A.2 Subjective relevance of the speech coder output bits

06.103GPPFull Rate Speech TranscodingTS

Since no valid objective quality criterion for speech signals is available, the only way to build up such a relevance table is to perform listening tests. The procedure described below was used to obtain the relevance classification given in table A.2.1 of the recommendation.

To classify a single bit, say bit i of parameter k, a short speech signal (2 sec) was encoded, then this bit was inverted in each frame (the other bits were left unchanged) and the resulting bit stream was fed into the speech decoder. The listeners had to compare the quality of the signal with the quality of six reference signals with different levels of distortion. Repeating this procedure for all bits would result in a subdivision of the 260 bits into six relevance classes. It can be observed that many of the bits have the same physical meaning and it can be expected that bits with the same meaning have the same relevance (e.g. the MSB’s of the RPE samples). Relying on this assumption, only one of the equivalent parameters was considered. Since there are 13 parameters with different physical meaning with 56 bits in total, the number of tests is reduced from 260 to 56.

The reference signals were the same speech signal distorted by inverting one of the six bits of LAR coefficient number one. This resulted in an adequate quantization of distortion levels ranging from "not intelligible" (MSB inverted) to "negligible distortion" (LSB inverted).

The test was carried out using three listeners and one female speaker. Since the three listeners came to rather similar results, no more listeners were considered to be required. Averaging the three outcomes led to the relevance table given in table A.2.1, where the order of all bits between two successive bits of the first parameter (LAR 1) are arbitrarily chosen.

Table A.2.1a: Subjective importance of encoded bits
(the parameter and bit numbers refer to table 1.1)

Importance

Parameter

Parameter

Bit number

class

name

number

1

Log.area ratio 1

1

b6

Block amplitude

12,29,46,63

b53,b109,b165,b221

Log.area ratio 1

1

b5

2

Log.area ratio 2

2

b12

Log.area ratio 3

3

b17

Log.area ratio 1

1

b4

Log.area ratio 2

2

b11

Log.area ratio 3

3

b16

Log.area ratio 4

4

b22

LTP lag

9,26,43,60

b43,b99,b155,b211

3

Block amplitude

12,29,46,63

b52,b108,b164,b220

Log.area ratio 2,5,6

2,5,6

b10,b26,b30

LTP lag

9,26,43,60

b42,b98,b154,b210

LTP lag

9,26,43,60

b41,b97,b153,b209

LTP lag

9,26,43,60

b40,b96,b152,b208

LTP lag

9,26,43,60

b39,b95,b151,b207

Block amplitude

12,29,46,63

b51,b107,b163,b219

Log.area ratio 1

1

b3

Log.area ratio 4

4

b21

Log.area ratio 7

7

b33

4

LTP lag

9,26,43,60

b38,b94,b150,b206

Log.area ratio 5,6

5,6

b25,b29

LTP gain

10,27,44,61

b45,b101,b157,b213

LTP lag

9,26,43,60

b37,b93,b149,b205

Grid position

11,28,45,62

b47,b103,b159,b215

Table A.2.1b: Subjective importance of encoded bits
(the parameter and bit numbers refer to table 1.1)

Importance

Parameter

Parameter

Bit number

class

name

number

Log.area ratio 1

1

b2

Log.area ratio 2,3,8,4

2,3,8,4

b9,b15,b36,b20

Log.area ratio 5,7

5,7

b24,b32

LTP gain

10,27,44,61

b44,b100,b156,b212

Block amplitude

12,29,46,63

b50,b106,b162,b218

RPE pulses

13..25

b56,b59,..,b92

RPE pulses

30..42

b112,b115,..,b148

RPE pulses

47..59

b168,b171,..,b204

5

RPE pulses

64..76

b224,b227,..,b260

Grid position

11,28,45,62

b46,b102,b158,b214

Block amplitude

12,29,46,63

b49,b105,b161,b217

RPE pulses

13..25

b55,b58,..,b91

RPE pulses

30..42

b111,b114,..,b147

RPE pulses

47..59

b167,b170,..,b203

RPE pulses

64..67

b223,b226,b229,b232

RPE pulses

68..76

b235,b238,..,b259

Log.area ratio 1

1

b1

Log.area ratio 2,3,6

2,3,6

b8,b14,b28

Log.area ratio 7

7

b31

Log.area ratio 8

8

b35

Log.area ratio 8,3

8,3

b34,b13

Log.area ratio 4

4

b19

6

Log.area ratio 4,5

4,5

b18,b23

Block amplitude

12,29,46,63

b48,b104,b160,b216

RPE pulses

13..25

b54,b57,..,b90

RPE pulses

30..42

b110,b113,..,b146

RPE pulses

47..59

b166,b169,..,b202

RPE pulses

64..76

b222,b225,..,b258

Log.area ratio 2,6

2,6

b7,b27