Speech quality assessment method and system
View Patent ↗In one embodiment, distortion in a received speech signal is estimated using at least one model trained based on subjective quality assessment data. A speech quality assessment for the received speech signal is then determined based on the estimated distortion.
1. A speech quality assessment method, comprising:
estimating, by a computer, objective frame distortion in a received speech signal at a portion of a speech quality system, the estimated objective frame distortion based on inputting at least one feature vector for each frame of the received speech signal into at least a first model trained based on subjective quality assessment data, the at least one feature vector including articulation power components;
estimating mute distortion caused by mutes in the received speech signal using a second model trained based on the subjective quality assessment data;
determining a speech quality assessment for the received speech signal based on the estimated objective frame distortion and the estimated mute distortion; and
storing the determined speech quality assessment at a portion of the speech quality system,
retraining the first model based on the estimated objective frame distortion and the estimated mute distortion.
2. The method of claim 1 , wherein the objective frame distortion estimating step, further includes:
estimating objective speech distortion in the received speech signal using the first model trained based on subjective quality assessment data.
3. The method of claim 2 , wherein the objective frame distortion estimating step, further includes:
estimating objective background noise distortion in the received speech signal using the first model trained based on subjective quality assessment data.
4. The method of claim 3 , further comprising:
determining an average articulation power and an average non-articulation power from the received speech signal; and wherein
the estimating objective speech distortion step estimates the speech distortion using the determined average articulation power, the determined average non-articulation power and the first model; and
the estimating objective background noise distortion step estimates the background noise distortion using the determined average articulation power, the determined average non-articulation power and the first model.
5. The method of claim 3 , wherein the first model models subjective determination of distortion in speech signals.
6. The method of claim 5 , wherein the first model is a neural network.
7. The method of claim 1 , wherein the estimating mute distortion step, further includes:
detecting mutes in the received speech signal; and
estimating distortion caused by the detected mutes.
8. The method of claim 7 , wherein,
the detecting step detects locations and durations of mutes in the received speech signal; and
the estimating distortion caused by the detected mutes step estimates the mute distortion based on the detected locations and durations.
9. The method of claim 1 , wherein the estimating distortion caused by the detected mutes step estimates the mute distortion such that mutes later in the received speech signal have a greater impact than mutes earlier in the received speech signal.
10. The method of claim 1 , wherein the first model models subjective determination of distortion in speech signals lacking mute distortion and the second model models subjective determination of mute distortion in speech signals.
11. The method of claim 1 , wherein the determining step maps the estimated objective frame distortion to a subjective quality assessment metric.
12. A processing device for speech quality assessment, comprising:
at least one processor to estimate objective frame distortion in a received speech signal, the estimated objective frame distortion based on inputting at least one feature vector for each frame of the received speech signal into at least a first model trained based on subjective quality assessment data, the at least one feature vector including articulation power components, the at least one processor to further estimate mute distortion in the received speech signal using a second model trained based on the subjective quality assessment data; and
a mapping unit mapping the estimated objective frame distortion to a speech quality metric,
wherein the first model is retrained based on the estimated objective frame distortion and the estimated mute distortion.
13. A method of estimating objective frame distortion, comprising:
receiving, by a computer, at least one feature vector for each frame of a received signal at a portion of a speech quality system, the at least one feature vector including articulation power components;
detecting mutes in the received signal;
estimating speech distortion in the received signal using the at least one feature vector for each frame and a first model trained based on subjective quality assessment data;
estimating background noise distortion in the received signal using the at least one feature vector for each frame and the first model trained based on subjective quality assessment data;
estimating mute distortion caused by the detected mutes such that mutes later in the received speech signal have a greater impact than mutes earlier in the received speech signal, using a second model trained based on the subjective quality assessment data;
combining the estimated speech distortion and the estimated background noise distortion to obtain an objective frame distortion estimate;
storing the objective frame distortion estimate at a receiver; and
storing the estimated mute distortion at the receiver,
retraining the first model based on the estimated objective frame distortion and the estimated mute distortion.
14. A method of training a speech quality assessment system, comprising:
training, by a computer, a first distortion estimation path of the system while excluding impact from a second distortion estimation path of the system using first subjective quality assessment data, the first subjective quality assessment data including first speech signals and first associated subjective quality metrics, the first speech signals lacking in mute distortion;
training the second distortion estimating path of the system using second subjective quality assessment data, the second subjective quality assessment data including second speech signals and second associated subjective quality metrics, the second speech signals including mute distortion,
wherein the first distortion estimating path is configured to estimate frame distortion and the second distortion estimating path is configured to estimate mute distortion; and
retraining the first distortion path while including the impact of the second distortion path using the first and second quality assessment data.