Method and apparatus for wavelet based data compression having adaptive bit rate control for compression of digital audio or other sensory data

• US 5,819,215 A
• Filed: 10/13/1995
• Issued: 10/06/1998
• Est. Priority Date: 10/13/1995
• Status: Expired due to Term

• Current Assignee: Hewlett Packard Enterprise Development LP (Hewlett-Packard Enterprise Company)
• Original Assignee: US Robotics Mobile Communication Corporation
• Inventors: Yang, Jack, Dobson, Kurt, Whitney, Nathan, Rigstad, Peter, Smart, Kevin
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Dorvil, Richemond
• Application Number:
  US08/543,205
• Time in Patent Office: 1,089 Days
• Field of Search: 395/2.38, 395/2.39, 395/2.12, 395/2.14, 345/396, 345/397, 382/232, 382/240, 382/244, 382/248, 704/229, 704/230, 704/203, 704/205, 704/500, 704/501, 704/504
• US Class Current: 704/230
• CPC Class Codes: G06T 9/007 Transform coding, e.g. discrete cosine transform
• CPC Class Codes: G10L 25/27 characterised by the analysis technique
• CPC Class Codes: G11B 20/00007 Time or data compression or expansion audio compression based on psychoacoustics G10L19/00; data processing for reproducing audio data at different playback speeds G10L21/04; video compression H04N19/00; data compression per se H03M7/30
• CPC Class Codes: H04L 27/0004 using wavelets

Abstract

The compression method utilizes a combination of both lossy and lossless compression to achieve significant compression while retaining very high subjective quality of the reconstructed or decompressed signal. Methods and apparatus for compression and decompression of digital audio data are provided. In one preferred embodiment, the compression method or apparatus has a bit rate control feedback loop particularly well suited to matching the output bit rate of the data compressor to the bandwidth capacity of a communication channel. Embodiments trade-off various error sources in order to keep perceptible distortion levels to a minimum for a fixed bit rate. Preferred embodiments also utilize a unique combination of run length and Huffman encoding methods in order to take advantage of both local and global statistics.

Claims

1. A method for compressing digitally sampled audio data into an average bit rate of compressed audio data, comprising the steps of:

a) performing a discrete wavelet transform on the digitally sampled audio data to obtain resultant wavelet coefficients;

b) eliminating according to a predetermined criteria a calculated fractional percentage of wavelet coefficients to achieve the average bit rate;

c) quantizing the wavelet coefficients using a selected quantization level;

d) entropy encoding the quantized coefficients to obtain the compressed audio data; and

e) feeding the number of bits used to represent the entropy encoded coefficients back into the calculation of the fractional percentage of wavelet coefficients which must be eliminated to achieve the average bit rate.

2. A method for compressing digitally sampled audio data as recited in claim 1 wherein the calculating a fractional percentage of wavelet coefficients which much be eliminated comprises changing the selected quantization level.

3. A method for compressing digitally sampled audio data as recited in claim 1 wherein the entropy encoding comprises run length encoding.

4. A method for compressing digitally sampled audio data as recited in claim 1 wherein the entropy encoding comprises Huffman encoding.

5. A method for compressing digitally sampled audio data as recited in claim 1 wherein the entropy encoding comprises a combination of entropy encoding methods.

6. A method for compressing digitally sampled audio data as recited in claim 5 wherein the combination of entropy encoding methods comprises a combination of run length encoding and Huffman encoding.

7. A method for compressing digitally sampled audio data as recited in claim 1 wherein the predefined criteria for eliminating the calculated fractional percentage of wavelet coefficients is to eliminate wavelet coefficients with a smaller magnitude first.

8. A method for compressing digitally sampled audio data as recited in claim 1 wherein the quantization level used to quantize the wavelet coefficients is selected based on a psychoacoustic model.

9. A method for compressing digitally sampled audio data as recited in claim 8 wherein the psychoacoustic model utilizes one or more parameters derived from the digitally sampled audio data.

10. A method for compressing digitally sampled audio data as recited in claim 9 wherein the one or more parameters derived from the digitally sampled audio data comprises the energy of a portion of the digitally sampled audio data.

11. A method for compressing digitally sampled audio data as recited in claim 9 wherein the one or more parameters derived from the digitally sampled audio data comprises a measure of the entropy of a portion of the digitally sampled audio data.

12. A method for compressing digitally sampled audio data as recited in claim 1 wherein the selected quantization level used to quantize the wavelet coefficients is selected based on the calculated percentage of wavelet coefficients to be eliminated.

