Methods, systems and apparatus, including computer programs encoded on computer storage media. One of the methods includes receiving an audio waveform that includes a respective audio sample for each of a plurality of time steps, processing the audio waveform using an encoder neural network to generate a plurality of feature vectors representing the audio waveform, generating a respective coded representation of each of the plurality of feature vectors using a plurality of vector quantizers that are each associated with a respective codebook of code vectors, wherein the respective coded representation of each feature vector identifies a plurality of code vectors, including a respective code vector from the codebook of each vector quantizer, that define a quantized representation of the feature vector, and generating a compressed representation of the audio waveform by compressing the respective coded representation of each of the plurality of feature vectors.

In various embodiments, a computer-implemented method of training a neural network for creating an output signal of different modality from an input signal is described. In embodiments, the first modality may be a sound signal or a visual image and where the output signal would be a visual image or a sound signal, respectively. In embodiments a model is trained using a first pair of visual and audio networks to train a set of codebooks using known visual signals and the audio signals and using a second pair of visual and audio networks to further train the set of codebooks using the augmented visual signals and the augmented audio signals. Further, the first and the second visual networks are equally weighted and where the first and the second audio networks are equally weighted. In aspects of the present disclosure, the set of codebooks comprise a visual codebook, an audio codebook and a correlation codebook. These codebooks are then used to create an visual image from a sound signal and/or a sound signal from a visual image.

Methods, systems and apparatus, including computer programs encoded on computer storage media. One of the methods includes receiving an audio waveform that includes a respective audio sample for each of a plurality of time steps, processing the audio waveform using an encoder neural network to generate a plurality of feature vectors representing the audio waveform, generating a respective coded representation of each of the plurality of feature vectors using a plurality of vector quantizers that are each associated with a respective codebook of code vectors, wherein the respective coded representation of each feature vector identifies a plurality of code vectors, including a respective code vector from the codebook of each vector quantizer, that define a quantized representation of the feature vector, and generating a compressed representation of the audio waveform by compressing the respective coded representation of each of the plurality of feature vectors.

A method performed by an encoder. The method comprises determining envelope representation residual coefficients as first compressed envelope representation coefficients subtracted from the input envelope representation coefficients. The method comprises transforming the envelope representation residual coefficients into a warped domain so as to obtain transformed envelope representation residual coefficients. The method comprises applying, at least one of a plurality of gain-shape coding schemes on the transformed envelope representation residual coefficients in order to achieve gain-shape coded envelope representation residual coefficients, where the plurality of gain-shape coding schemes have mutually different trade-offs in one or more of gain resolution and shape resolution for one or more of the transformed envelope representation residual coefficients. The method comprises transmitting, over a communication channel to a decoder, a representation of the first compressed envelope representation coefficients, the gain-shape coded envelope representation residual coefficients, and information on the at least one applied gain-shape coding scheme.

A high quality speech is reproduced with a small data amount in speech coding and decoding for performing compression coding and decoding of a speech signal to a digital signal. In speech coding method according to a code-excited linear prediction (CELP) speech coding, a noise level of a speech in a concerning coding period is evaluated by using a code or coding result of at least one of spectrum information, power information, and pitch information, and various excitation codebooks are used based on an evaluation result.

Methods, an encoder and a decoder are configured for transition between frames with different internal sampling rates. Linear predictive (LP) filter parameters are converted from a sampling rate S1 to a sampling rate S2. A power spectrum of a LP synthesis filter is computed, at the sampling rate S1, using the LP filter parameters. The power spectrum of the LP synthesis filter is modified to convert it from the sampling rate S1 to the sampling rate S2. The modified power spectrum of the LP synthesis filter is inverse transformed to determine autocorrelations of the LP synthesis filter at the sampling rate S2. The autocorrelations are used to compute the LP filter parameters at the sampling rate S2.

Methods, systems and apparatus, including computer programs encoded on computer storage media. One of the methods includes receiving an audio waveform that includes a respective audio sample for each of a plurality of time steps, processing the audio waveform using an encoder neural network to generate a plurality of feature vectors representing the audio waveform, generating a respective coded representation of each of the plurality of feature vectors using a plurality of vector quantizers that are each associated with a respective codebook of code vectors, wherein the respective coded representation of each feature vector identifies a plurality of code vectors, including a respective code vector from the codebook of each vector quantizer, that define a quantized representation of the feature vector, and generating a compressed representation of the audio waveform by compressing the respective coded representation of each of the plurality of feature vectors.

An apparatus for decoding an encoded audio signal to obtain a reconstructed audio signal includes a receiving interface for receiving one or more frames comprising information on a plurality of audio signal samples of an audio signal spectrum of the encoded audio signal, and a processor for generating the reconstructed audio signal. The processor is configured to generate the reconstructed audio signal by fading a modified spectrum to a target spectrum, if a current frame is not received by the receiving interface or if the current frame is received by the receiving interface but is corrupted, wherein the modified spectrum includes a plurality of modified signal samples, wherein, for each of the modified signal samples of the modified spectrum, an absolute value of the modified signal sample is equal to an absolute value of one of the audio signal samples of the audio signal spectrum.

An apparatus for selecting one of a first encoding algorithm and a second encoding algorithm includes a filter configured to receive the audio signal, to reduce the amplitude of harmonics in the audio signal and to output a filtered version of the audio signal. First and second estimators are provided for estimating first and second quality measures in the form of SNRs of segmented SNRs associated with the first and second encoding algorithms without actually encoding and decoding the portion of the audio signal using the first and second encoding algorithms. A controller is provided for selecting the first encoding algorithm or the second encoding algorithm based on a comparison between the first quality measure and the second quality measure.

A frame loss compensation processing method and apparatus is presented, where the method includes, when a i.sup.th frame is a lost frame, estimating a spectrum frequency parameter, a pitch period, and a gain of the i.sup.th frame according to at least one of an inter-frame relationship between first N frames of the i.sup.th frame or an intra-frame relationship between first N frames of the i.sup.th frame. A parameter of the i.sup.th frame is determined using the signal correlation between the first N frames, the signal energy stability between the first N frames, intra-frame signal correlation of each frame, and intra-frame signal energy stability of each frame.