Abstract: A speech synthesis system may be configured to be robust against variations and errors in spelling and/or punctuation in the input text. A text modifier may generate a parallel training dataset by modifying text from a training dataset to include variations in spelling, punctuation, and/or formatting. The speech synthesis system may generate synthesized speech based on the modified text in the parallel training dataset. A robustness tester may compare audio from the original training dataset with synthesized speech generated using the modified text. The results may be used to update parameters of one or more speech generation models of the speech synthesis system. The results may also be used to adjust the frequency of modifications generated by the text modifier to, for example, ensure that performance of the speech synthesis system on unmodified text is not adversely affected by the training using the modified text.

Abstract: As described herein, a system, method, and computer program provide real-time language translation using generative artificial intelligence. An input in a first spoken language is received. The input in the first spoken language is processed, using a generative artificial intelligence model, to generate a translation of the input in a second spoken language. The translation is output.

Abstract: A computer-implemented method comprising receiving, by a computer, at least one input corresponding to at least one of an attribute of a user or an interest of the user; executing, by the computer, a machine-learning model to generate at least one topic associated with an essay based on the at least one input; in response to outputting, by the computer, the at least one topic, receiving, by the computer, an electronic document; executing, by the computer, the machine-learning model to identify a score associated with the electronic document using the at least one topic; and displaying, by the computer, the score.

Abstract: A system and method for mitigating generative artificial intelligence errors during inter-agent communications. A method includes populating an intent field of a schema with an intent value representing an intent of an inter-agent communication session including a first artificial intelligence (AI) agent and a second AI agent, wherein the intent value is determined based on a communication from the first AI agent; comparing each of inputs from the first AI agent and the second AI agent to the schema with respect to the intent value; detecting a misalignment between a first input and the schema when a dissimilarity between the first input and at least a portion of the schema including the intent value exceeds a threshold; and performing a mitigation action based on the detected misalignment in order to mitigate an effect of generative artificial intelligence error on the first input.

Abstract: A method includes receiving training data that includes a set of unspoken textual utterances. For each respective unspoken textual utterance, the method includes, tokenizing the respective textual utterance into a sequence of sub-word units, generating a first higher order textual feature representation for a corresponding sub-word unit tokenized from the respective unspoken textual utterance, receiving the first higher order textual feature representation generated by a text encoder, and generating a first probability distribution over possible text units. The method also includes training an encoder based on the first probability distribution over possible text units generated by a first-pass decoder for each respective unspoken textual utterance in the set of unspoken textual utterances.

Abstract: Provided is a reverberation removal device that is highly accurate even in noisy environments and underdetermined conditions. Reverberation is removed by applying a plurality of reverberation prediction filters to an observation signal while switching the plurality of reverberation prediction filters according to each time frequency bin of the observation signal.

Abstract: A speech signal enhancement method includes: performing noise reduction processing on a first speech signal according to a first time-frequency spectrum and a first power spectrum to obtain a second speech signal, where the first time-frequency spectrum is used to indicate a time domain feature and a frequency domain feature of the first speech signal, and the first power spectrum is a power spectrum of a noise signal in the first speech signal; determining a voiced signal in the second speech signal, and performing gain compensation on the voiced signal; and determining a damage compensation gain of the second speech signal according to the voiced signal on which the gain compensation has been performed, and performing gain compensation on the second speech signal based on the damage compensation gain.

Abstract: A computer implemented method for translating a document. A number of processor units separate the document into elements having media types. The number of processor units determine attributes for the elements. The number of processor units create a virtual map identifying relationships between the elements using the attributes. The number of processor units translate the elements into a target language based on media types for the elements. The number of processor units adjust translations for the elements based on the relationships between the elements using the virtual map to create adjusted translations for the elements. The number of processor units generate the translated document using the adjusted translations for the elements and the virtual map.