13. A method for compressing digitally sampled audio data as recited in claim 1 comprising the further step of scaling the wavelet coefficients so they lie between -1.0 and +1.0 prior to the step of calculating a fractional percentage of wavelet coefficients which must be eliminated to achieve the average bit rate.

14. A method for compressing digitally sampled audio data which has been divided into data frames containing a predefined number of digital audio data samples into an average bit rate of compressed audio data, the method comprising the steps of:

a) performing a discrete wavelet transform on the data frame to obtain the corresponding wavelet coefficients;

b) calculating a fractional percentage of wavelet coefficients which must be eliminated in this frame in order to achieve the average bit rate;

c) eliminating the calculated fractional percentage of wavelet coefficients according to a predefined criteria;

d) selecting a quantization level for the wavelet coefficients based on a psychoacoustic model which uses one or more parameters derived from the data in the data frame;

e) quantizing the wavelet coefficients at the selected quantization level;

f) entropy encoding the quantized wavelet coefficients to obtain the compressed audio data; and

g) feeding the number of bits used to represent the entropy encoded coefficients back into the calculation of the fractional percentage of wavelet coefficients which must be eliminated to achieve the average bit rate.

15. A method for compressing digitally sampled audio data as recited in claim 14 wherein the predefined criteria to eliminate the calculated fractional percentage of wavelet coefficients comprises eliminating wavelet coefficients having a smaller magnitude first.

16. A method for compressing digitally sampled audio data as recited in claim 15 wherein calculating a fractional percentage of wavelet coefficients which must be eliminated comprises changing the selected quantization level.

17. A method for compressing digitally sampled audio data as recited in claim 16 further comprising the step of calculating the energy in the data frame before the step of performing a discrete wavelet transform.

18. A method for compressing digitally sampled audio data as recited in claim 17 wherein the one or more parameters derived from the data in the data frame comprises the calculated energy of the data in the data frame.

19. A method for compressing digitally sampled audio data as recited in claim 18 further comprising the step of dividing each wavelet coefficient by a fixed gain prior to the step of quantizing the wavelet coefficients.

20. A method for compressing digitally sampled audio data as recited in claim 19 wherein the gain is the a filtered estimate of the maximum wavelet coefficient in the data frame.

21. A method for compressing digitally sampled audio data as recited in claim 20 wherein the entropy encoding comprises run length encoding which produces at least a run length encoding symbol and a run length encoding count for each run of contiguous eliminated coefficients.

22. A method for compressing digitally sampled audio data as recited in claim 21 wherein the entropy encoding further comprises Huffman encoding.

23. A method for compressing digitally sampled audio data as recited in claim 22 wherein the entropy encoding comprises an adaptive codebook entropy encoding method.

24. A method for compressing digitally sampled audio data as recited in claim 22 wherein the selected quantization level is entropy encoded using a first codebook.

25. A method for compressing digitally sampled audio data as recited in claim 24 wherein the gain is entropy encoded using a second codebook.

26. A method for compressing digitally sampled audio data as recited in claim 25 wherein the run length encoding count is encoded using a third codebook.

27. A method for compressing digitally sampled audio data as recited in claim 26 wherein the quantized wavelet coefficients are entropy encoded using a fourth codebook.

28. A method for compressing digitally sampled audio data as recited in claim 27 wherein the fourth codebook is selected, based on the quantization level, from a set of wavelet coefficient codebooks.

29. A method for compressing digitally sampled audio data as recited in claim 28 wherein the one or more parameters derived from the data in the data frame comprises a measure of the entropy of the data in the data frame.

30. A method for compressing digitally sampled audio data as recited in claim 29 further comprising the step of scaling the wavelet coefficients so they lie between -1.0 and +1.0 prior to the step of eliminating the calculated fractional percentage of wavelet coefficients needed to achieve the bit rate.

