Abstract: Introduced here are approaches to training and then employing computer-implemented models designed to upsample discrete audio signals to higher sampling rates. Assume, for example, that a media production platform obtains a first discrete signal at a relatively low sampling rate. The relatively low sampling frequency may make the first discrete audio signal unsuitable for inclusion in media compilations, so the media production platform may attempt to improve its quality through upsampling. To accomplish this, the media production platform can apply a transform to the first discrete signal to produce a first magnitude spectrogram. Then, the media production platform can apply a computer-implemented model to the first magnitude spectrogram to produce a second magnitude spectrogram. Thereafter, the media production platform can apply an inverse transform to the second magnitude spectrogram to create a second discrete signal that has a higher sampling rate than the first discrete audio signal.

Abstract: A method, computer program product, and computing system for generating a plurality of acoustic relative transfer functions between a plurality of audio acquisition devices of an audio recording system based upon, at least in part, one or more of a predefined speech processing application and a predefined acoustic environment. An acoustic relative transfer function codebook may be generated using the plurality of acoustic relative transfer functions. One or more channels from the plurality of audio acquisition devices of the audio recording system may be encoded using the acoustic relative transfer function codebook.

Abstract: An accompaniment classification method and apparatus is provided. The method includes the following. A first type of audio features of a target accompaniment is obtained (S301, S401). Data normalization is performed on each kind of audio features in the first type of audio features of the target accompaniment to obtain a first feature-set of the target accompaniment and the first feature-set is input into a first classification model for processing (S302, S402). A first probability value output by the first classification model for the first feature-set is obtained (S303, S403). An accompaniment category of the target accompaniment is determined to be a first category of accompaniments when the first probability value is greater than a first classification threshold (S404). The accompaniment category of the target accompaniment is determined to be other categories of accompaniments when the first probability value is less than or equal to the first classification threshold.

Abstract: A data processing method based on simultaneous interpretation, applied to a server in a simultaneous interpretation system, including: obtaining audio transmitted by a simultaneous interpretation device; processing the audio by using a simultaneous interpretation model to obtain an initial text; transmitting the initial text to a user terminal; receiving a modified text fed back by the user terminal, the modified text being obtained after the user terminal modifies the initial text; and updating the simultaneous interpretation model according to the initial text and the modified text.

Abstract: Embodiments described herein provide for end-to-end joint determination of degradation parameter scores for certain types of degradation. Degradation parameters include degradation describing additive noise and multiplicative noise such as Signal-to-Noise Ratio (SNR), reverberation time (T60), and Direct-to-Reverberant Ratio (DRR). Various neural network architectures are described such that the inherent interplay between the degradation parameters is considered in both the degradation parameter score and degradation score determination. The neural network architectures are trained according to computer generated audio datasets.

Abstract: Systems and methods related to partial interlace frequency domain resource allocations for uplink transmissions from a wireless device are disclosed. In some embodiments, a method performed by a wireless device comprises receiving a resource allocation for an uplink transmission that allocates resources in one or more partially allocated interlaces and performing an uplink transmission on the allocated resources in the one or more partially allocated interlaces in accordance with the resource allocation. In this manner, a low-complexity approach to support flexible frequency domain resource allocation is provided. In addition, using this approach, an interlace may be shared by two or more wireless devices, which may increase spectral efficiency and reduce latency.

Abstract: Methods, apparatus, systems and articles of manufacture are disclosed for watermarking using starting phase modulation. An example apparatus includes at least one memory, machine readable instructions, and processor circuitry to execute the machine readable instructions to at least access a media signal, access a watermark symbol to be encoded into the media signal, determine bit values of the watermark symbol, determine a common starting phase value for watermark components of the watermark symbol, the common starting phase value to represent at least one of the bit values of the watermark symbol, and embed the watermark components into the media signal based on the common starting phase value.

