Abstract: A stereo signal encoding method includes performing spectrum broadening on a quantized line spectral frequency (LSF) parameter of a primary channel signal in a current frame in a stereo signal to obtain a spectrum-broadened LSF parameter of the primary channel signal, determining a prediction residual of an LSF parameter of a secondary channel signal in the current frame based on an original LSF parameter of the secondary channel signal and the spectrum-broadened LSF parameter of the primary channel signal, and performing a quantization on the prediction residual of the LSF parameter of the secondary channel signal.

Abstract: Described herein are devices, methods, and systems for detecting a phrase from uttered speech. A processing device may determine a first model for phrase recognition based on a likelihood ratio using a set of training utterances. The set of utterances may be analyzed by the first model to determine a second model, the second model comprising a training state sequence for each of the set of training utterances, and wherein each training state sequence indicates a likely state for each time interval of a corresponding training utterance. A determination of whether a detected utterance corresponds to the phrase may be based on a concatenation of the first model and the second model.

Abstract: Method and system are provided for audio quality feedback during live transmission from a source that is received at multiple audience devices. The method carried out at a server includes: obtaining audio information of an audio signal as received by at least some of the audience devices in a transmission session; classifying one or more subsets of the audience devices by one or more common factors per subset; and analyzing the obtained audio information from the audience devices in conjunction with the classifications of the subsets of the audience devices to determine one or more common factors that affect received audio quality at an identified subset of the audience devices classified by the one or more common factors. The method provides feedback of the one or more common factors to at least one of the audience devices in the identified subset or to the source device, or to both.

Abstract: An encoding method includes determining an adaptive broadening factor based on a quantized line spectral frequency (LSF) vector of a first channel of a current frame of an audio signal and an LSF vector of a second channel of the current frame, and writing the quantized LSF vector and the adaptive broadening factor into a bitstream.

Abstract: Embodiments may include determination, for each of a plurality of speech frames associated with an acoustic feature, of a phonetic feature based on the associated acoustic feature, generation of one or more two-dimensional feature maps based on the plurality of phonetic features, input of the one or more two-dimensional feature maps to a trained neural network to generate a plurality of speaker embeddings, and aggregation of the plurality of speaker embeddings into a speaker embedding based on respective weights determined for each of the plurality of speaker embeddings, wherein the speaker embedding is associated with an identity of the speaker.

Abstract: According to an embodiment, an electronic device comprises: a communication module comprising communication circuitry, a memory, and at least one processor.

Abstract: A system and a method are disclosed for identifying a subjectively interesting moment in a transcript. In an embodiment, a device receives a transcription of a conversation, and identifies a participant of the conversation. The device accesses a machine learning model corresponding to the participant, and applies, as input to the machine learning model, the transcription. The device receives as output from the machine learning model a portion of the transcription having relevance to the participant, and generates for display, to the participant, information pertaining to the portion.

Abstract: Digitized audio command is decoded to generate audio features. An in-domain confidence score is calculated for a model trained by a limited set of peripheral device commands. An out-domain confidence score is calculated for a model trained without the peripheral device commands. The best score determines whether to process the audio locally or at a remote server. In some embodiments, a likelihood ratio (LR) is calculated of the in-domain and out-domain confidence scores. Based on the likelihood ratio, a locally decoded audio command is performed, or the audio features are sent to a remote server for processing to determine the audio command.

Abstract: Systems, methods, and devices for human-machine interfaces for utterance-based playlist selection are disclosed. In one method, a list of playlists is traversed and a portion of each is audibly output until a playlist command is received. Based on the playlist command, the traversing is stopped and a playlist is selected for playback. In examples, the list of playlists is modified based on a modification input.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for performing long-term prediction (LTP) are described. One example of the methods includes determining a pitch gain and a pitch lag of an input audio signal for at least a predetermined number of frames. It is determined that the pitch gain of the input audio signal has exceeded a predetermined threshold and that a change of the pitch lag of the input audio signal has been within a predetermined range for at least the predetermined number of frames. In response to determining that the pitch gain of the input audio signal has exceeded the predetermined threshold and that the change of the third pitch lag has been within the predetermined range for at least the predetermined number of frames, a pitch gain is set for a current frame of the input audio signal.

Abstract: A system providing a multi-persona social agent includes a computing platform having a hardware processor, a system memory storing a software code, and multiple neural network (NN) based predictive models accessible by the software code. The hardware processor executes the software code to receive input data corresponding to an interaction with a user, determine a generic expression for use in the interaction, and identify one of the character personas as a persona to be assumed by the multi-persona social agent. The software code also generates, using the generic expression and one of the NN based predictive models corresponding to the persona to be assumed by the multi-persona social agent, a sentiment driven personified response for the interaction with the user based on a vocabulary, phrases, and one or more syntax rules idiosyncratic to the persona to be assumed, and renders the sentiment driven personified response using the multi-persona social agent.