31. A method for compressing digitally sampled audio data which has been divided into data frames containing a predefined number of digital audio data samples into an average bit rate of compressed audio data, the method comprising the steps of:

a) calculating the energy of the data frame;

b) performing a discrete wavelet transform on the data frame to obtain the corresponding wavelet coefficients;

c) scaling the wavelet coefficients so they lie between -1.0 and +1.0;

d) using a number of bits output last data frame to calculate a fractional percentage of wavelet coefficients which must be eliminated in this frame in order to achieve the average bit rate;

e) eliminating all wavelet coefficients that fall below a predefined threshold;

f) eliminating wavelet coefficients starting with the smallest remaining values until the calculated fractional percentage of wavelet coefficients needed to achieve the average bit rate have been eliminated;

g) dividing all wavelet coefficients by a gain, the gain comprising a filtered estimate of the maximum wavelet coefficient in the frame;

i) selecting a quantization level for the wavelet coefficients based on a psychoacoustic model which uses the calculated energy of the data frame;

j) quantizing the wavelet coefficients at the selected quantization level; and

k) entropy encoding the quantized wavelet coefficients using run length encoding of contiguous eliminated wavelet coefficients followed by Huffman encoding to obtain the compressed audio data.

32. A method for compressing digitally sampled audio data as recited in claim 31 further comprising changing the fractional percentage of wavelet coefficients which must be eliminated by changing the selected quantization level if the fractional percentage of wavelet coefficients which must be eliminated falls outside a predetermined range.

33. A method for compressing digitally sampled audio data as recited in claim 31 wherein the selected quantization level is included in the data frame.

34. A method for compressing digitally sampled audio data as recited in claim 31 wherein the selected quantization level is entropy encoded using a first codebook and included in the data frame.

35. A method for compressing digitally sampled audio data as recited in claim 34 wherein the gain is entropy encoded using a second codebook and included in the data frame.

36. A method for compressing digitally sampled audio data as recited in claim 35 wherein the run length encoding count is encoded using a third codebook.

37. A method for compressing digitally sampled audio data as recited in claim 36 wherein the quantized wavelet coefficients are entropy encoded using a fourth codebook.

38. A method for compressing digitally sampled audio data as recited in claim 37 wherein the fourth codebook is selected, based on the quantization level, from a set of wavelet coefficient codebooks.

39. A method of reconstructing digitally sampled audio information which has been previously compressed into compressed digitally sampled audio information, the method comprising the steps of:

a) entropy decoding the compressed digitally sampled audio information to obtain entropy decoded information comprising wavelet coefficients wherein the wavelet coefficients are resultant from decimation of a calculated fractional percentage of wavelet coefficients necessary to achieve an average bit rate of the compressed digitally sampled audio information;

b) performing an inverse discrete wavelet transform on the entropy decoded information to obtain inverse wavelet transformed data, a portion of which are comprised of interpolated data for mitigating the effects of decimation of wavelet coefficients; and

c) resealing the inverse wavelet transformed data into the time domain to obtain the digitally sampled audio information.

40. A method of reconstructing digitally sampled audio information which has been previously compressed as recited in claim 39 comprising the further step of multiplying each wavelet coefficient by a gain factor prior to performing the inverse discrete wavelet transform.

41. A method of reconstructing digitally sampled audio information which has been previously compressed as recited in claim 40 wherein the previously compressed data has been formed into one or more data frames.

42. A method of reconstructing digitally sampled audio information which has been previously compressed as recited in claim 41 wherein the gain factor is found in the one or more data frames.

43. A method of reconstructing digitally sampled audio information which has been previously compressed as recited in claim 39 wherein rescaling the transformed data into the time domain comprises converting the transformed values into integer values.

44. An article of manufacture comprising:

a computer usable medium having computer readable program code means embedded therein for compressing digitally sampled audio data, the computer readable program code means comprising;

a) means for performing a discrete wavelet transform on the digitally sampled audio data to obtain the resultant wavelet coefficients;

b) means for eliminating according to a predetermined criteria a calculated fractional percentage of wavelet coefficients to achieve the average bit rate;

c) means for quantizing the wavelet coefficients using a selected quantization level;

d) means for entropy encoding the quantized coefficients to obtain the compressed audio data; and

e) means for feeding the number of bits used to represent the entropy encoded coefficients back into the means for calculating the fractional percentage of wavelet coefficients which must be eliminated to achieve the average bit rate.

45. An article of manufacture as recited in claim 44 wherein the means for calculating a fractional percentage of wavelet coefficients which much be eliminated comprises means for changing the selected quantization level.

46. An article of manufacture as recited in claim 44 wherein the means for entropy encoding the quantized coefficients comprises means for run length encoding.

47. An article of manufacture as recited in claim 44 wherein the means for entropy encoding comprises means for Huffman encoding.

48. An article of manufacture as recited in claim 44 wherein the means for entropy encoding comprises means for entropy encoding the quantized wavelet coefficients using a combination of entropy encoding methods.

