# US 2019/0172479 A1 · Devices and methods for evaluating speech quality

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2019/0172479 A1
- Document ID: US20190172479A1
- Status: Published application
- Kind: A1
- Application number: 16/270,274
- Filed: 2019-02-07
- Published: 2019-06-06
- Earliest claimed: 2016-08-09
- Assignees: Huawei Technologies Co., Ltd.
- Applicants: Huawei Technologies Co., Ltd.
- Inventors: Wei XIAO; Mona HAKAMI; Willem Bastiaan KLEIJN
- Web page: https://patentgrep.com/patent/US20190172479A1

## CPC

- G10L 25/60: for measuring the quality of voice signals
- G06F 17/18: for evaluating statistical data {, e.g. average values, frequency distributions, probability functions, regression analysis (forecasting specially adapted for a specific administrative, business or logistic context G06Q10/04)}
- G10L 15/02: Feature extraction for speech recognition; Selection of recognition unit
- G10L 25/03: characterised by the type of extracted parameters
- G10L 25/30: using neural networks
- G10L 25/69: for evaluating synthetic or decoded voice signals
- G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

The disclosure relates to an apparatus for determining a quality score (MOS) for an audio signal sample, the apparatus comprising: an extractor configured to extract a feature vector from the audio signal sample, wherein the feature vector comprises a plurality of feature values and wherein each feature value is associated to a different feature of the feature vector; a pre-processor configured to pre-process a feature value of the feature vector based on a cumulative distribution function associated to the feature represented by the feature value to obtain a pre-processed feature value; and a processor configured to implement a neural network and to determine the quality score (MOS) for the audio signal sample based on the pre-processed feature value and a set of neural network parameters for the neural network associated to the cumulative distribution function.

## Claims

### Claim 1

An apparatus for determining a quality score for an audio signal sample, the apparatus comprising:
a pre-processor configured to pre-process a feature value of a feature vector extracted from the audio signal sample to obtain a pre-processed feature value, wherein the feature vector comprises a plurality of feature values, and wherein each feature value is associated with a feature of the feature vector, wherein pre-processing the feature value is based on a cumulative distribution function (CDFm) associated with the feature represented by the feature value; and
a processor configured to:
implement a neural network, and
determine the quality score for the audio signal sample based on the pre-processed feature value and a set of neural network parameters associated with the CDFm.

### Claim 2 (depends on 1)

The apparatus according to claim 1, wherein the pre-processor is configured to pre-process the feature value by mapping the feature value of the feature to a function value of the CDFm associated with the feature.

### Claim 3 (depends on 1)

The apparatus according to claim 1, wherein the CDFm is a uniform distribution function.

### Claim 4 (depends on 1)

The apparatus according to claim 1, wherein the CDFm is a uniform distribution function on the interval [0, 1].

### Claim 5 (depends on 1)

The apparatus according to claim 1, wherein the processor is configured to determine the quality score for the audio signal sample during a training phase.

### Claim 6 (depends on 1)

The apparatus according to claim 1, wherein the apparatus is configured to determine, based on the pre-processed feature value, the set of neural network parameters for determining the quality score for the audio signal sample.

### Claim 7 (depends on 1)

The apparatus according to claim 1, wherein the processor is configured to implement the neural network as an auto-encoder neural network for unsupervised learning or multi-layer perception OILY) neural network for supervised learning.

### Claim 8

An apparatus for determining a set of neural network parameters for determining a quality score for an audio signal sample, the apparatus comprising:
a pre-processor configured to pre-process a feature value of a feature vector extracted from the audio signal sample to obtain a pre-processed feature value, wherein the feature vector comprises a plurality of feature values, and wherein each feature value is associated with a feature of the feature vector, wherein pre-processing the feature value is based on a cumulative distribution function (CDFm) associated with the feature represented by the feature value to obtain a pre-processed feature value; and
a processor configured to:
implement a neural network, and
determine based on the pre-processed feature value the set of neural network parameters for determining the quality score for the audio signal sample.

### Claim 9 (depends on 8)

The apparatus according to claim 8, wherein the pre-processor is configured to pre-process the feature value by mapping the feature value of the feature to a function value of the CDFm associated with the feature.

### Claim 10 (depends on 8)

The apparatus according to claim 8, wherein the CDFm is a uniform distribution function.

### Claim 11 (depends on 8)

The apparatus according to claim 8, wherein the CDFm. is a uniform distribution function on the interval [0, 1].

### Claim 12 (depends on 8)

The apparatus according to claim 8, wherein the pre-processor is further configured to:
determine a histogram of feature values associated with a first feature of the plurality of features based on feature values associated with the first feature over a plurality of feature vectors; and
determine the CDFm associated with the first feature based on the histogram of feature values.

### Claim 13 (depends on 12)

The apparatus according to claim 12, wherein the pre-processor is further configured to:
determine the CDFm associated with the first feature by summing over histogram bins of the histogram of the feature values.

### Claim 14 (depends on 8)

The apparatus according to claim 8, wherein the pre-processor is further configured to:
determine a Gaussian distribution (μ, σ) for feature values associated with a first feature of the plurality of features based on feature values associated with the first feature over a plurality of feature vectors; and
determine the CDFm associated with the first feature based on the Gaussian distribution (μ, σ) for the feature values.

### Claim 15 (depends on 14)

The apparatus according to claim 14, wherein the pre-processor is further configured to determine the CDFm associated with the first feature by using a mixture of Gaussian distributions (μ, σ) for feature values associated with the first feature of the plurality of features.

### Claim 16 (depends on 14)

The apparatus according to claim 14, wherein the pre-processor is further configured to determine parameters of the Gaussian distribution (μ, σ) or a mixture of the Gaussian distributions for feature values associated with the first feature of the plurality of features by using an expectation-minimization algorithm.

### Claim 17

An apparatus for determining a quality score for an audio signal sample, wherein the apparatus comprises:
a pre-processor configured to pre-process a feature value of the feature vector extracted from the audio signal sample to obtain a pre-processed feature value, wherein the feature vector comprises a plurality of feature values, and wherein each feature value is associated with a feature of the feature vector, wherein processing the feature value is based on a cumulative distribution function (CDFm) associated with the feature represented by the feature value to obtain a pre-processed feature value, and
(i) map the feature value of the feature to a function value of the CDFm associated with the feature to obtain an intermediate feature value, and to map the intermediate feature value to a function value of an inverse distribution function associated with the feature to obtain the pre-processed feature value, wherein the inverse distribution function is an inverse distribution function of the CDFm. associated with the feature, which is different to an original distribution function associated with the feature, or,
(ii) map the feature value of the feature vector to a function value of a combined distribution function to obtain the pre-processed feature value, wherein function values of the distribution function correspond to function values obtainable by firstly mapping the feature value to the CDFm associated with the feature and secondly mapping the resulting intermediate feature value to a functional value of the inverse distribution function associated with the feature; and
a processor configured to:
implement a neural network, and
determine the quality score for the audio signal sample based on the pre-processed feature value and a set of neural network parameters for the neural network associated with the CDFm.

### Claim 18 (depends on 17)

The apparatus according to claim 17, wherein the processor is configured to determine, based on the pre-processed feature value, the set of neural network parameters for determining the quality score for the audio signal sample.