Methods and apparatus for detecting harmonic structure in a waveform

• US 5,809,453 A
• Filed: 01/25/1996
• Issued: 09/15/1998
• Est. Priority Date: 01/25/1995
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: Dragon Systems, Inc. (Microsoft Corporation)
• Inventors: Hunt, Melvyn John
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/591,766
• Time in Patent Office: 964 Days
• Field of Search: 395/2.17, 395/2.15, 395/2.12, 395/2.42, 704/208, 704/206, 704/207, 704/203, 704/233, 704/211, 704/214, 704/216, 704/217, 704/218
• US Class Current: 704/214
• CPC Class Codes: G10L 2025/937 Signal energy in various frequency bands
• CPC Class Codes: G10L 25/06 the extracted parameters being correlation coefficients
• CPC Class Codes: G10L 25/24 the extracted parameters being the cepstrum
• CPC Class Codes: G10L 25/27 characterised by the analysis technique
• CPC Class Codes: G10L 25/93 Discriminating between voiced and unvoiced parts of speech signals G10L25/90 takes precedence
• CPC Class Codes: H04Q 1/46 comprising means for distinguishing between a signalling current of predetermined frequency and a complex current containing that frequency, e.g. speech current

Abstract

The periodicity of a signal from a voice channel is determined by sampling the signal, computing the log power spectrum, optionally thresholding and differencing the power spectrum, and then performing an autocorrelation function of limited order to confine the search for periodicity to spans of up to about 400 Hz to 500 Hz.

Claims

1. A method of using a computer processor to analyze electrical signals and input data representative of sounds to determine the periodicity of a waveform represented by said data, which method comprises using the computer processor to process said input data to obtain log power spectrum data representing a log power spectrum of said input data, using the computer processor to apply an autocorrelation function of limited order to said log power spectrum data to obtain autocorrelation data, and thereby determining an estimate of the periodicity of said waveform.

2. A method according to claim 1, wherein said step of processing said input data includes computing the complex Fourier Transform of the input data and thereafter computing the log power spectrum.

3. A method according to claim 1, wherein said log power spectrum data is thresholded before application of said autocorrelation function.

4. A method according to claim 1, wherein said log power spectrum data is differenced before application of said autocorrelation function.

5. A method according to claim 1, wherein the said limited order of the autocorrelation function corresponds to searches for periodicity to spans of up to above 400-500 Hz.

6. A method according to claim 1, wherein said step of determining an estimate of the periodicity includes computing the cosine transform of said autocorrelation data and thereafter determining at least one of a height and frequency of a greatest peak in said cosine transform.

7. A method according to claim 6, wherein the height of the greatest peak is used as an indicator of the speech-related harmonicity.

8. A method according to claim 6, wherein the autocorrelation data is extended with zeroes before computing said cosine transform.

9. A method according to claim 1, wherein, having determined an estimate of the periodicity of said waveform, said estimate is refined by applying a cepstrum processing technique to said input data to obtain cepstrum data and monitoring said cepstrum data in the neighbourhood corresponding to said estimate to obtain a refined estimate.

10. Apparatus for digitally processing input data representative of sounds to estimate periodicity of a waveform of the input data, said apparatus comprising:

input means for inputting the input data representative of sounds in the form of electrical signals,

processing means comprising a computer processor for processing the input data from said input means to obtain log power spectrum data representative of a log power spectrum of said input data,

autocorrelation means comprising the computer processor for autocorrelating said log spectrum data; and

means responsive to said autocorrelation data comprising the computer processor to determine an estimate of the periodicity of said waveform.