Abstract: Systems and methods of the present disclosure enable authentication and/or anomaly detection using machine learning-based modelling. Audio recordings that represent audio from a forced cough vocalizations are received from a user device. One or more audio filters extract forced cough vocalization recordings from the audio recordings and signal data signatures representative of the forced cough vocalization recordings are generated. Gaussian mixture models are produced for each unique combination of the signal data signatures, where each unique combination include a group of model baselines and a test match baseline. Each Gaussian mixture model is used to produce a match value for the associated test match baseline based on the associated model baselines, and a statistical score is determined for each match value. One or more baseline Gaussian mixture models are determined based on the statistical score and stored in a user profile.

Abstract: There is provided a sound synthesis apparatus. The apparatus comprises a transceiver configured to obtain a plurality of sound samples; and a processor, wherein the processor is configured to: preprocess each sound sample to convert each sound sample into a spectrogram; generate a plurality of latent codes by inputting the spectrogram of each sound sample to an encoder of an artificial neural network pre-trained to output a latent code that maximizes timbre information; generate one synthesized latent code by synthesizing the plurality of latent codes based on a weight present for each sound sample; and generate a synthesized sound by inputting the synthesized latent code to a decoder of the pre-trained artificial neural network.

Abstract: An embodiment may involve: obtaining textual content including a plurality of token strings, wherein each of the plurality of token strings includes one or more tokens; determining, for the plurality of token strings, respectively corresponding sets of n-gram tuples; assigning respective weights to the plurality of token strings, wherein, for each of the plurality of token strings, the assignment is based on the respectively corresponding set of n-gram tuples; identifying a subset of the plurality of token strings, wherein each of the subset of the plurality of token strings is characterized by a respective weight that exceeds a predetermined threshold weight; and storing sets of n-gram tuples respectively corresponding to the subset of the plurality of token strings.

Abstract: The present invention relates to a method and an apparatus for synthesizing the voice based on brain waves during imagined speech. The method for synthesizing the voice based on brain waves during imagined speech according to an embodiment of the present invention may include the following steps: a step to obtain the user's brain waves during imagined speech; a step to convert the above-mentioned brain waves of imagined speech into embedding vectors; a step to generate the mel-spectrograms based on the above-mentioned embedding vectors; a step to generate the voice using the above-mentioned mel-spectrograms; a step to output the above-mentioned voice.

Abstract: A speech synthesis method includes: obtaining an acoustic feature sequence of a text to be processed; processing the acoustic feature sequence by using a non-autoregressive computing model in parallel to obtain first audio information of the text to be processed, wherein the first audio information comprises audio corresponding to each segment; processing the acoustic feature sequence and the first audio information by using an autoregressive computing model to obtain a residual value corresponding to each segment; and obtaining second audio information corresponding to an i-th segment based on the first audio information corresponding to the i-th segment and the residual values corresponding to a first to an (i?1)-th segment, wherein a synthesized audio of the text to be processed comprises each of the second audio information, i=1, 2 . . . n, n is a total number of the segments.

Abstract: Techniques for adjusting outlier datasets for training chatbot systems in natural language processing are disclosed. In one particular aspect, a method is provided that includes receiving a dataset that includes training or inference data. An initial set of outlier data points can be identified within the dataset based on a score of the outlier data points being above or below a threshold. The initial set can be adjusted by identifying one or more nearest neighbors, which can be included in the dataset. Outlier data points that include a label that matches a number of labels of the nearest neighbors that exceeds a predetermined threshold can be removed from the initial set of outlier data points to generate a final set. Outlier data points of the final set can be adjusted with respect to the dataset to generate a set of training data that is used to train a machine-learning model.

Abstract: A concealed text feature corresponding to a text data block of a plurality of text data blocks included in the text data and at least one concealed text feature corresponding to at least one text data block subsequent to the text data block are generated. A coarse fusion is performed on (i) the concealed text feature corresponding to the text data block and (ii) the at least one concealed text feature corresponding to the at least one text data block subsequent to the text data block to obtain at least one coarse fusion text feature. A fine fusion is performed on the at least one coarse fusion text feature to obtain a fine fusion text feature corresponding to the text data block. A length corresponding to the fine fusion text feature is regulated. The fine fusion text feature with the regulated length is transformed into the acoustic feature.