Microsoft Technology Licensing, LLC

Alibaba Group Holding Limited

Min Chu

Yong Zhao

Hu Peng

Jian Qiao Sun

Frank Kao-Ping Soong

Jian Wang

Yi Chen

Jie Su

Lei Hou

Wenli Zhu

Lijuan Wang

Yusheng Li

Yining Chen

Frank Kao-ping Soong

Jing Ming Tang

Xiao Ling Liao

Bing Jing Xu

Ren Gang Peng

Yang Yang

Yifan Gong

Xiaodong He

Sheng Zhao

Jian-Lai Zhou

Chao Huang

Jiali You

Yao Qian

Xin Zou

Shiun-Zu Kuo

Kevin E Feige

Megan Riley

Huixing Jiang

Analyzing textual data is disclosed, including by: receiving textual data; determining that the textual data is a candidate for analogy analysis based at least in part on at least a portion of the textual data matching an analogical question template; extracting a source substantive from the textual data; using the source substantive to determine a target substantive from a word vector model that is trained on a set of training data; and generating an answer including the target substantive based at least in part on an analogical answer template corresponding to the analogical question template.

Analyzing textual data

Text processing includes: segmenting received text based on a lexicon of smallest semantic units to obtain medium-grained segmentation results; merging the medium-grained segmentation results to obtain coarse-grained segmentation results, the coarse-grained segmentation results having coarser granularity than the medium-grained segmentation results; looking up in the lexicon of smallest semantic units respective search elements that correspond to segments in the medium-grained segmentation results; and forming fine-grained segmentation results based on the respective search elements, the fine-grained segmentation results having finer granularity than the medium-grained segmentation results.

Text segmentation with multiple granularity levels

An automated method of providing a pronunciation of a word to a remote device is disclosed. The method includes receiving an input indicative of the word to be pronounced. The method further includes searching a database having a plurality of records. Each of the records has an indication of a textual representation and an associated indication of an audible representation. At least one output is provided to the remote device of an audible representation of the word to be pronounced.

Name synthesis

Described is a technology by which synthesized speech generated from text is evaluated against a prosody model (trained offline) to determine whether the speech will sound unnatural. If so, the speech is regenerated with modified data. The evaluation and regeneration may be iterative until deemed natural sounding. For example, text is built into a lattice that is then (e.g., Viterbi) searched to find a best path. The sections (e.g., units) of data on the path are evaluated via a prosody model. If the evaluation deems a section to correspond to unnatural prosody, that section is replaced, e.g., by modifying/pruning the lattice and re-performing the search. Replacement may be iterative until all sections pass the evaluation. Unnatural prosody detection may be biased such that during evaluation, unnatural prosody is falsely detected at a higher rate relative to a rate at which unnatural prosody is missed.

Unnatural prosody detection in speech synthesis

Described is a technology in which a finite state machine-based method segments original sequential data into high level units. Segments with similar sub-sequences are treated as the same unit. In general, the finite state machine indentifies sub-sequences in the sequential data that have similar tree-like microstructures. One described finite state machine operates by entering a path state when an input is a path command corresponding to a non-leaf node; the state machine remains in the path state until another command comprises an action command corresponding to a leaf node, whereby it outputs an action unit, or until the other command comprises a path command that is not the parent or the sibling of a next command, whereby a browsing unit is output. Also described is aggregating sub-sequences of a same unit to produce an access structure corresponding to the unit.

Segmenting sequential data with a finite state machine

Described is a technology in which sequential data, such as application program command sequences, are processed into patterns, such as for use in analyzing program usage. In one aspect, sequential data may be first transformed via state machines that remove repeated data, group similar data into sub-sequences, and/or remove noisy data. The transformed data is then segmented into units. A pattern extraction mechanism extracts patterns from the units into a pattern set, by calculating a stability score (e.g., a mutual information score) between succeeding units, selecting the pair of units having the most stability (e.g., the highest score), and adding corresponding information for that pair into the pattern set. Pattern extraction is iteratively repeated until a stopping criterion is met, e.g., the pattern set reaches a defined size, or when the stability score is smaller than a pre-set threshold.

Extracting patterns from sequential data

The language of origin of a word is determined by analyzing non-uniform letter sequence portions of the word.

Identifying language origin of words

Methods are disclosed for automatic accent labeling without manually labeled data. The methods are designed to exploit accent distribution between function and content words.

Unsupervised labeling of sentence level accent

A method for synthesizing speech from text includes receiving one or more waveforms characteristic of a voice of a person selected by a user, generating a personalized voice font based on the one or more waveforms, and delivering the personalized voice font to the user's computer, whereby speech can be synthesized from text, the speech being in the voice of the selected person, the speech being synthesized using the personalized voice font. A system includes a text-to-speech (TTS) application operable to generate a voice font based on speech waveforms transmitted from a client computer remotely accessing the TTS application.

Providing personalized voice font for text-to-speech applications

The language of origin of a word or named entity is predicted using estimates of frequency of occurrence of the word or named entity in different languages. In one embodiment, the normalized frequency of occurrence of the word or named entity in a variety of different languages is estimated and the values are used as features in a feature vector which is scored and used to identify language of origin.

Identifying language of origin for words using estimates of normalized appearance frequency

Described is a voice persona service by which users convert text into speech waveforms, based on user-provided parameters and voice data from a service data store. The service may be remotely accessed, such as via the Internet. The user may provide text tagged with parameters, with the text sent to a text-to-speech engine along with base or custom voice data, and the resulting waveform morphed based on the tags. The user may also provide speech. Once created, a voice persona corresponding to the speech waveform may be persisted, exchanged, made public, shared and so forth. In one example, the voice persona service receives user input and parameters, and retrieves a base or custom voice that may be edited by the user via a morphing algorithm. The service outputs a waveform, such as a .wav file for embedding in a software program, and persists the voice persona corresponding to that waveform.

Voice persona service for embedding text-to-speech features into software programs

A phone set for use in speech processing such as speech recognition or text-to-speech conversion is used to model or form syllables of a tonal language having a plurality of different tones. Each syllable includes an initial part that can be glide dependent and a final part. The final part includes a plurality of phones. Each phones carries partial tonal information such that the phones taken together implicitly and jointly represent the different tones.

Segmental tonal modeling for tonal languages

A text processing system for processing multi-lingual text for a speech synthesizer includes a first language dependent module for performing at least one of text and prosody analysis on a portion of input text comprising a first language. A second language dependent module performs at least one of text and prosody analysis on a second portion of input text comprising a second language. A third module is adapted to receive outputs from the first and second dependent module and performs prosodic and phonetic context abstraction over the outputs based on multi-lingual text.

Front-end architecture for a multi-lingual text-to-speech system

A method and apparatus are provided for refining segmental boundaries in speech waveforms. Contextual acoustic feature similarities are used as a basis for clustering adjacent phoneme speech units, where each adjacent pair phoneme speech units include a segmental boundary. A refining model is trained for each cluster and used to refine boundaries of contextual phoneme speech units forming the clusters.

Growing community of inventors

Beijing, China

Min Chu