Abstract: Systems and methods related to partial interlace frequency domain resource allocations for uplink transmissions from a wireless device are disclosed. In some embodiments, a method performed by a wireless device comprises receiving a resource allocation for an uplink transmission that allocates resources in one or more partially allocated interlaces and performing an uplink transmission on the allocated resources in the one or more partially allocated interlaces in accordance with the resource allocation. In this manner, a low-complexity approach to support flexible frequency domain resource allocation is provided. In addition, using this approach, an interlace may be shared by two or more wireless devices, which may increase spectral efficiency and reduce latency.

Abstract: In one aspect, a method includes detecting a fingerprint match between query fingerprint data representing at least one audio segment within podcast content and reference fingerprint data representing known repetitive content within other podcast content, detecting a feature match between a set of audio features across multiple time-windows of the podcast content, and detecting a text match between at least one query text sentences from a transcript of the podcast content and reference text sentences, the reference text sentences comprising text sentences from the known repetitive content within the other podcast content. The method also includes responsive to the detections, generating sets of labels identifying potential repetitive content within the podcast content. The method also includes selecting, from the sets of labels, a consolidated set of labels identifying segments of repetitive content within the podcast content, and responsive to selecting the consolidated set of labels, performing an action.

Abstract: An apparatus for decoding an encoded audio signal, includes a spectral domain audio decoder for generating a first decoded representation of a first set of first spectral portions, the decoded representation having a first spectral resolution; a parametric decoder for generating a second decoded representation of a second set of second spectral portions having a second spectral resolution being lower than the first spectral resolution; a frequency regenerator for regenerating every constructed second spectral portion having the first spectral resolution using a first spectral portion and spectral envelope information for the second spectral portion; and a spectrum time converter for converting the first decoded representation and the reconstructed second spectral portion into a time representation.

Abstract: Innovations in phase quantization during speech encoding and phase reconstruction during speech decoding are described. For example, to encode a set of phase values, a speech encoder omits higher-frequency phase values and/or represents at least some of the phase values as a weighted sum of basis functions. Or, as another example, to decode a set of phase values, a speech decoder reconstructs at least some of the phase values using a weighted sum of basis functions and/or reconstructs lower-frequency phase values then uses at least some of the lower-frequency phase values to synthesize higher-frequency phase values. In many cases, the innovations improve the performance of a speech codec in low bitrate scenarios, even when encoded data is delivered over a network that suffers from insufficient bandwidth or transmission quality problems.

Abstract: A measurement method includes generating a plurality of second measurement signals by disposing a plurality of first measurement signals corresponding to each of the plurality of speakers in respective different time zones on a time axis, generating a plurality of third measurement signals by copying a portion of a back end of each of the plurality of second measurement signals and adding the portion to a front end of each of the plurality of second measurement signals, outputting sounds according to each of the plurality of third measurement signals from each of the plurality of speakers, collecting the sounds with a microphone, and calculating a plurality of impulse responses corresponding to the plurality of first measurement signals, based on the collected sound signal collected with the microphone and the plurality of third measurement signals.

Abstract: An audio packet error concealment system includes an encoding unit for encoding an audio signal consisting of a plurality of frames, and an auxiliary information encoding unit for estimating and encoding auxiliary information about a temporal change of power of the audio signal. The auxiliary information is used in packet loss concealment in decoding of the audio signal. The auxiliary information about the temporal change of power may contain a parameter that functionally approximates a plurality of powers of subframes shorter than one frame, or may contain information about a vector obtained by vector quantization of a plurality of powers of subframes shorter than one frame.

Abstract: A keyword detection method includes: obtaining an enhanced speech signal of a to-be-detected speech signal, the enhanced speech signal corresponding to a target speech speed; performing speed adjustment on the enhanced speech signal to obtain a first speed-adjusted speech signal having a first speech speed, the first speech speed being different from the target speech speed; obtaining a first speech feature signal according to the first speed-adjusted speech signal; obtaining a detection result according to a first keyword detection result corresponding to the first speech feature signal, the detection result indicating whether a target keyword exists in the to-be-detected speech signal; and performing an operation corresponding to the target keyword in response to determining that the target keyword exists according to the detection result.

