Amazon Technologies, Inc.

Arvindh Krishnaswamy

Mehmet Umut Isik

Karim Helwani

Jean-Marc Valin

Srikanth Venkata Tenneti

Fangzhou Cheng

Neerad Dilip Phansalkar

Siddhartha Shankara Rao

John Joseph Dunne

Michael Klingbeil

Michael Mark Goodwin

Aparna Pandey

Erfan Soltanmohammadi

Hassan Sawaf

Andrew Paul Breen

Yaser Al-Onaizan

Roberto Barra-Chicote

Richard Newman

Marcello Federico

Robert Enyedi

Shrikant Venkataramani

Sid Shankara Rao

An audio assessment machine learning model, for example including one or more neural networks, may be trained for human audio quality assessment estimation, wherein the training comprises performing comparisons of a plurality of training machine learning audio quality assessments of training audio content to a plurality of human audio quality assessments of the training audio content and adjusting the audio assessment machine learning model based on the comparisons. After the training, a first audio analysis of first audio content may be performed by the audio assessment machine learning model. A first machine learning audio quality assessment of the first audio content may be provided, by the audio assessment machine learning model, based on the first audio analysis. The first machine learning audio quality assessment may include a quality score for the first audio content and a quality degradation reason for the first audio content.

Machine learning model that estimates human audio quality assessments

Systems, methods, and apparatuses for selecting a model are described. A method includes receiving a request to perform model selection and evaluating multiple models by generating various metrics for each trained model. These metrics include a forewarning time (how much in advance of a failure an alert can be raised), an event recall metric (how many failure events were alerted to in advance), an event precision metric (ratio of true and false positives, and an area under a receiver operating characteristic (ROC) curve. For each trained model, a weighted harmonic mean is calculated from these metrics. A model is then selected based on the calculated means and a report on the selected model is generated and provided.

Prognostics and health management service

Implementations for compositing two input signals to form a higher quality signal are described. A first input signal is received from a first input device and a second input signal is received from a second input device. The first input signal and the second input signal are combined into a composite input signal. It is then determined that the composite input signal has a higher quality than either the first input signal or the second input signal individually. Based on that determination, the composite input signal is selected for use by the media conferencing service as part of a media conference.

Automatic signal composition for media conferences

Real-time audio enhancement for a target speaker may be performed. An embedding of a sample of speaker audio is created using a trained neural network that performs voice identification. The embedding is then concatenated with the input features of a trained machine learning model for audio enhancement. The audio enhancement model can recognize and enhance a target speaker's speech in a real-time implementation, as the embedding is in the same feature space of the audio enhancement model.

Real-time target speaker audio enhancement

A group masked autoencoder may be implemented for anomaly detection. An autoencoder network model may be trained without supervision and applied to output an estimated joint probability distribution of normality for a group of frames of time series data. The estimated joint probability distribution may be used to determine an anomaly score for the time series data. An anomaly may be detected according to the anomaly score and a result that indicates a detected anomaly may be provided.

Group masked autoencoder for anomaly detection

Systems, methods, and apparatuses for detecting anomalies using clusters are described. In some examples, a method includes receiving a request to perform anomaly detection using a plurality of clusters; receiving a data point; determining when the received data point is a part of one of the plurality of clusters utilizing a distance to centers of the one or more clusters, wherein: when the received data point is determined to belong to a normal cluster, assigning the received data point to the determined cluster, updating the cluster, and updating a history for the cluster, when the received data point is determined to belong to an anomalous cluster, raising an anomaly, updating the cluster, and updating a history for the cluster, and when the received data point is determined to not belong to any cluster, raising an anomaly.

Implementations for selecting an input device based on characteristics of the input signals from those input devices are described. A first input signal is received from a first input device of a participant device participating in a media conference and a second input signal is received from a second input device of the participant device. A first characteristic of the first input signal and a second characteristic of the second input signal are determined. The first characteristic is compared to the second characteristic. It is determined that a quality of the second input signal is greater than a quality of the first input signal based on comparing the first characteristic to the second characteristic. The second input device is selected based on determining that the quality of the second input signal is greater than the quality of the first input signal.

Automatic input device selection for media conferences

Techniques for training and using a machine learning model for estimation of reverberation in a multi-task learning framework are described. According to some embodiments, the multi-task learning framework improves the performance of the machine learning model by estimating the amount of reverberation present in an input audio recording as a secondary task to the primary task of generating a clean speech portion of the input audio recording. In one embodiment, a model architecture is selected that takes a noisy reverberant recording as an input and outputs an estimate of a clean (e.g., de-reverberated) signal, an estimate of noise (e.g., background noise), and an estimate of the reverb only portion, with the secondary task of estimating the reverb only portion acting as a regularizer that improves the machine learning model's performance in enhancing the reverberant (e.g., and noisy) input speech.

Speech enhancement machine learning model for estimation of reverberation in a multi-task learning framework

Audio processing may be performed with a convolutional neural network that includes positional embeddings. Audio data may be received at an audio processing system. A convolutional neural network that concatenates frequency-positional embeddings at an input layer may be used to process the audio data. A result of processing the audio data through the convolutional neural network may be used to perform an audio processing task.

Convolutional neural network with positional embeddings for audio processing

Joint noise and echo suppression may be performed for enhancing two-way audio communications. Audio data is captured at a communication device and audio data transmitted to the communication device from another communication device are used as input features to a trained machine learning model that uses the transmitted audio data as a reference signal to eliminate residual echo in the captured audio data when also suppressing noise in the captured audio data.

Joint noise and echo suppression for two-way audio communication enhancement

Techniques for the generation of dubbed audio for an audio/video are described. An exemplary approach is to receive a request to generate dubbed speech for an audio/visual file; and in response to the request to: extract speech segments from an audio track of the audio/visual file associated with identified speakers; translate the extracted speech segments into a target language; determine a machine learning model per identified speaker, the trained machine learning models to be used to generate a spoken version of the translated, extracted speech segments based on the identified speaker; generate, per translated, extracted speech segment, a spoken version of the translated, extracted speech segments using a trained machine learning model that corresponds to the identified speaker of the translated, extracted speech segment and prosody information for the extracted speech segments; and replace the extracted speech segments from the audio track of the audio/visual file with the spoken versions spoken version of the translated, extracted speech segments to generate a modified audio track.

Multilingual speech translation with adaptive speech synthesis and adaptive physiognomy

Post-filtering may be performed for ratio masks as part of audio enhancement. Audio data may be received. A machine learning model may be applied to generate gain values for different spectrum bands of the audio data. The gain values may then be modified using an envelope post-filter according to a monotonically increasing function applied to the gain values to produce modified gain values used to generate an enhanced version of the audio data.

Growing community of inventors

Palo Alto, CA, United States of America

Arvindh Krishnaswamy