Method and system for generating training data for an automatic speech recognizer
View Patent ↗The invention describes a method and a system for generating training data (D T ) for an automatic speech recogniser ( 2 ) for operating at a particular first sampling frequency (f H ), comprising steps of deriving spectral characteristics (S L ) from audio data (D L ) sampled at a second frequency (f L ) lower than the first sampling frequency (f H ), extending the bandwidth of the spectral characteristics (S L ) by retrieving bandwidth extending information OBE) from a codebook ( 6 ), and processing the bandwidth extended spectral characteristics (S LE ) to give the required training data (D T ). Moreover a method and a system ( 5 ) for generating a codebook ( 6 ) for extending the bandwidth of spectral characteristics (S L ) for audio data (D L ) sampled at a second sampling frequency (f L ) to spectral characteristics (S H ) for a first sampling frequency (f H ) higher than the second sampling frequency (f L ) are described.
1. A method for generating training data for an automatic speech recognizer operating at a first sampling frequency, comprising:
deriving spectral characteristics from audio data sampled at a second frequency lower than the first sampling frequency;
extending a bandwidth of the spectral characteristics by retrieving bandwidth extending information from a codebook so that the audio data sampled at the second frequency is compatible with the automatic speech recognizer operating at the first sampling frequency; and
processing the bandwidth extended spectral characteristics to give the required training data,
wherein extending the bandwidth includes augmenting the spectral characteristics of the audio data sampled at the second frequency with a set of spectral characteristics of audio data sampled at the first sampling frequency obtained from the codebook, wherein each entry in the codebook is generated by:
deriving a first set of spectral characteristics from audio data sampled at the first sampling frequency;
performing a sampling rate transformation on the audio data to the second frequency and deriving a corresponding second set of spectral characteristics; and
generating a codebook entry comprising information from the second set of spectral characteristics, together with additional bandwidth extending information extracted from the first set of spectral characteristics.
2. A method according to claim 1 , where the conversion of audio data into sets of spectral characteristics comprises calculating the FFT of the audio data to give a set of Fourier coefficients and filtering the output of the FFT with a filterbank to give a set of filterbank power values.
3. A method according to claim 2 , where the conversion of audio data into sets of spectral characteristics comprises processing the FFT coefficients or the filterbank power values to give a set of log-spectral coefficients.
4. A method according to claim 1 , where the processing of bandwidth extended spectral characteristics comprises a step of altering the spectrum to adjust signal properties of the audio data.
5. A method according to claim 4 , where the step of altering the spectrum to adjust the signal properties of the audio data is performed in the linear domain.
6. A method according to claim 1 , where the derivation of the spectral characteristics from audio data is followed by a step subtracting the mean spectrum from the spectral characteristics.
7. A method for training an automatic speech recognition system wherein the data used for training are at least partially generated using a method according to claim 1 .
8. A system for generating training data for an automatic speech processor operating at a first sampling frequency, comprising:
a hardware converter for deriving spectral characteristics from audio data sampled at a second frequency lower than the first sampling frequency;
a hardware retrieval unit for retrieving bandwidth extending information for the spectral characteristics from a codebook so that the audio data sampled at the second frequency is compatible with the automatic speech processor operating at the first sampling frequency; and
a hardware processing module for processing the bandwidth-extended spectral characteristics to give the required training data,
wherein the hardware retrieval unit augments the spectral characteristics of the audio data sampled at the second frequency with the set of spectral characteristics of audio data sampled at the first sampling frequency obtained from the codebook, wherein each entry in the codebook is generated by:
deriving a first set of spectral characteristics from audio data sampled at the first sampling frequency;
performing a sampling rate transformation on the audio data to the second frequency and deriving a corresponding second set of spectral characteristics; and
generating a codebook entry comprising information from the second set of spectral characteristics, together with additional bandwidth extending information extracted from the first set of spectral characteristics.
9. A method for generating training data for an automatic speech recognizer operating at a first sampling frequency, comprising:
deriving spectral characteristics from audio data sampled at a second frequency lower than the first sampling frequency;
extending a bandwidth of the spectral characteristics by retrieving bandwidth extending information from a codebook so that the audio data sampled at the second frequency is compatible with the automatic speech recognizer operating at the first sampling frequency; and
processing the bandwidth extended spectral characteristics to give the required training data,
wherein extending the bandwidth includes augmenting the spectral characteristics of the audio data sampled at the second frequency with a set of spectral characteristics of audio data sampled at the first sampling frequency obtained from the codebook, wherein each entry in the codebook is generated by:
deriving a first set of spectral characteristics from audio data sampled at the first sampling frequency;
performing a sampling rate transformation on the audio data to the second frequency and deriving a corresponding second set of spectral characteristics; and
generating a codebook entry comprising information from the first and second sets of spectral characteristics, stored such that the first and second sets of spectral characteristics are associated with each other and yet remain separate.