Abstract: An audio system for an elevator includes two or more speaker cabinets arranged inside a suspended ceiling fixed to a ceiling board of a car of the elevator, an input device to which sound content radiated to an inside of the car from each of the two or more speaker cabinets are input, and a sound field control device configured to conduct phase control and reverberation time control for the sound content and thereby cause a sound wave based on the sound content to be radiated from the speaker cabinet to the inside of the car. Each of the speaker cabinets includes a casing arranged inside the suspended ceiling, and a speaker unit arranged inside the casing and having a radiation surface that radiates the sound wave.

Abstract: A target speech signal extraction method for robust speech recognition includes: initializing a steering vector for a target speech source and an adaptive vector, setting a real output channel of the target speech source as an output by the adaptive vector, initializing adaptive vectors for a noise and setting a dummy channel as an output by the adaptive vectors for the noise; setting a cost function for minimizing dependency between a real output for the target speech source and a dummy output for the noise; setting an auxiliary function to the cost function, and updating the adaptive vector for the target speech source and the adaptive vectors for the noise by using the auxiliary function and the steering vector; estimating the target speech signal by using the adaptive vector thereby extracting the target speech signal from the input signals; and updating the steering vector for the target speech source.

Abstract: An audio encoder has a first information sink oriented encoding branch such as a spectral domain encoding branch, a second information source or SNR oriented encoding branch such as an LPC-domain encoding branch, and a switch for switching between the first and second encoding branches, the second encoding branch having a converter into a specific domain different from the spectral domain such as an LPC analysis stage generating an excitation signal, and the second encoding branch having a specific domain coding branch such as LPC domain processing branch, and a specific spectral domain coding branch such as LPC spectral domain processing branch, and an additional switch for switching between the specific domain coding branch and the specific spectral domain coding branch. An audio decoder has a first domain decoder, a second domain decoder, and a third domain decoder as well as two cascaded switches for switching between the decoders.

Abstract: An audio encoder has a first information sink oriented encoding branch such as a spectral domain encoding branch, a second information source or SNR oriented encoding branch such as an LPC-domain encoding branch, and a switch for switching between the first and second encoding branches, the second encoding branch having a converter into a specific domain different from the spectral domain such as an LPC analysis stage generating an excitation signal, and the second encoding branch having a specific domain coding branch such as LPC domain processing branch, and a specific spectral domain coding branch such as LPC spectral domain processing branch, and an additional switch for switching between the specific domain coding branch and the specific spectral domain coding branch. An audio decoder has a first domain decoder, a second domain decoder, and a third domain decoder as well as two cascaded switches for switching between the decoders.

Abstract: A device capable of autonomous motion may move in an environment and may receive audio data from a microphone. A model may be trained to process the audio data to determine mask data, which may be used to mask noise in the audio data. Training data for the model may be normalized before training, and different loss functions may be used for different types of training data.

Abstract: A system for detection of at least one designated wake-up word for at least one speech-enabled application. The system comprises at least one microphone; and at least one computer hardware processor configured to perform: receiving an acoustic signal generated by the at least one microphone at least in part as a result of receiving an utterance spoken by a speaker; obtaining information indicative of the speaker's identity; interpreting the acoustic signal at least in part by determining, using the information indicative of the speaker's identity and automated speech recognition, whether the utterance spoken by the speaker includes the at least one designated wake-up word; and interacting with the speaker based, at least in part, on results of the interpreting.

Abstract: An encoder operable to filter audio signals into a plurality of frequency band components, generate quantized digital components for each band, identify a potential for pre-echo events within the generated quantized digital components, generate an approximate signal by decoding the quantized digital components using inverse pulse code modulation, generate an error signal by comparing the approximate signal with the sampled audio signal, and process the error signal and quantized digital components. The encoder operable to process the error signal by processing delayed audio signals and Q band values, determining the potential for pre-echo events from the Q band values, and determining scale factors and MDCT block sizes for the potential for pre-echo events.