49. An article of manufacture as recited in claim 48 wherein the means for entropy encoding the quantized wavelet coefficients comprises means for run length encoding the quantized wavelet coefficients and means for Huffman encoding the quantized wavelet coefficients.

50. An article of manufacture as recited in claim 49 further comprising means to divide the wavelet coefficients by a gain.

51. An article of manufacture as recited in claim 50 further comprising means for entropy encoding the gain using a first codebook.

52. An article of manufacture as recited in claim 51 further comprising means for entropy encoding the selected quantization level using a second codebook.

53. An article of manufacture as recited in claim 52 further comprising means for entropy encoding the quantized wavelet coefficients using a third codebook.

54. An article of manufacture as recited in claim 53 wherein the means for run length encoding the quantized wavelet coefficients produces at least a run length marker and a run length count for each run length encoded sequence of coefficients.

55. An article of manufacture as recited in claim 54 further comprising means for entropy encoding the run length count using a fourth codebook.

56. An article of manufacture as recited in claim 44 wherein the predefined criteria for eliminating the calculated fractional percentage of wavelet coefficients is to eliminate wavelet coefficients with a smaller magnitude first.

57. An article of manufacture as recited in claim 44 wherein the quantization level used to quantize the wavelet coefficients is selected based on a psychoacoustic model.

58. An article of manufacture as recited in claim 57 wherein the psychoacoustic model utilizes one or more parameters derived from the digitally sampled audio data.

59. An article of manufacture as recited in claim 58 wherein the one or more parameters derived from the digitally sampled audio data comprises the energy of a portion of the digitally sampled audio data.

60. An article of manufacture as recited in claim 59 wherein the one or more parameters derived from the digitally sampled audio data comprises a measure of the entropy of a portion of the digitally sampled audio data.

61. An article of manufacture as recited in claim 44 wherein the selected quantization level used to quantize the wavelet coefficients is selected based on the calculated percentage of wavelet coefficients to be eliminated.

62. An article of manufacture as recited in claim 44 further comprising means to divide the wavelet coefficients by a gain.

63. An article of manufacture as recited in claim 62 wherein the gain represents a filtered estimate of the maximum wavelet coefficient value.

64. An article of manufacture as recited in claim 62 wherein the gain provides small signal suppression when the maximum wavelet coefficient value drops rapidly.

65. An article of manufacture comprising:

a computer usable medium having computer readable program code means embedded therein for compressing digitally sampled data into an average bit rate of compressed sampled data, the computer readable program code means comprising;

a) means for dividing the digitally sampled data into data frames containing a predefined number of digital data samples;

b) means for performing a discrete wavelet transform on the data frame to obtain the corresponding wavelet coefficients;

c) means for calculating a fractional percentage of wavelet coefficients which must be eliminated in a current data frame in order to achieve a average bit rate;

d) means for eliminating the calculated fractional percentage of wavelet coefficients according to a predefined criteria;

e) means for selecting a quantization level for the wavelet coefficients based on a psychoacoustic model which uses one or more parameters derived from the data in the current data frame;

f) means for quantizing the wavelet coefficients at the selected quantization level;

g) means for entropy encoding the quantized wavelet coefficients to obtain the compressed sampled data; and

h) means for feeding the number of bits used to represent the entropy encoded coefficients back into the means for calculating the fractional percentage of wavelet coefficients which must be eliminated to achieve the average bit rate.

66. An article of manufacture as recited in claim 65 wherein the predefined criteria to eliminate the calculated fractional percentage of wavelet coefficients comprises eliminating wavelet coefficients having a smaller magnitude first.

67. An article of manufacture as recited in claim 66 wherein the means to calculate a fractional percentage of wavelet coefficients which must be eliminated comprises means for changing the selected quantization level.

68. An article of manufacture as recited in claim 67 further comprises means for scaling the wavelet coefficients so they lie between -1.0 and +1.0.

69. An article of manufacture as recited in claim 68 further comprising means for calculating the energy in the data frame.

70. An article of manufacture as recited in claim 69 wherein the one or more parameters derived from the data in the data frame comprises the calculated energy of the data in the data frame.

71. An article of manufacture as recited in claim 70 means for dividing each wavelet coefficient by a fixed gain prior to quantizing the wavelet coefficients.

72. An article of manufacture as recited in claim 71 wherein the gain is a filtered estimate of the maximum wavelet coefficient in the data frame.

73. An article of manufacture as recited in claim 72 wherein the means for entropy encoding comprises an adaptive codebook entropy encoding means.

