High-efficiency digital signal processing of streaming media
A phase vocoder executes a fast-Fourier transform (FFT) with respect to an input audio data stream to generate an array of frequency-domain values corresponding to respective frequencies that are nominally uniformly distributed across a frequency range of interest, each of the frequency-domain values being representative of amplitude and phase of a spectral component of the input audio data stream at the respective frequency. The phase vocoder scales the nominally uniform distribution of the respective frequencies to reduce a cumulative error across the frequency distribution resulting from finite precision of a digital representation and then implements at least one of a time-stretching operation or a pitch-shifting operation with respect to the input data stream by manipulating the frequency-domain values with respect to one another within the array.
1. A method of operation within a phase vocoder, the method comprising:
executing a fast-Fourier transform (FFT) with respect to an input audio data stream to generate an array of frequency-domain values corresponding to respective frequencies that are nominally uniformly distributed across a frequency range of interest, each of the frequency domain values being representative of amplitude and phase of a spectral component of the input audio data stream at the respective frequency;
scaling the nominally uniform distribution of the respective frequencies to reduce a cumulative error across the frequency distribution resulting from finite precision of a digital representation;
implementing at least one of a time-stretching operation or a pitch-shifting operation with respect to the input data stream by manipulating the frequency-domain values with respect to one another within the array; and
executing an inverse-FFT with respect to the manipulated frequency-domain values to generate a time-domain audio output.
2. The method of claim 1 wherein scaling the nominally uniform distribution of the respective frequencies to reduce a cumulative error comprises compressing the frequency distribution by omitting at least one of a division or a multiplication by Pi in the generation of the frequency distribution.
3. The method of claim 1 wherein at least one of executing the FFT, scaling the nominally uniform distribution of the respective frequencies, implementing the time-stretching operation, implementing the pitch-shifting operation, or executing the inverse-FFT comprises executing a sequence of program code instructions within one or more processors.
4. The method of claim 3 wherein executing the sequence of program code instructions comprises executing floating point operations in which Pi is implemented as a floating point number with finite precision such that, absent scaling the nominally uniform distribution, a substantial cumulative error results across the frequency distribution due to the finite precision of the floating point implementation of Pi.
5. The method of claim 1 wherein scaling the nominally uniform distribution of the respective frequencies comprises scaling by a factor of Pi such that at least one multiplication or division by Pi is obviated.
6. A phase vocoder comprising:
a Fourier transform engine to (i) generate, through execution of a fast-Fourier transform (FFT) with respect to an input audio data stream, an array of frequency-domain values corresponding to respective frequencies that are nominally uniformly distributed across a frequency range of interest, each of the frequency domain values being representative of amplitude and phase of a spectral component of the input audio data stream at the respective frequency, and (ii) to scale the nominally uniform distribution of the respective frequencies such that a cumulative error across the frequency distribution resulting from finite precision of a digital representation is reduced;
a scaling module to execute at least one of a time-stretching operation or a pitch-shifting operation with respect to the input data stream by manipulating the frequency-domain values with respect to one another within the array; and
an inverse Fourier transform engine to generate, through execution of an inverse-FFT with respect to the manipulated frequency-domain values, a time-domain audio output.
7. The phase vocoder of claim 6 wherein the Fourier transform engine to scale the nominally uniform distribution of the respective frequencies to reduce a cumulative error comprises logic to compress the frequency distribution by omitting at least one of a division or a multiplication by Pi in the generation of the frequency distribution.
8. The phase vocoder of claim 6 wherein at least one of the Fourier transform engine, scaling module and inverse Fourier transform engine comprises one or more programmed processors.
9. The phase vocoder of claim 8 wherein the one or more programmed processors comprises a processor and a memory having a sequence of program code instructions stored therein and wherein the processor executes floating point operations in which Pi is implemented as a floating point number with finite precision such that, absent scaling the nominally uniform distribution, a substantial cumulative error results across the frequency distribution due to the finite precision of the floating point implementation of Pi.
10. The phase vocoder of claim 6 wherein the Fourier transform engine to scale the nominally uniform distribution of the respective frequencies comprises logic to scale by a factor of Pi such that at least one multiplication or division by Pi is obviated.