Methods for reconstructing an audio signal
A system configured to reconstruct audio signals. The system may identify missing audio samples due to packet loss or detect distortion caused by audio clipping and may reconstruct the audio data. The system may employ a forward-looking neural network that recursively predicts audio samples based on previous audio samples and/or a backward-looking neural network that recursively predicts audio samples based on subsequent audio samples. The system may generate audio data using only the forward-looking neural network for low latency applications or may generate audio data using both neural networks for mid to high latency applications. To reduce distortion in output audio data, the system may generate the audio data by cross-fading between outputs of the neural networks and/or may cross-fade between the generated audio data and the input audio data.
1. A computer-implemented method, comprising:
receiving input audio data comprising a plurality of audio samples;
detecting distortion in a first portion of the input audio data associated with a first period of time, the distortion caused by at least one of the plurality of audio samples missing from the input audio data or a magnitude value of one or more of the plurality of audio samples being equal to a saturation threshold value;
determining that a second portion of the input audio data following the first portion is not distorted, the second portion corresponding to a second period of time that begins at a first time;
performing, based on a magnitude of signal values of the input audio data, a quantization process to generate first audio data by mapping the signal values of the input audio data to discrete states corresponding to respective quantization intervals;
generating, based on the first audio data, two or more first audio data predictions corresponding to at least part of the first period of time, the two or more first audio data predictions determined using a first generative model that receives the first audio data as input features and predicts a magnitude of signal values for audio samples recursively in a first direction in time;
determining a first audio sample in the first audio data corresponding to the first time;
determining a magnitude value associated with the first audio sample;
selecting, based on at least the magnitude value associated with the first audio sample, a first data prediction of the two or more first audio data predictions;
generating, based on the first data prediction, second audio data corresponding to at least part of the first period of time;
generating, based on at least the first audio data and the second audio data, output audio data, the output audio data including the second audio data followed by a third portion of the first audio data that includes the first audio sample; and
doing at least one of (a) causing audio corresponding to the output audio data to be output by at least one speaker, or (b) causing a function corresponding to a voice command represented by the output audio data to be executed.
2. The computer-implemented method of claim 1 , further comprising:
determining the third portion of the first audio data, the third portion including the first audio sample and corresponding to the second period of time;
selecting the first data prediction based on the third portion and the magnitude value associated with the first audio sample;
generating the output audio data, the output audio data cross-fading from the second audio data to the third portion of the first audio data; and
training, using the output audio data, a neural network included in the first generative model.
3. The computer-implemented method of claim 1 , further comprising:
generating, based on the first audio data, a second data prediction corresponding to at least part of the first period of time, the second data prediction determined using a second generative model that predicts a magnitude of signal values for audio samples recursively in a second direction in time opposite the first direction beginning at the first time; and
generating, based on the first data prediction and the second data prediction, the second audio data, the second audio data cross-fading between the first data prediction and the second data prediction, the second audio data corresponding to the first period of time.
4. The computer-implemented method of claim 1 , further comprising:
performing, based on a magnitude of signal values for the input audio data, the quantization process to generate the first audio data, the quantization process having nonuniform quantization intervals, wherein a first quantization interval corresponds to a first range of signal values and a second quantization interval corresponds to a second range of signal values that is smaller than the first range.
5. A computer-implemented method, comprising:
receiving input audio data comprising a plurality of audio samples;
detecting distortion in a first portion of the input audio data associated with a first period of time;
determining that a second portion of the input audio data following the first portion is not distorted, the second portion corresponding to a second period of time that begins at a first time;
performing a quantization process on the input audio data to generate first audio data by mapping signal values of the input audio data to discrete states corresponding to respective quantization intervals;
generating, based on the first audio data, two or more first audio data predictions corresponding to at least part of the first period of time, the two or more first audio data predictions determined using a first generative model that receives the first audio data as input features and predicts audio samples recursively in a first direction in time;
generating, based on the two or more first audio data predictions, second audio data corresponding to at least part of the first period of time;
generating, based on at least the first audio data and the second audio data, output audio data; and
doing at least one of (a) causing audio corresponding to the output audio data to be output by at least one speaker, or (b) causing a function corresponding to a voice command represented by the output audio data to be executed.
6. The computer-implemented method of claim 5 , further comprising:
determining a first audio sample in the first audio data, the first audio sample corresponding to the first time;
determining a magnitude value associated with the first audio sample;
generating the second audio data by selecting, based on at least the magnitude value associated with the first audio sample, a first data prediction of the two or more first audio data predictions, the second audio data corresponding to the first period of time; and
generating the output audio data, the output audio data including the second audio data followed by a third portion of the first audio data that includes the first audio sample.
