[flac-dev] FLAC as a format for archiving non-audio (SDR) sample data?
a.j.buxton at gmail.com
Wed Apr 7 13:00:07 UTC 2021
Thanks for the hints everyone.
It looks like the metadata interface has everything I need, however I am
having trouble with the encoder itself.
We have found that the reference encoder can't handle more than 2^36
samples properly. It encodes the data and calculates md5sum correctly but
truncates the total samples, and then the decoder thinks there is a
signature mismatch because it only checks the truncated number of samples.
I have reported this here: https://github.com/xiph/flac/issues/237
Building the seektable is also a problem since it goes at the beginning of
the file and we won't know how big to make it when recording from a live
source. We will need up to 20MB of seek table for the longest recordings,
but this means every recording will be at least 20MB regardless of length.
Adding new seekpoints after encoding is also extremely slow as metaflac
seems to decode the entire file rather than using the existing seekpoints.
We will also need parallel encoding to reach the target of encoding up to
80MB/s of data. I can't see any way of doing this with the public API since
the stream encoder callback delivers arbitrary buffers rather than frames,
meaning the results of two encoders can't easily be combined.
I know I could avoid all these problems by splitting the recording at
arbitrary points but having to manage a lot of small files is something I
specifically want to avoid.
On Thu, 1 Apr 2021 at 16:58, Roy Galaasen <roy.galaasen at gmail.com> wrote:
> I was going to answer this question yesterday, so I was refreshing my mind
> a bit by doing some research, but life has a habit of getting in the way.
> My response would pretty much be like the one that was given by Martijn
> van Beurden, but differ a bit regarding the sample rate:
> There are two places where sample rate is given:
> In the STREAMINFO block and in each frame header themselves.
> It says in the documentation that the STREAMINFO sample rate is a 20 bit
> value, but it is limited by the structure of the frame headers to 655350
> Hz. As opposed to the seemingly possible 2^20 Hz.
> If you look at the frame header there is a four bit field combination to
> select from a set number of fixed sample rates. If most significant bit is
> set, you can specify a more flexible value in terms of 8 bit sample rate in
> terms of kHz, or 16 bits value in terms of Hz or tens of Hz, and lastly to
> use the value that is given in the STREAMINFO header.
> That won’t of course work if you start reading in the middle of the
> stream, but as an archival format, this value will always be available,
> should it not somehow get corrupted.
> I am curious, though about the limitation mentioned, since it seems to me
> that you could put any value (except 0) into the STREAMINFO header and then
> just put %0000 into the sample rate field of each frame header to reference
> the STREAMINFO value instead, even values that are higher than the
> mentioned maximum of 655350 Hz.
> That makes me wonder why the documentation mentions a limitation, when it
> in theory should work with values up to 1048576 -1. Is there something that
> is not coming out from the documentation?
> There is something called “subset”, which is a set of criteria which is
> set to ensure that the file is streamable. You can’s specify to use the
> streaminfo sample rate if you want to confirm to the subset. Maybe that is
> what the documentation is referring to? Your need is already stated as
> being “archival”, hence you do not need to conform to the subset.
> Roy Galaasen.
> I am in no way affiliated with this project, just an enthusiast that has
> written a complete decoder some time in the past.
> On 1 Apr 2021, at 14:06, Martijn van Beurden <mvanb1 at gmail.com> wrote:
> Considering the sample rate, I'm pretty sure you'll have to resort to
> metadata to store that information. The FLAC format can't be extended to
> store such high sample rates in the frame headers. However, the normal
> vorbis tags accept any key=value pair you want to use, no restrictions.
> For inserting cue marks, using a cuesheet is a rather standard way to do
> this for audio, however, as these use CDDA frames for location, this might
> not be the best way. Fortunately, you can define your own metadata format
> as an 'application' metadata block. See https://xiph.org/flac/format.html
> for more information, see METADATA_BLOCK_APPLICATION. I don't think using
> an OGG container will add anything not already possible with FLAC, except
> when using multiple streams, which I don't think is the case here.
> For not rewriting a whole file, one can add a padding metadata block at
> capture, and use this padding to write metadata afterwards.
> Considering Python, there are quite a few options actually. I've been
> using https://pypi.org/project/SoundFile/, but I think it is a good idea
> to try a few and see which one suits your needs best. I only used it for
> reading FLAC files, so I can't really comment there.
> Kind regards,
> Martijn van Beurden
> flac-dev mailing list
> flac-dev at xiph.org
a.j.buxton at gmail.com
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the flac-dev