PCM16 distortion

[chdh]

For the PCM16 format, negative sample values are multiplied by 32768, but positive sample values are multiplied by 32767. This leads to a small distortion.

Consider the following symmetrical float signal without a DC value: 0, 0.5, 0, -0.5, 0
The values written to the WAV file are: 0, 16383, 0, -16384, 0

(0.5 * 32767)|0 ==> 16383
(-0.5 * 32768)|0 ==> -16384

[chdh]

I don't understand the fix. This doesn't solve the problem.

Example:
Math.round(0.99 * 32767) ==> 32439
Math.round(-0.99 * 32768) ==> -32440

Why not just use:
value = Math.round(value * 32767) | 0
(without a distinction whether its negative or positive)

[chdh]

I had a look at how Google Chrome and Firefox decode WAV files with AudioContext.decodeAudioData():

WAV hex	decimal	Chrome	Firefox
0x7FFF	32767	1	0.999969482421875
0x7FFE	32766	0.999969482421875	0.99993896484375
0x4000	16384	0.5000152587890625	0.5
0x3FFF	16383	0.4999847412109375	0.499969482421875
0x0000	0	0	0
0xBFFF	-16385	-0.500030517578125	-0.500030517578125
0xC000	-16384	-0.5	-0.5
0x8002	-32766	-0.99993896484375	-0.99993896484375
0x8001	-32767	-0.999969482421875	-0.999969482421875
0x8000	-32768	-1	-1

Chrome uses an asymmetric scheme, like you currently do, and can encode the float value 1.
Firefox uses a symmetric scheme, but cannot encode the float value 1.

Why not map the integer range -32768 .. +32767 to the float range -1.0000305185 .. +1? This would be symmetric and encoding of the float value 1 would be possible.

[mohayonao]

It seems to me that is decoder problem.

[chdh]

I agree, The problem occurs when the encoder and decoder do not use the same mapping between float (-1 to 1) and int (-32768 to 32767). Your encoder and decoder use the same mapping as the decoder of Google Chrome, but Firefox is different.

[mohayonao]

This issue move to mohayonao/wav-decoder#14.

[chdh]

The same "symmetric" option would also make sense for the encoder.

[mohayonao]

How do encoder map values with symmetric options with int16?

signal	default	symmetric
1	32767
0.5	16384
0	0
-0.5	-16384
-1	-32768

[chdh]

signal	default	symmetric V1	symmetric V2 (Firefox)
1	32767	32767	32767 (clipped)
0.9999694824	32766	32766	32767
0.75	24575	24575	24576
0.5	16384	16384	16384
0	0	0	0
-0.5	-16384	-16384	-16384
-1	-32768	-32767	-32768
-1.0000305185	-32768 (clipped)	-32768	-32768 (clipped)

[mohayonao]

Thanks.

I can't accept "symmetric V1".
Because, "symmetric V1" encode -1 to -32767, but decoder does not decode -32767 to -1.

The code of "symmetric V2" is below? I agree that is not bad as optional variation.

function encode_pcm16(value) {
  value = Math.max(-1, Math.min(value, +1));

  return Math.min(Math.round(value * 32768), 32767);
}

abs signal	positive	negative
1	32768	-32767
0.9999847412109375	32768	-32767
0.9999846816062927	32767	-32767

[chdh]

or shorter:

function encode_pcm16(value) {
   return Math.max(-32768, Math.min(32767, Math.round(value * 32768)));
}

[mohayonao]

Okay. I've implemented and merged it to v1.3.0. Thanks.

[chdh]

Thanks for your work.