74. An article of manufacture as recited in claim 72 wherein the means for entropy encoding comprises means for run length encoding which produces at least a run length encoding symbol and a run length count for each run of contiguous eliminated coefficients.

75. An article of manufacture as recited in claim 74 wherein the means for entropy encoding further comprises means for Huffman encoding.

76. An article of manufacture as recited in claim 75 comprising means for entropy encoding the selected quantization level using a first codebook.

77. An article of manufacture as recited in claim 76 comprising means for entropy encoding the gain using a second codebook.

78. An article of manufacture as recited in claim 77 comprising means for entropy encoding the run length coding count using a third codebook.

79. An article of manufacture as recited in claim 78 wherein the means for entropy encoding the quantized wavelet coefficients comprises means for entropy encoding the quantized wavelet coefficients using a fourth codebook.

80. An article of manufacture as recited in claim 79 comprising means for selecting the fourth codebook, based on the quantization level, from a set of wavelet coefficient codebooks.

81. An article of manufacture as recited in claim 80 wherein the one or more parameters derived from the data in the data frame comprises a measure of the entropy of the data in the data frame.

82. An article of manufacture comprising:

a computer readable medium having computer readable program code means embedded therein for compressing digitally sampled audio data into an average bit rate of compressed audio data, the computer readable program code means comprising;

a) means for dividing the digitally sampled audio data into data frames containing a predefined number of digital audio data samples;

b) means for calculating the energy of the data frame;

c) means for performing a discrete wavelet transform on the data frame to obtain the corresponding wavelet coefficients;

d) means for scaling the wavelet coefficients so they lie between -1.0 and +1.0;

e) means for using a number of bits output last data frame to calculate a fractional percentage of wavelet coefficients which must be eliminated in this frame in order to achieve the average bit rate;

f) means for eliminating all wavelet coefficients that fall below a predefined threshold;

g) means for eliminating wavelet coefficients starting with the smallest remaining values until the calculated fractional percentage of wavelet coefficients needed to achieve the average bit rate have been eliminated;

h) means for dividing all wavelet coefficients by a fixed gain, the fixed gain comprising a filtered estimate of the maximum wavelet coefficient in the frame;

i) means for selecting a quantization level for the wavelet coefficients based on a psychoacoustic model which uses the calculated energy of the data frame;

j) means quantizing the wavelet coefficients at the selected quantization level; and

k) means for entropy encoding the quantized wavelet coefficients using run length encoding of contiguous eliminated wavelet coefficients followed by Huffmnan encoding to obtain the compressed audio data.

83. An article of manufacture as recited in claim 82 further comprising means for changing the fractional percentage of wavelet coefficients which must be eliminated by changing the selected quantization level if the fractional percentage of wavelet coefficients which must be eliminated falls outside a predetermined range.

84. An article of manufacture as recited in claim 83 wherein the selected quantization level is included in the data frame.

85. An article of manufacture as recited in claim 83 comprising means for entropy encoding the selected quantization level using a first codebook.

86. An article of manufacture as recited in claim 85 comprising means for entropy encoding the gain using a second codebook.

87. An article of manufacture as recited in claim 86 comprising means for entropy encoding the run length count using a third codebook.

88. An article of manufacture as recited in claim 87 comprising means for entropy encoding the quantized wavelet coefficients using a fourth codebook.

89. An article of manufacture as recited in claim 88 comprising means to select the fourth codebook, based on the quantization level, from a set of wavelet coefficient codebooks.

90. An article of manufacture comprising:

a computer usable medium having computer readable code means embedded therein for reconstructing digitally sampled information which has been previously compressed into compressed digitally sampled information, the computer readable code means comprising;

a) means for entropy decoding the compressed digitally sampled information to obtain entropy decoded information comprising wavelet coefficients;

b) means for performing an inverse discrete wavelet transform on the entropy decoded information to obtain inverse wavelet transformed data; and

c) means for rescaling the inverse wavelet transformed data into the time domain to obtain the digitally sampled information.

91. An article of manufacture as recited in claim 90 comprising means for multiplying each wavelet coefficient by a gain factor before the wavelet coefficients are transformed by the means for performing an inverse discrete wavelet transform.

92. An article of manufacture as recited in claim 91 comprising means for forming the previously compressed data into one or more data frames.

93. An article of manufacture as recited in claim 92 wherein the gain factor is found in the one or more data frames.

94. An article of manufacture as recited in claim 90 wherein the means for rescaling the inverse wavelet transformed data into the time domain comprises means for converting the transformed data into integer values.