Michael Daniel Monkowski

Language model adaptation via minimum discrimination information

Statistical language models improve the performance of speech recognition systems by providing estimates of a priori probabilities of word sequences. The commonly used trigram language models obtain the conditional probability estimate of a word given the previous two words, from a large corpus of text. The text corpus is often a collection of several small diverse segments such as newspaper articles, or conversations on different topics. Knowledge of the current topic could be utilized to adapt the general trigram language models to match that topic closely. For example, an interpolation of the general language model with one built on the topic data could be used. The authors first discuss the adaptation of general trigram language models to a known topic using the minimum discrimination information (MDI) method. They then present results on the switchboard corpus which consists of telephone conversations on several topics.

Speech recognition on Mandarin Call Home: a large-vocabulary, conversational, and telephone speech corpus

We describe IBMs most recent efforts for speech recognition on a conversational-speech database, the Mandarin Call Home corpus. While it is similar to the well-known Switchboard corpus, the Call Home task addresses several major challenges in the domain of spoken language systems, including spontaneous dialogue with no pre-specified topics, limited-bandwidth telephone signal, and recognition of other languages than English. We particularly describe the methodology used in Mandarin Call Home corpus to address language-specific issues. We also examine and compare our results with those of the English Switchboard corpus. Preliminary experiments show that a 58.7% character error rate can be achieved in the context of April 95 Mandarin Call Home data set. The experimental results are comparable to those of the state-of-the-art IBM Switchboard system with similar amount of training data.

E-beam inspection system for comparison of wafer and design data

Effectively patterning the intended design on the wafer for all possible geometries allowed by the design rule document is one of the most critical challenges for semiconductor manufacturing. Despite new lithography techniques like OPC, double patterning and the latest patterning simulation methods, and on-wafer evaluation using brightfield inspection and SEM review tools, patterning problems still occur and can result in a major delay in the qualification of a technology or product. Of particular concern are shorts and opens that cause product chip failure. Initial discovery of yield issues when a chip is being functionally tested is highly undesirable. A system for in-line, die to database (D2DB) comparison using E-beam inspection has been developed to address this risk. This system offers a substantial new line of defense against these patterning issues. The D2DB system is described along with a methodology for applying it for pattern fidelity inspection. Some examples illustrating the system operation are presented.

Context dependent phonetic duration models for decoding conversational speech

Conversational speech provides a particularly difficult task for speech recognition. It provides much more variability than either dictation, read speech, or isolated commands. Phonetic context was used to predict the durations of phones using a decision tree. These predictions were used to calculate context dependent HMM transition probabilities for these phone models, which were used to decode telephone conversations from the SwitchBoard corpus. We observed that the duration models do not appreciably improve the word error rate; that more can be gained by modeling phone durations within words than by adjusting for local average speaking rates; and conclude that local or global variations in speaking rate are not major contributors to the observed high error rates for SwitchBoard.

A continuous speaker-independent putonghua dictation system

We describe new methods for continuous putonghua speech recognition. We have augmented the IBM HMM-based continuous speech recognition system with the following features. First, we treat tones in putonghua as attributes of certain phonemes, instead of syllables. We call those phonemes with tone tonemes. Second, instantaneous pitch is treated as a variable in the acoustic feature vector, in the same way as cepstra or energy. Third, by designing a set of word-segmentation rules to convert the continuous Chinese text into segmented text, the trigram language model works effectively. By applying those new methods, a speaker-independent, very-large-vocabulary continuous putonghua dictation system can be constructed.

E-Beam Hot Spot Inspection for Early Detection of Systematic Patterning Problems for a 22 nm SOI Technology

Early detection of systematic patterning problems can provide a major boost for a technology team. Often in the past, these type defects might only be detected after functional test and subsequent failure analysis. At this point, three to six months of process development time have been lost and three to six months of defective hardware have been wasted. In this paper, a methodology for in-line detection of systematic patterning problems using E-beam hot spot inspection (EBHI) is introduced. Pattern simulation tools and other sources are used to recommend X, Y locations with challenging geometries for evaluation. EBHI evaluates the patterning capability for these locations using modulated wafers. A multifunction team addresses the hot spots that fail within the process window. EBHI is then used to evaluate the solutions proposed by this team. Often, additional data is necessary to determine the full yield impact. This methodology provided tremendous value for IBMs 22 nm SOI technology. Several examples illustrating this point are presented. Line monitoring after the process windows have been established is also discussed.