7. The computer-implemented method of claim 5 , further comprising:
determining a third portion of the first audio data, the third portion corresponding to at least part of the second period of time;
generating the second audio data by selecting, based on the third portion of the first audio data, a first data prediction of the two or more first audio data predictions, the second audio data corresponding to the first period of time and at least part of the second period of time;
generating, based on at least part of the second audio data and the third portion of the first audio data, third audio data, the third audio data cross-fading from the second audio data to the third portion of the first audio data; and
generating the output audio data, the output audio data including a part of the second audio data followed by the third audio data and part of the third portion of the first audio data.
8. The computer-implemented method of claim 5 , further comprising:
determining a third portion of the first audio data, the third portion corresponding to at least part of the second period of time;
generating the second audio data by averaging signal values of audio samples included in the two or more first audio data predictions, the second audio data corresponding to the first period of time and at least part of the second period of time;
generating, based on at least part of the second audio data and the third portion of the first audio data, third audio data, the third audio data cross-fading from the second audio data to the third portion of the first audio data; and
generating the output audio data, the output audio data including part of the second audio data followed by the third audio data and part of the third portion of the first audio data.
9. The computer-implemented method of claim 5 , further comprising:
selecting a first data prediction of the two or more first audio data predictions;
generating, based on the first audio data, a second data prediction corresponding to at least part of the first period of time, the second data prediction determined using a second generative model that predicts audio samples recursively in a second direction in time opposite the first direction beginning at the first time; and
generating, based on the first data prediction and the second data prediction, the second audio data, the second audio data cross-fading between the first data prediction and the second data prediction, the second audio data corresponding to the first period of time.
10. The computer-implemented method of claim 5 , further comprising:
selecting a first data prediction of the two or more first audio data predictions;
generating, based on the first audio data, two or more second audio data predictions corresponding to at least part of the first period of time, the two or more second audio data predictions determined using a second generative model that predicts audio samples recursively in a second direction in time opposite the first direction beginning at the first time;
selecting a second data prediction of the two or more second audio data predictions; and
generating, based on the first data prediction and the second data prediction, the second audio data, the second audio data cross-fading between the first data prediction and the second data prediction, the second audio data corresponding to the first period of time.
11. The computer-implemented method of claim 5 , further comprising:
generating, based on the first audio data, two or more second audio data predictions corresponding to at least part of the first period of time, the two or more second audio data predictions determined using a second generative model that predicts audio samples recursively in a second direction in time opposite the first direction beginning at the first time;
determining a plurality of similarity metrics, wherein the determining the plurality of similarity metrics further comprises:
determining a first similarity metric between a first data prediction of the two or more first audio data predictions and a second data prediction of the two or more second audio data predictions, and
determining a second similarity metric between the first data prediction of the two or more first audio data predictions and a third data prediction of the two or more second audio data predictions;
determining that the second similarity metric is a highest similarity metric of the plurality of similarity metrics; and
generating the second audio data, the second audio data cross-fading between the first data prediction and the third data prediction, the second audio data corresponding to the first period of time.
12. The computer-implemented method of claim 5 , further comprising:
performing, based on a magnitude of signal values for the input audio data, the quantization process to generate the first audio data, the quantization process having nonuniform quantization intervals, wherein a first quantization interval corresponds to a first range of signal values and a second quantization interval corresponds to a second range of signal values that is smaller than the first range.
13. The computer-implemented method of claim 5 , wherein:
the distortion is caused by at least one of the plurality of audio samples missing from the input audio data or a magnitude value of one or more of the plurality of audio samples being equal to a saturation threshold value.
14. A system comprising:
at least one processor; and
memory including instructions operable to be executed by the at least one processor to perform a set of actions to configure the system device to:
receive input audio data comprising a plurality of audio samples;
detect distortion in a first portion of the input audio data associated with a first period of time;
determine that a second portion of the input audio data following the first portion is not distorted, the second portion corresponding to a second period of time that begins at a first time;
perform a quantization process on the input audio data to generate first audio data by mapping signal values of the input audio data to discrete states corresponding to respective quantization intervals;
generate, based on the first audio data, two or more first audio data predictions corresponding to at least part of the first period of time, the two or more first audio data predictions determined using a first generative model that receives the first audio data as input features and predicts audio samples recursively in a first direction in time;
generate, based on the two or more first audio data predictions, second audio data corresponding to at least part of the first period of time;
generate, based on at least the first audio data and the second audio data, output audio data; and
do at least one of (a) cause audio corresponding to the output audio data to be output by at least one speaker, or (b) cause a function corresponding to a voice command represented by the output audio data to be executed.
