaker variability, such as gender, accent, age, speech rate, and phones realizations, are important issues in speech recognition. It is known that existing feature representations describing speaker variations can be of very high dimension. In this paper, we introduce two powerful multivariate statistical analysis methods, namely, principal component analysis (PCA) and independent component analysis (ICA), as tools for analysis of such variability and extraction of low dimensional feature representation. Our findings are the following: (1) the first two principal components correspond to the gender and accent, respectively. The result that the second component corresponding to the accent has never been reported before, to the best of our knowledge. (2) It is shown that ICA based features yield better classification performance than PCA ones. Using 2dimensional ICA representation, we achieved about 6.1% and 13.3% error rate in gender and accent classification, respectively, for 980 speakers.

It is well known that speaker dependent acoustic models can achieve an error rate that is up to a factor of two smaller compared to well trained speaker independent acoustic models [1]. Thus, for improved accuracy, many modern dictation systems require the user to perform enrollment sessions to adapt the acoustic model of the system. In this paper, we present an approach that uses as few as three sentences from the test speaker to select N closest speakers (cohorts) from both the original training set and newly available training speakers to construct customized models. By using such an approach, our adaptation scheme can be updated online without reconfiguring anything that has been determined or calculated before, i.e. speaker independent model. When applying this approach to address dialectal differences, the cohort based user specific models constructed with 3 user sentences can obtain a lower error rate even when compared to user-adapted models based on 170 user sentences. On average, a relative error rate reduction of 22% was achieved.

Some stochastic models like Markov decision process (MDP) are used to model the dialogue manager. MDP-based system degrades fast when uncertainty about user’s intention increases. We propose a novel dialogue model based on the partially observable Markov decision process (POMDP). We use hidden system states and user intentions as the state set, parser results and low-level information as the observation set, domain actions and dialogue repair actions as the action set. Here the low-level information is extracted from different input modals using Bayesian networks. Because of the limitation of exact algorithms, we focus on heuristic methods and their applicability in dialogue management.

The use of discriminative training methods in sp eech recognition is a promising approach. The minimum classification error (MCE) based discriminative methods have been extensively studied and successfully applied to speech recognition [1][2][3], speaker recognition [4], and utterance verification [5][6]. Our goal is to modify the embedded string model based MCE algorithm to train a large number of crosssyllable triphones in a large vocabulary recognition system. In this paper, selective strategy about MCE based discriminative training method, in particular for Mandarin sp eech syllable loop recognition, is introduced. Here, we use syllable loop recognition task to evaluate performance of the acoustic model of an established large vocabulary continuous sp eech recognition system. Since decoding errors only occur in parts of the whole decoded sentence, it is reasonable to adjust only the parameters of the wrong models. As a result, we introduce the weighted MCE formulation, which can provide more effective convergence property and about 10% relative error rate reduction for a large training set. On the other hand, from our experiments, we observed that although the overall performance of recognition system is improved, some originally correct recognition results are misrecognized after discriminative training. To address this issue, we propose a divide and conquer strategy. The acoustic feature sp ace is divided into two or more sub-spaces according to discriminative training procedure. Combining the above two methods, we got more than 14.5% error rate reduction in syllable loop recognition experiments.

In this paper, we present a cartoon face animation system for multimedia HCI applications. We animate face cartoons not only from input speech, but also based on emotions derived from speech signal. Using a corpus of over 700 utterances from different speakers, we have trained SVMs (support vector machines) to recognize four categories of emotions: neutral, happiness, anger and sadness. Given each input speech phrase, we identify its emotion content as a mixture of all four emotions, rather than classifying it into a single emotion. Then, facial expressions are= generated from the recovered emotion for each phrase, by morphing different cartoon templates that correspond to various emotions. To ensure smooth transitions in the animation, we apply low-pass filtering to the recovered (and possibly jumpy) emotion sequence. Moreover, lip-syncing is applied to produce the lip movement from speech, by recovering a statistical audio-visual mapping. Experimental results demonstrate that cartoon animation sequences generated by our system are of good and convincing quality.

Large vocabulary continuous Mandarin speech recognition has been an important problem for speech recognition researchers for several reasons [1], [3]. First of all, it is a tonal language that requires special treatment for the modeling of tones. There are five tones in Mandarin which are necessary to disambiguate between confusable words. Secondly, the difficulty of entering Chinese by keyboard presents a great opportunity for speech recognition to improve computer usability. Previous approaches to modeling tones have included using a separate tone classifier [1] and incorporating pitch directly into the feature vector [3]. In this paper, we describe a large vocabulary Mandarin speech recognition system based on Microsoft's Whisper system. Several alternatives in modeling tones and their error rates on continuous speech are compared.

A method of accent modeling through Pronunciation Dictionary Adaptation (PDA) is presented. We derive the pronunciation variation between canonical speaker groups and accent groups and add an encoding of the differences to a canonical dictionary to create a new, adapted dictionary that reflects the accent characteristics. The pronunciation variation information is then integrated with acoustic and language models into a one-pass search framework. It is assumed that acoustic deviation and pronunciation variation are independent but complementary phenomena that cause poor performance among accented speakers. Therefore, MLLR, an efficient model adaptation technique, is also presented both alone and in combination with PDA. It is shown that when PDA, MLLR and PDA+MLLR are used, error rate reductions of 13.9%, 24.1% and 28.4% respectively are achieved.