Abstract: A method and device for automatically increasing the spectral bandwidth of an audio signal including generating a “mapping” (or “prediction”) matrix based on the analysis of a reference wideband signal and a reference narrowband signal, the mapping matrix being a transformation matrix to predict high frequency energy from a low frequency energy envelope, generating an energy envelope analysis of an input narrowband audio signal, generating a resynthesized noise signal by processing a random noise signal with the mapping matrix and the envelope analysis, high-pass filtering the resynthesized noise signal, and summing the high-pass filtered resynthesized noise signal with the original an input narrowband audio signal. Other embodiments are disclosed.

Abstract: The disclosure discloses a voice signal enhancing method and device, which divide a voice signal at the present scene into multiple frame signals based on a preset time interval; feed multiple frame signals into a trained neural network based on a preset step size, perform convolution operations on multiple frame signals through skip-connected convolutional layers to obtain multiple enhanced frame signals; superpose each enhanced frame signal according to the time domain of each enhanced frame signal to obtain an enhanced voice signal. Compared with the prior art, the present disclosure automatically enhances voice signals through the neural network without manual interference, so the effects and the application scenes of voice enhancement is not necessary to be limited by the preset method and method designers, thereby reducing the occurrence frequency of signal distortion and extra noises, which in turn improves the effects of the voice signal enhancement.

Abstract: A method, system, and computer readable medium for decomposing an audio signal into different isolated sources. The techniques and mechanisms convert an audio signal into K input spectrogram fragments. The fragments are sent into a deep neural network to isolate for different sources. The isolated fragments are then combined to form full isolated source audio signals.

Abstract: A speech synthesis model training device includes one or more hardware processors configured to perform the following. Storing, in a speech corpus storing unit, speech data, and pitch mark information and context information of the speech data. From the speech data, analyzing acoustic feature parameters at each pitch mark timing in pitch mark information. From the acoustic feature parameters analyzed, training a statistical model which has a plurality of states and which includes an output distribution of acoustic feature parameters including pitch feature parameters and a duration distribution based on timing parameters.

Abstract: Disclosed are methods and systems to help disambiguate channel identification in a scenario where a video fingerprint of media content matches multiple reference video fingerprints corresponding respectively with multiple different channels. Given such a multi-match situation, an entity could disambiguate based on an audio component of the media content, such as by further determining that an audio fingerprint of the media content at issue matches an audio fingerprint of just one of the multiple channels, thereby establishing that that is the channel on which the media content being rendered by the media presentation device is arriving.

Abstract: The present disclosure describes various examples of a method, an apparatus, and a computer readable medium for signaling synchronization block patterns in wireless communications (e.g., 5th Generation New Radio (5G NR)). For example, one of the methods described may include receiving, by a user equipment (UE), a message including information of a configuration. The configuration includes at least a group of repetitions of one or more synchronization signal (SS) blocks in an SS burst set, and the repetitions of the one or more SS blocks are configured into at least two groups. The method may further include determining, by the UE, which group of the at least two groups to search for during a synchronous neighbor cell search based on the information and at least one condition at the UE.

Abstract: An electronic device includes a display, a microphone, a memory, a communication circuit, and a processor. The processor is configured to display a user interface for adjusting voice recognition sensitivity of each of a plurality of voice recognizing devices configured to start a voice recognition service in response to the same start utterance, through the display, to transmit a value of the changed sensitivity to at least part of the plurality of voice recognizing devices when the voice recognition sensitivity is changed through the user interface, to transmit a signal for waiting to receive a first utterance of a user, to the plurality of voice recognizing devices, to receive utterance information corresponding to the first utterance from the plurality of voice recognizing devices, and to update the user interface based on the utterance information.