15. The system of claim 14 , wherein the memory includes additional instructions operable to be executed by the at least one processor to further configure the system to:
determine a first audio sample in the first audio data, the first audio sample corresponding to the first time;
determine a magnitude value associated with the first audio sample;
generate the second audio data by selecting, based on at least the magnitude value associated with the first audio sample, a first data prediction of the two or more first audio data predictions, the second audio data corresponding to the first period of time; and
generate the output audio data, the output audio data including the second audio data followed by a third portion of the first audio data that includes the first audio sample.
16. The system of claim 14 , wherein the memory includes additional instructions operable to be executed by the at least one processor to further configure the system to:
determine a third portion of the first audio data, the third portion corresponding to at least part of the second period of time;
generate the second audio data by selecting, based on the third portion of the first audio data, a first data prediction of the two or more first audio data predictions, the second audio data corresponding to the first period of time and at least part of the second period of time;
generate, based on at least part of the second audio data and the third portion of the first audio data, third audio data, the third audio data cross-fading from the second audio data to the third portion of the first audio data; and
generate the output audio data, the output audio data including a part of the second audio data followed by the third audio data and part of the third portion of the first audio data.
17. The system of claim 14 , wherein the memory includes additional instructions operable to be executed by the at least one processor to further configure the system to:
determine a third portion of the first audio data, the third portion corresponding to at least part of the second period of time;
generate the second audio data by averaging signal values of audio samples included in the two or more first audio data predictions, the second audio data corresponding to the first period of time and at least part of the second period of time;
generate, based on at least part of the second audio data and the third portion of the first audio data, third audio data, the third audio data cross-fading from the second audio data to the third portion of the first audio data; and
generate the output audio data, the output audio data including part of the second audio data followed by the third audio data and part of the third portion of the first audio data.
18. The system of claim 14 , wherein the memory includes additional instructions operable to be executed by the at least one processor to further configure the system to:
select a first data prediction of the two or more first audio data predictions;
generate, based on the first audio data, a second data prediction corresponding to at least part of the first period of time, the second data prediction determined using a second generative model that predicts audio samples recursively in a second direction in time opposite the first direction beginning at the first time; and
generate, based on the first data prediction and the second data prediction, the second audio data, the second audio data cross-fading between the first data prediction and the second data prediction, the second audio data corresponding to the first period of time.
19. The system of claim 14 , wherein the memory includes additional instructions operable to be executed by the at least one processor to further configure the system to:
select a first data prediction of the two or more first audio data predictions;
generate, based on the first audio data, two or more second audio data predictions corresponding to at least part of the first period of time, the two or more second audio data predictions determined using a second generative model that predicts audio samples recursively in a second direction in time opposite the first direction beginning at the first time;
select a second audio prediction of the two or more second audio data predictions; and
generate, based on the first data prediction and the second data prediction, the second audio data, the second audio data cross-fading between the first data prediction and the second data prediction, the second audio data corresponding to the first period of time.
20. The system of claim 14 , wherein the memory includes additional instructions operable to be executed by the at least one processor to further configure the system to:
generate, based on the first audio data, two or more second audio data predictions corresponding to at least part of the first period of time, the two or more second audio data predictions determined using a second generative model that predicts audio samples recursively in a second direction in time opposite the first direction beginning at the first time;
determine a plurality of similarity metrics, wherein determining the plurality of similarity metrics further comprises:
determining a first similarity metric between a first data prediction of the two or more first audio data predictions and a second data prediction of the two or more second audio data predictions, and
determining a second similarity metric between the first data prediction of the two or more first audio data predictions and a third data prediction of the two or more second audio data predictions;
determine that the second similarity metric is a highest similarity metric of the plurality of similarity metrics; and
generate the second audio data, the second audio data cross-fading between the first data prediction and the third data prediction, the second audio data corresponding to the first period of time.
21. The system of claim 14 , wherein the memory includes additional instructions operable to be executed by the at least one processor to further configure the system to:
perform, based on a magnitude of signal values for the input audio data, the quantization process to generate the first audio data, the quantization process having nonuniform quantization intervals, wherein a first quantization interval corresponds to a first range of signal values and a second quantization interval corresponds to a second range of signal values that is smaller than the first range.
22. The system of claim 14 , wherein the input audio data corresponds to an utterance, and the memory includes additional instructions operable to be executed by the at least one processor to further configure the system to:
cause a voice command represented by the utterance to be determined; and
cause the function to be performed based at least in part on the voice command.
23. The system of claim 14 , wherein the memory includes instructions operable to be executed by the at least one processor to configure the system to:
cause the audio corresponding to the output audio data to be output by the at least one speaker.