Microsoft Technology Licensing, LLC

Other

Max Leung

Jian Luan

Kaisheng N Yao

Bo Yan

Pei Zhao

Lei He

Yu Shi

Mei-Yuh Hwang

Sheng Zhao

Edward Un

Malone Ma

Peng Wang

Fileno A Alleva

Shenghua Ye

Chi-Ho Li

Binggong Ding

Baoxun Wang

Difei Gao

Yijian Wu

An example intent-recognition system comprises a processor and memory storing instructions. The instructions cause the processor to receive speech input comprising spoken words. The instructions cause the processor to generate text results based on the speech input and generate acoustic feature annotations based on the speech input. The instructions also cause the processor to apply an intent model to the text result and the acoustic feature annotations to recognize an intent based on the speech input. An example system for adapting an emotional text-to-speech model comprises a processor and memory. The memory stores instructions that cause the processor to receive training examples comprising speech input and receive labelling data comprising emotion information associated with the speech input. The instructions also cause the processor to extract audio signal vectors from the training examples and generate an emotion-adapted voice font model based on the audio signal vectors and the labelling data.

Intent recognition and emotional text-to-speech learning

Techniques for generating voice with predetermined emotion type. In an aspect, semantic content and emotion type are separately specified for a speech segment to be generated. A candidate generation module generates a plurality of emotionally diverse candidate speech segments, wherein each candidate has the specified semantic content. A candidate selection module identifies an optimal candidate from amongst the plurality of candidate speech segments, wherein the optimal candidate most closely corresponds to the predetermined emotion type. In further aspects, crowd-sourcing techniques may be applied to generate the plurality of speech output candidates associated with a given semantic content, and machine-learning techniques may be applied to derive parameters for a real-time algorithm for the candidate selection module.

Voice generation with predetermined emotion type

Techniques for selecting an emotion type code associated with semantic content in an interactive dialog system. In an aspect, fact or profile inputs are provided to an emotion classification algorithm, which selects an emotion type based on the specific combination of fact or profile inputs. The emotion classification algorithm may be rules-based or derived from machine learning. A previous user input may be further specified as input to the emotion classification algorithm. The techniques are especially applicable in mobile communications devices such as smartphones, wherein the fact or profile inputs may be derived from usage of the diverse function set of the device, including online access, text or voice communications, scheduling functions, etc.

Emotion type classification for interactive dialog system

Multi-voice font interpolation is provided. A multi-voice font interpolation engine allows the production of computer generated speech with a wide variety of speaker characteristics and/or prosody by interpolating speaker characteristics and prosody from existing fonts. Using prediction models from multiple voice fonts, the multi-voice font interpolation engine predicts values for the parameters that influence speaker characteristics and/or prosody for the phoneme sequence obtained from the text to spoken. For each parameter, additional parameter values are generated by a weighted interpolation from the predicted values. Modifying an existing voice font with the interpolated parameters changes the style and/or emotion of the speech while retaining the base sound qualities of the original voice. The multi-voice font interpolation engine allows the speaker characteristics and/or prosody to be transplanted from one voice font to another or entirely new speaker characteristics and/or prosody to be generated for an existing voice font.

Voice font speaker and prosody interpolation

The technology relates to converting text to speech utilizing recurrent neural networks (RNNs). The recurrent neural networks may be implemented as multiple modules for determining properties of the text. In embodiments, a part-of-speech RNN module, letter-to-sound RNN module, a linguistic prosody tagger RNN module, and a context awareness and semantic mining RNN module may all be utilized. The properties from the RNN modules are processed by a hyper-structure RNN module that determine the phonetic properties of the input text based on the outputs of the other RNN modules. The hyper-structure RNN module may generate a generation sequence that is capable of being converting to audible speech by a speech synthesizer. The generation sequence may also be optimized by a global optimization module prior to being synthesized into audible speech.

Hyper-structure recurrent neural networks for text-to-speech

An example text-to-speech learning system performs a method for generating a pronunciation sequence conversion model. The method includes generating a first pronunciation sequence from a speech input of a training pair and generating a second pronunciation sequence from a text input of the training pair. The method also includes determining a pronunciation sequence difference between the first pronunciation sequence and the second pronunciation sequence; and generating a pronunciation sequence conversion model based on the pronunciation sequence difference. An example speech recognition learning system performs a method for generating a pronunciation sequence conversion model. The method includes extracting an audio signal vector from a speech input and applying an audio signal conversion model to the audio signal vector to generate a converted audio signal vector. The method also includes adapting an acoustic model based on the converted audio signal vector to generate an adapted acoustic model.

Speech recognition and text-to-speech learning system

Techniques for converting text to speech having emotional content. In an aspect, an emotionally neutral acoustic trajectory is predicted for a script using a neutral model, and an emotion-specific acoustic trajectory adjustment is independently predicted using an emotion-specific model. The neutral trajectory and emotion-specific adjustments are combined to generate a transformed speech output having emotional content. In another aspect, state parameters of a statistical parametric model for neutral voice are transformed by emotion-specific factors that vary across contexts and states. The emotion-dependent adjustment factors may be clustered and stored using an emotion-specific decision tree or other clustering scheme distinct from a decision tree used for the neutral voice model.

Text-to-speech with emotional content

A text-to-speech (TTS) engine combines recorded speech with synthesized speech from a TTS synthesizer based on text input. The TTS engine receives the text input and identifies the domain for the speech (e.g. navigation, dialing, . . . ). The identified domain is used in selecting domain specific speech recordings (e.g. pre-recorded static phrases such as 'turn left', 'turn right' . . . ) from the input text. The speech recordings are obtained based on the static phrases for the domain that are identified from the input text. The TTS engine blends the static phrases with the TTS output to smooth the acoustic trajectory of the input text. The prosody of the static phrases is used to create similar prosody in the TTS output.

Growing community of inventors

Beijing, China

Max Leung