Abstract: A system is typically configured for customizing interconnectivity of one or more layers associated with a neural network architecture, wherein the neural network architecture is associated with an application, customizing functional transformation of the one or more layers associated with the neural network architecture, wherein each of the one or more layers comprises a custom transformation function, and generating a custom neural network architecture based on customizing the interconnectivity and the functional transformation of the one or more layers.

Abstract: In various embodiments, a voice command is transmitted while the user is performing an activity. The user’s activity is determined, along with follow on actions for the request. If the user is performing an interaction-limiting activity, subsequent actions may be presented in a simplified display to reduce the likelihood that user interacts with the display. Additionally, certain follow on actions may be considered complex. If a complex follow on action occurs during an interaction-limiting activity, the request may be flagged and delayed until the user has completed the interaction-limiting activity.

Abstract: Quantizing speech model parameters includes, for each of multiple vectors of quantized excitation strength parameters, determining first and second errors between first and second elements of a vector of excitation strength parameters and, respectively, first and second elements of the vector of quantized excitation strength parameters, and determining a first energy and a second energy associated with, respectively, the first and second errors. First and second weights for, respectively, the first error and the second error, are determined and are used to produce first and second weighted errors, which are combined to produce a total error. The total errors of each of the multiple vectors of quantized excitation strength parameters are compared and the vector of quantized excitation strength parameters that produces the smallest total error is selected to represent the vector of excitation strength parameters.

Abstract: In one aspect, a playback device includes at least one microphone configured to detect a voice input and generate sound input data. The playback device detects a first command keyword in the detected sound and, in response, makes a first determination, via a first local natural language unit (NLU), whether the input sound data includes at least one keyword within a first predetermined library of keywords. The playback device receives an indication of a second determination made by a second NLU that the input sound data includes at least one keyword from a second predetermined library of keywords. The playback device compares the results of the first determination and the second determination and, based on the comparison, foregoes further processing of the input sound data.

Abstract: Deep Neural Networks (DNNs) for forecasting future data are provided. In one embodiment, a non-transitory computer-readable medium is configured to store computer logic having instructions that, when executed, cause one or more processing devices to receive, at each of a plurality of Deep Neural Network (DNN) forecasters, an input corresponding to a time-series dataset of a plurality of input time-series datasets. The instructions further cause the one or more processing devices to produce, from each of the plurality of DNN forecasters, a forecast output and provide the forecast output from each of the plurality of DNN forecasters to a DNN mixer for combining the forecast outputs to produce one or more output time-series datasets.

Abstract: Playback devices can support audio encoded using various encoding schemes. Playing back such content includes receiving, at a playback device, audio data from an audio source; and receiving an indication from the audio source that the audio data is encoded in the compressed audio format. The device determines, independently of receiving the indication from the audio source that the audio data is encoded in the compressed audio format, whether the audio data is encoded in a compressed audio format. If the audio data is determined to be encoded in the compressed audio format: the device selects a decoder from among a plurality of decoders; decodes the audio data using the selected decoder; and plays back the decoded audio data via the playback device. If the audio data is determined not to be encoded in the compressed audio format, the device inhibits playback of the audio data.

Abstract: To obtain an appropriate evaluation value in an acoustic quality evaluation by a conversational test. An acoustic quality evaluation apparatus 3 evaluates the acoustic quality of a call performed between a near-end terminal 1 and a far-end terminal 2 via a voice communication network 4. An evaluation value presenting unit 31 displays, on a display unit 13, evaluation categories obtained by classifying each of a plurality of evaluation viewpoints into a predetermined number of levels. An input unit 14 transmits the evaluation category selected by the evaluator for each of the evaluation viewpoints, to an evaluation value determination unit 32. The evaluation value determination unit 32 determines the lowest evaluation value among evaluation values assigned to the evaluation category received from the input unit 14 as a subjective evaluation value for acoustic quality.

Abstract: An audio signal encoding and decoding method using a learning model, a training method of the learning model, and an encoder and decoder that perform the method, are disclosed. The audio signal decoding method may include extracting a first residual signal and a first linear prediction coefficient by decoding a bitstream received from an encoder, generating a first audio signal from the first residual signal using the first linear prediction coefficient, generating a second linear prediction coefficients and a second residual signal from the first audio signal, obtaining a third linear prediction coefficient by inputting the second linear prediction coefficient into a trained learning model, and generating a second audio signal from the second residual signal using the third linear prediction coefficient.

Abstract: The present disclosure is generally directed a system to detect activation phrases within input audio signals transmitted over a low-bandwidth network. The system can use a two-stage activation phrase detection process. First a sensing device, which can include a plurality of microphones for detecting an input audio signal, can detect an input audio signal that includes a candidate activation phrase. Second, the sensing device can transmit the recordings of the input audio signal to a client device for confirmation that the input audio signal includes the activation phrase.