Speech feature extraction apparatus, speech feature extraction method, and computer-readable storage medium
A speech feature extraction apparatus 100 includes a voice activity detection unit 103 that drops non-voice frames from frames corresponding to an input speech utterance, and calculates a posterior of being voiced for each frame, a voice activity detection process unit 106 calculates a function value as weights in pooling frames to produce an utterance-level feature, from a given a voice activity detection posterior, and an utterance-level feature extraction unit 112 that extracts an utterance-level feature, from the frame on a basis of multiple frame-level features, using the function values.
1. A speech feature extraction apparatus comprising:
a processor; and
a memory device storing instructions executable by the processor to:
drop non-voice frames from frames corresponding to an input speech utterance, and calculate posterior of voice for the frames;
calculate function values from the posteriors; and
extract an utterance-level feature from the frames based on multiple frame-level features by using the function values as weights for pooling the frames in a pooling layer of a neural network.
2. The speech feature extraction apparatus according to claim 1 , wherein the instructions are executable by the processor to further train extraction of the utterance-level feature to generate utterance-level feature extraction parameters using the multiple frame-level features and the weights.
3. The speech feature extraction apparatus according to claim 2 ,
wherein from function values of the second posteriors used as weights for dropping the non-voice frames are further used to train the extraction.
4. The speech feature extraction apparatus according to claim 2 ,
wherein voice activity detection used for obtaining weights for dropping the non-voice frames is further used to train the extraction.
5. The speech feature extraction apparatus according to claim 1 , wherein the instructions are executable by the processor to further:
calculate second posteriors of voice for the frames; and
wherein function values of the second posteriors are used as weights for dropping the non-voice frames.
6. The speech feature extraction apparatus according to claim 5 ,
wherein voice activity detection is further used for obtaining the weights for dropping the non-voice frames.
7. The speech feature extraction apparatus according to claim 1 ,
wherein the a monotonically increasing and non-linear function defined as one of normalized Odds, and normalized log Odds, is used to calculate the function values from the posteriors, an i-vector is extracted as the utterance-level feature.
8. The speech feature extraction apparatus according to claim 1 ,
wherein the a monotonically increasing function is used to calculate the function values from the posteriors, and the utterance-level feature is extracted using the neural network.
9. A speech feature extraction method comprising:
dropping non-voice frames from frames corresponding to an input speech utterance, and calculating posteriors of voice for the frames;
calculating function values as weights from the posteriors;
extracting an utterance-level feature from the frames based on multiple frame-level features by using the function values as weights for pooling the frames in a pooling layer of a neural network.
10. A non-transitory computer-readable storage medium storing a program that includes commands for causing a computer to execute:
dropping non-voice frames from frames corresponding to an input speech utterance, and calculating posteriors of voice for the frames;
calculating function values as weights from the posteriors;
extracting an utterance-level feature from the frames based on multiple frame-level features by using the function values as weights for pooling the frames in a pooling layer of a neural network.