Abstract: A data generating apparatus for generating noise environment noisy data is disclosed. The data generating apparatus according to the present application comprises a signal conversion unit configured to convert each of a noisy signal obtained in real environment and an original sound signal for the noisy signal into a noisy signal spectrum and an original sound signal spectrum in a short-time frequency domain; and a noisy signal generation training unit configured to train deep neural network to output the noisy signal spectrum corresponding to each short-time using the original sound signal spectrum as an input.

Abstract: Methods, apparatus, systems and articles of manufacture are disclosed for watermarking using starting phase modulation. An example apparatus includes memory, and processor circuitry to execute instructions to at least determine a first analyzed phase value for a watermark component of a watermarked media signal at a first time, determine a sum of differences for analyzed phase values with respect to a first one of a plurality of possible starting phase values, the analyzed phase values associated with the watermarked media signal, the analyzed phase values including the first analyzed phase value, in response to the sum of differences satisfying a threshold, decode a first data value corresponding to the first one of the possible starting phase values, and determine a watermark payload based on the first data value.

Abstract: The present disclosure relates to a device for processing an audio signal. The device may include a first acoustic-electric transducer and a second acoustic-electric transducer. The first acoustic-electric transducer may have a first frequency response, and may be configured to detect the audio signal and generate a first sub-band signal according to the detected audio signal. The second acoustic-electric transducer may have a second frequency response, the second frequency response being different from the first frequency response. The second acoustic-electric transducer may be configured to detect the audio signal and generate a second sub-band signal according to the detected audio signal.

Abstract: An echo and near-end cross-talk (NEXT) cancellation system includes a time-domain processing module and a frequency-domain processing module. The time-domain processing module is configured to receive an unprocessed signal after an analog-to-digital conversion, remove at least one time-domain dominant component of interference from the unprocessed signal, and accordingly cancel a time-domain processed signal. The frequency-domain processing module is connected to the time-domain processing module, and configured to receive the time-domain processed signal, cancel at least one frequency-domain component of the interference from the unprocessed signal, and accordingly generate a processed signal.

Abstract: A method, system, and computer readable medium for decomposing an audio signal into different isolated sources. The techniques and mechanisms convert an audio signal into K input spectrogram fragments. The fragments are sent into a deep neural network to isolate for different sources. The isolated fragments are then combined to form full isolated source audio signals.

Abstract: To provide a bandwidth extension method which allows reduction of computation amount in bandwidth extension and suppression of deterioration of quality in the bandwidth to be extended. In the bandwidth extension method: a low frequency bandwidth signal is transformed into a QMF domain to generate a first low frequency QMF spectrum; pitch-shifted signals are generated by applying different shifting factors on the low frequency bandwidth signal; a high frequency QMF spectrum is generated by time-stretching the pitch-shifted signals in the QMF domain; the high frequency QMF spectrum is modified; and the modified high frequency QMF spectrum is combined with the first low frequency QMF spectrum.

Abstract: The present disclosure relates to a device for processing an audio signal. The device may include a first acoustic-electric transducer and a second acoustic-electric transducer. The first acoustic-electric transducer may have a first frequency response, and may be configured to detect the audio signal and generate a first sub-band signal according to the detected audio signal. The second acoustic-electric transducer may have a second frequency response, the second frequency response being different from the first frequency response. The second acoustic-electric transducer may be configured to detect the audio signal and generate a second sub-band signal according to the detected audio signal.

Abstract: A method for ascertaining features in an environment of at least one mobile unit for implementation of a localization and/or mapping by a control unit. In the course of the method, sensor measurement data of the environment are received, the sensor measurement data received are transformed by an alignment algorithm into a cost function and a cost map is generated with the aid of the cost function, a convergence map is generated based on the alignment algorithm. At least one feature is extracted from the cost map and/or the convergence map and stored, the at least one feature being provided in order to optimize a localization and/or mapping. A control unit, a computer program, and a machine-readable storage medium are also described.

