Average Magnitude Difference Function Research Articles

A pitch extraction algorithm based on LPC inverse filtering and AMDF

We present an efficient algorithm for determining fundamental frequency and voiced/unvoiced (V/UV) decision of speech. The pitch extractor utilizes the cross-correlation average magnitude difference function (AMDF) waveform that is obtained from the linear prediction residual signal. The decision logic used in pitch extraction is simple and reliable. The periodicity and null depth of AMDF waveforms, together with the average residual energy and the past pitch information, are used in the decision logic for fundamental frequency and V/UV decision. Computer simulation of the algorithm yielded accurate results, even for difficult phonemes for pitch extraction.

Pitch-Adaptive DPCM Coding of Speech With Two-Bit Quantization and Fixed Spectrum Prediction

This paper is concerned with the utilization of speech waveform periodicities in differential pulse code modulation (dpcm) coding with 2-bit adaptive quantization and time-invariant spectrum prediction. Our work is based on computer simulations of dpcm codes. We have studied pitch detectors based on autocorrelation and an average magnitude difference function (amdf), and we have measured the benefits of predicting from a previous pitch period as functions of pitch-period-updating frequency and periodicity-indicating thresholds (for autocorrelation and the amdf). We have compared several alternative methods of utilizing past quantized samples (in the present and previous pitch periods) for providing speech sample predictions. We find the following combination to be attractive for waveform coding at bit rates in the neighborhood of 16 kb/s: 2-bit adaptive quantization with a one-word (2-bit dpcm word) memory, pitch detection performed on unquantized speech (preferably with an amdf criterion) and a prediction scheme that uses fixed three-tap (short-term) prediction for nonperiodic waveform segments, but switches to an appropriate one-tap (long-term) predictor upon the detection of strong periodicity. With four sample utterances, the latter procedure results in an average SNR (signal-to-noise ratio) gain of 3.75 dB over a non-pitch-adaptive encoder.

A comparative performance study of several pitch detection algorithms

A comparative performance study of seven pitch detection algorithms was conducted. A speech data base, consisting of eight utterances spoken by three males, three females, and one child was constructed. Telephone, close talking microphone, and wideband recordings were made of each of the utterances. For each of the utterances in the data base; a pitch contour was semiautomatically measured using a highly sophisticated interactive pitch detection program. The pitch contour was then compared with the pitch contour that was obtained from each of the seven programmed pitch detectors. The algorithms used in this study were 1) a center clipping, infinite-peak clipping, modified autocorrelation method (AUTOC), 2) the cepstral method (CEP), 3) the simplified inverse filtering technique (SIFT) method, 4) the parallel processing time-domain method (PPROC), 5) the data reduction method (DARD), 6) a spectral flattening linear predictive coding (LPC) method, and 7) the average magnitude difference function (AMDF) method. A set of measurements was made on the pitch contours to quantify the various types of errors which occur in each of the above methods. Included among the error measurements were the average and standard deviation of the error in pitch period during voiced regions, the number of gross errors in the pitch period, and the average number of voiced-unvoiced classification errors. For each of the error measurements, the individual pitch detectors could be rank ordered as a measure of their relative performance as a function of recording condition, and pitch range of the various speakers. Performance scores are presented for each of the seven pitch detectors based on each of the categories of error.

A Real-Time Adaptive Predictive Coder Using Small Computers

This paper describes a real-time computer implementaation of a fourth-order adaptive predictive coder (APC) which transmits speech at 6400 bits/s. To reduce the demands on the central processor, the terminal uses the average magnitude difference function (AMDF) for pitch extraction and calculates the predictor coefficients and reflection coefficients assuming stationarity of the input data. With 16-bit fixed-point arithmetic processing, the coder achieves an intelligibility score of 87 percent on the diagnostic rhyme tests (DRT) and produces speech with acceptable voice naturalness even in the presence of acoustic background noise.

Average magnitude difference function pitch extractor

This paper describes a method for using the average magnitude difference function (AMDF) and associated decision logic to estimate the pitch period of voiced speech sounds. The AMDF is a variation on autocorrelation analysis where, instead of correlating the input speech at various delays (where multiplications and summations are formed at each value of delay), a difference signal is formed between the delayed speech and the original and, at each delay, the absolute magnitude of the difference is taken. The difference signal is always zero at delay = π, and exhibits deep nulls at delays corresponding to the pitch period of voiced sounds. Some of the reasons the AMDF is attractive include the following. 1) It is a simple measurement which gives a good estimate of pitch contour, 2) it has no multiply operations, 3) its dynamic range characteristics are suitable for implementation on a 16-bit machine, and 4) the nature of its operations makes it suitable for implementation on a programmable processor or in special purpose hardware. The implementation of the AMDF pitch extractor (nonreal-time simulation and real-time) is described and experimental results presented to illustrate its basic measurement properties.

Real-Time All-Digital FFT Vocoder

In this paper, we describe a real-time all-digital spectrum channel vocoder implemented on a programmable signal processor. The vocoder system efficiently utilizes a radix four, 64-point FFT algorithm for spectral analysis, while the pitch extractor selects the pitch period using decision logic on the average magnitude difference function (AMDF) of the input speech. The receiver linearly interpolates both the pitch period and the synthesized impulse response which is generated by using the inverse Fourier transform. The vocoder frame updating rate is 22 msec, and the input speech sampling rate is 7040 samples/sec. The configuration of the vocoder and the logic of the AMDF pitch extractor is discussed and the system demonstrated with audio tapes.

Real-Time Autocorrelation Predictive Coder

We describe a real-time autocorrelation predictive coding system for speech compression that has been implemented on a programmable signal processor having a word size of 16 bits and a multiply time of 750 nsec. This system employs the normal equations to generate an all-pole filter which, when impulsed, produces the least-squares approximation to the input signal. The technique uses a 6.4-kHz sampling rate on an input bandwidth of 3 kHz, a lOth-order predictive filter, an average magnitude difference function (AMDF) pitch extractor, and pitch synchronous interpolation of the gain and pitch parameters to produce transmission rates of 2.4 and 3.6 k bits/sec. The paper emphasizes the modifications needed to achieve real-time operation and the accuracy and timing requirements to extract pitch, calculate the autocorrelation and filter coefficients, and synthesize the output.

AMDF Pitch Extractor

A method is described for estimating the fundamental frequency component of voiced speech using the Average Magnitude Difference Function (AMDF). The technique, a variation of autocorrelation analysis, uses a difference signal formed by delaying the input speech various amounts and subtracting the delayed waveform from the original. The difference signal is always zero at delay = ø, and exhibits deep nulls at delays corresponding to the pitch period for a voiced sound having a quasi-periodic structure. Unlike the generation of the autocorrelation function, the AMDF calculations require no multiplications, a desirable property for real-time speech processing. Also, since the computations involve numerous repetitive operations, they are very suitable for fabrication in special-purpose hardware. The implementation of the AMDF pitch extractor will be described and experimental results presented to illustrate the basic measurement properties.