[Libav-user] Handling of 24 bit audio in libav* and libswresample

[Paul B Mahol]

On 6/7/13, Hendrik Schreiber <hs at tagtraum.com> wrote:
> On Jun 4, 2013, at 1:34 PM, Paul B Mahol wrote:
>
>> On 6/4/13, Hendrik Schreiber <hs at tagtraum.com> wrote:
>>>
>>> 1. (SHIFTING) When decoding, 24bit audio is apparently shifted, i.e.
>>> 24bit
>>> become 32bit, as there is no 24bit AVSampleFormat. Am I right to assume
>>> that
>>> the data is shifted toward the most significant byte? I.e. the most
>>> significant 3 bytes are the same as the original 24bit?
>>> Or is the most significant byte simply "sign-extended" and the three
>>> least
>>> significant bytes are the original 24bit?
>
> The first statement is true.
>
> I did some tests and all libav does it shift the data toward the most
> significant byte. I.e. the least significant byte is 0. This means, that one
> has to apply dithering, *if* one wants to use this 4 bytes representation
> for anything other than extracting the most significant three bytes.
>
> If one just wants to dump 3-byte for each 24bit sample, one has to simply
> cut off that extra byte added before (encoding with, AV_CODEC_ID_PCM_S24LE
> see below). No dithering necessary.
>
>>>
>>> 2. (SWRESAMPLE) I'm using libswresample to, well, resample data, get rid
>>> of
>>> planar formats etc. It's working great. libswresample also accepts
>>> AVSampleFormat parameters for input and output format. This implies that
>>> it
>>> does not support any conversion to true 24bit, represented by 3 bytes.
>>> Correct?
>
> Yes. I fiddled with it some more. swresample does not support any true 24bit
> (i.e. 3byte per sample) output. It works strictly on the intermediate
> dataformats defined in AVSampleFormat.
>
>>>
>>> 3. (CODEC) What is the recommend way to produce 24bit audio? After
>>> decoding
>>> (and potentially resampling), should I use the corresponding codec (e.g.
>>> AV_CODEC_ID_PCM_S24LE) to produce the data in the format I'm interested
>>> in?
>>> Or is there another, better way?
>>
>> There should be dithering applied, see output_sample_bits option.
>
> I guess there is no other way (expect for perhaps filtering). It turns out
> that to produce 24bit audio in true a 24bit format, one has to use an
> appropriate encoder, e.g. AV_CODEC_ID_PCM_S24LE for signed 24bit little
> endian.

Note that dithering should be done when doing 32bit to 24bit case
and source audio have >24bits used.

>
> Dithering is only necessary, when converting the data somewhere in between
> (e.g. changing the sample rate while it's in 32bit format), as the code in
> pcm.c (macro ENCODE) simply shifts the 32bit representation by 8bit,
> essentially just dropping the last 8bits.

Because last 8bits are always zero for that particular case.

>
> Since I got the answers to all my questions - I figured, I might as well
> post them. Hope it's useful to someone else.
>
> Cheers,
>
> -hendrik
> _______________________________________________
> Libav-user mailing list
> Libav-user at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/libav-user
>