Abstract: According to one embodiment, a signal generation device includes one or more processors. The processors convert an acoustic signal and output amplitude and phase at a plurality of frequencies. The processors, for each of a plurality of nodes of a hidden layer included in a neural network that treats the amplitude and the phase as input, obtain frequency based on a plurality of weights used in arithmetic operation of the node. The processors generate an acoustic signal based on the plurality of obtained frequencies and based on amplitude and phase corresponding to each of the plurality of nodes.

Abstract: A decoding device includes: a separating unit separating first encoded data, a spectrum including a low-band spectrum of audio signals having been encoded, and second encoded data, a high-band spectrum of a higher band having been encoded, based on the first encoded data; a first decoding unit decoding the first encoded data and generating a first decoded spectrum; a first amplitude normalizer dividing amplitude of the first decoded spectrum into sub-bands, normalizing the spectrum of each sub-band by the largest amplitude of the first decoded spectrum within each sub-band, and generating a normalized spectrum; an addition unit adding noise spectrum to the normalized spectrum and generating a noise-added normalized spectrum; a second decoding unit decoding the second encoded data using the noise-added normalized spectrum, and generating a second noise-added spectrum; and a converter performing time-frequency conversion regarding a spectrum coupled based on the first decoded spectrum and second noise-added spe

Abstract: A system and method of recording and transmitting compressed audio signals over a network is disclosed. The end node device first converts the audio signal to a spectrogram, which is commonly used by machine learning algorithms to perform speech recognition. The end node device then compresses the spectrogram prior to transmission. In certain embodiments, the compression is performed using Discrete Cosine Transforms (DCT). Furthermore, in some embodiments, the DCT is performed on the difference between two columns of the spectrogram. Further, in some embodiments, a function that replaces values below a predetermined threshold with zeroes in the Encoded Spectrogram is utilized. These functions may be performed in hardware or software.

Abstract: The invention provides methods and devices for outputting a stereo audio signal having a left channel and a right channel. The apparatus includes a demultiplexer, decoder, and upmixer. The upmixer is configured operate either in a prediction mode or a non-prediction mode based on a parameter encoded in the audio bitstream.

Abstract: Systems and methods for identifying patterns of symbols in standardized system diagrams are disclosed. Disclosed implementations obtain or synthetically generate a symbol recognition training data set including multiple training images, generate a symbol recognition model based on the symbol recognition training data set, obtain an image comprising a pattern of symbols, group symbols into process loops based on the logical relationships captured by process loop identification algorithm, apply a character classification model to image contours to identify the characters and group characters into tags via hierarchical clustering, and store the identified tags, symbols and identified process loops in a relational database.

Abstract: A speech processing device of an embodiment includes a spectrum parameter calculation unit, a phase spectrum calculation unit, a group delay spectrum calculation unit, a band group delay parameter calculation unit, and a band group delay compensation parameter calculation unit. The spectrum parameter calculation unit calculates a spectrum parameter. The phase spectrum calculation unit calculates a first phase spectrum. The group delay spectrum calculation unit calculates a group delay spectrum from the first phase spectrum based on a frequency component of the first phase spectrum. The band group delay parameter calculation unit calculates a band group delay parameter in a predetermined frequency band from a group delay spectrum. The band group delay compensation parameter calculation unit calculates a band group delay compensation parameter to compensate a difference between a second phase spectrum reconstructed from the band group delay parameter and the first phase spectrum.

Abstract: A method for generating a psychoacoustic model from an audio signal transforms a block of samples of an audio signal into a frequency spectrum comprising frequency components. From this frequency spectrum, it derives group masking energies. These group masking energies each correspond to a group of neighboring frequency components in the frequency spectrum. For a group of frequency components, the method allocates the group masking energy to the frequency components in the group in proportion to energy of the frequency components within the group to provide adapted mask energies for the frequency components within the group, the adapted mask energies providing masking thresholds for the psychoacoustic model of the audio signal.

Abstract: The invention relates to a method, a circuit, and an apparatus for detecting distortion in spread spectrum signals. An edge in a spread spectrum clock signal is identified based on a reference clock signal. The edge data is then provided to a set of counters which are incremented corresponding to an identified edge. Each bit of a respective output of the counters are provided to a respective OR gate of a set of OR gates. An OR gate from the set of OR gates corresponding to a selected bit then outputs an indication of whether distortion exists in the spread spectrum clock signal.

Abstract: An audio encoding method includes dividing an energy spectrum of a current audio frame into P FFT energy spectrum coefficients; determining a minimum bandwidth of distribution, on spectrum, of first-preset-proportion energy of the current audio frame according to the energy of the P FFT energy spectrum coefficients of the current audio frame, wherein the minimum bandwidth of distribution, on spectrum, of first preset proportion energy of the current audio frame indicates sparseness of distribution, on the spectrum, of energy of the current audio frame; and determining to use a linear-prediction-based encoding method to encode the current audio frame in response to the minimum bandwidth of distribution is greater than a first preset value.

Abstract: An electronic device for identity verification includes a memory and a processor; the system of identity verification is stored in the memory, and executed by the processor to implement: after receiving current voice data of a target user, carrying out framing processing on the current voice data according to preset framing parameters to obtain multiple voice frames; extracting preset types of acoustic features in all the voice frames by using a predetermined filter, and generating multiple observed feature units corresponding to the current voice data according to the extracted acoustic features; pairwise coupling all the observed feature units with pre-stored observed feature units respectively to obtain multiple groups of coupled observed feature units; inputting the multiple groups of coupled observed feature units into a preset type of identity verification model generated by pre-training to carry out the identity verification on the target user.

Abstract: Provided are a voice denoising method and apparatus, a server and a storage medium. The voice denoising method comprises: acquiring voice signals synchronously collected by an acoustic microphone and a non-acoustic microphone (S100); carrying out voice activity detection according to the voice signal collected by the non-acoustic microphone to obtain a voice activity detection result (S110); and according to the voice activity detection result, denoising the voice signal collected by the acoustic microphone to obtain a denoised voice signal (S120). The effect of denoising can be enhanced, and the quality of voice signals can be improved.

Abstract: The present disclosure proposes a speech processing method and a cloud-based speech processing apparatus. The speech processing method includes: acquiring a piece of speech to be recognized collected by a terminal; performing a speech recognition on the piece of speech to be recognized; detecting whether the piece of speech to be recognized ends during the speech recognition; and feeding back a recognized result of the piece of speech to be recognized to the terminal when it is detected that the piece of speech to be recognized ends.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for coding speech using neural networks. One of the methods includes obtaining a bitstream of parametric coder parameters characterizing spoken speech; generating, from the parametric coder parameters, a conditioning sequence; generating a reconstruction of the spoken speech that includes a respective speech sample at each of a plurality of decoder time steps, comprising, at each decoder time step: processing a current reconstruction sequence using an auto-regressive generative neural network, wherein the auto-regressive generative neural network is configured to process the current reconstruction to compute a score distribution over possible speech sample values, and wherein the processing comprises conditioning the auto-regressive generative neural network on at least a portion of the conditioning sequence; and sampling a speech sample from the possible speech sample values.

Abstract: A speech processing method for estimating a pitch frequency includes: specifying, for each determination result of a speech-like-frame, a fundamental sound by using a plurality of local maximum values included in a spectrum of a respective frame determined as the speech-like-frame; obtaining a learned value by performing learning processing on a magnitude of the fundamental sound specified from each determination result of the speech-like-frame, the learned value including an average value and a variance of the magnitude of the fundamental sound specified from each determination result of the speech-like-frame; and executing a detection process by using the learned value, the detection process including detecting a pitch frequency of the respective frame determined as the speech-like-frame by using a threshold, the threshold being obtained by subtracting the variance included in the learned value from the average value included in the learned value.