[FFmpeg-devel] [PATCH] Set correct frame_size for Speex decoding

[Justin Ruggles]

Michael Niedermayer wrote:

> On Sun, Aug 16, 2009 at 10:01:53PM -0400, Justin Ruggles wrote:
>> Michael Niedermayer wrote:
>>
>>> On Sun, Aug 16, 2009 at 12:26:32PM -0400, Justin Ruggles wrote:
>>>> Michael Niedermayer wrote:
>>>>
>>>>> On Sun, Aug 16, 2009 at 05:07:18PM +0200, Michael Niedermayer wrote:
>>>>>> On Sat, Aug 15, 2009 at 07:55:46PM -0400, Justin Ruggles wrote:
>>>>>>> Michael Niedermayer wrote:
>>>>> [...]
>>>>>>>>>> what exactly is the argument you have that speex should not be handled like
>>>>>>>>>> every other codec?!
>>>>>>>>>> split it in a parser, the muxer muxes ONLY a single speex packet per
>>>>>>>>>> container packet. Any extension from that is low priority and "patch welcome"
>>>>>>>>>> IMHO ...
>>>>>>>>> The downside for Speex is the container overhead since individual frames
>>>>>>>>> are very small.
>>>>>>>> this is true for many (most?) speech codecs
>>>>>>>>
>>>>>>>> also if we write our own speex encoder, it will only return one frame at a
>>>>>>>> time.
>>>>>>> Why would it have to?  
>>>>>> because the API requires it, both encoders and decoders have to implement the
>>>>>> API, a video encoder also cannot return 5 frames in one packet.
>>>>>> APIs can be changed but you arent arguing for a API change you argue for
>>>>>> ignoring the API and just doing something else.
>>>>>>
>>>>>>
>>>>>>> If the user sets frame_size before initialization
>>>>>>> to a number of samples that is a multiple of a single frame, it could
>>>>>>> return multiple speex frames at once, properly joined together and
>>>>>>> padded at the end.  With libspeex this is very easy to do because the
>>>>>>> functionality is built into the API.
>>>>>>>
>>>>>>> I understand the desire to keep what are called frames as separate
>>>>>>> entities, but in the case of Speex I see it as more advantagous to allow
>>>>>>> the user more control when encoding.  If frames are always split up,
>>>>>>> this limits the users options for both stream copy *and* for encoding to
>>>>>>>  just 1 frame per packet.
>>>>>>>
>>>>>>> If you're dead-set against this idea, then I will finish the parser that
>>>>>>> splits packets in order to continue with my other Speex patches, but I
>>>>>>> don't like how limiting it would be for the user.
>>>>>> i am againt speex handling things different than other speech codecs
>>>>>> based on arguments that apply to other speech codecs as well.
>>>>>> Also iam against passing data between muxer and codec layers in a way
>>>>>> that violates the API.
>>>>> ffmpeg seperates muxer and codec layers, writing a demuxer & decoder
>>>>> that depend on things beyond the API (frames per frame) is going to
>>>>> break things. We had similar great code (passing structs in AVPacket.data
>>>>> because it was convenient) that also didnt turn out to be that great
>>>>> and required a complete rewrite ...
>>>>>
>>>>>
>>>>> ive alraedy said nut doesnt allow multiple frames per packet, but its
>>>>> not just nut, avi as well doesnt allow multiple frames per packet
>>>>> for VBR and either way avi needs to have its headers set up properly,
>>>>> not with fake frame size and such and flv as we alaredy know has a 
>>>>> issue with >8 frames per frame. All that is just what we know of will
>>>>> break if you implement your hack, what else will break is something we
>>>>> only would learn after some time.
>>>>>
>>>>> IMHO, demuxer->parser->splited frames [unless it is not possible to split]
>>>>> if a muxer can store multiple frames it can merge several depending on its
>>>>> abilities and user provided parameters, that merging could also be done
>>>>> as a bitstream filter.
>>>>> But just skiping the spliting and merging and hoping that every container
>>>>> would allow anyting that any other container allowed is just not going to
>>>>> work. And even more so as we already know of many combinations that would
>>>>> noz work
>>>> I do understand your point.  There is a subtle difference with speex
>>>> though.  The process of merging of frames into groups of frames is
>>>> something that is specified by the codec itself, not the container.  To
>>>> the container, it would be as transparent as for an audio codec that
>>>> allows different numbers of samples/duration for a frame.  Nut would
>>>> support it just fine, as it does with FLAC with different numbers of
>>>> samples per frame.  As for FLV, it would be the same as if it doesn't
>>>> allow over a certain number of samples per frame before getting choppy
>>>> and/or crashing.
>>> hmmm, could you quote the part of the speex spec that supports your view?
>>> and provide a link to that spec. (iam asking for a quote instead of just
>>> url cuz of the ML archive ...)
>> unfortunately there is not a speex specification per-se, but a speex
>> manual that gives some hints and guidance.  the libspeex code is
>> definitive reference.
>>
>> http://speex.org/docs/manual/speex-manual/node7.html
>>
>> "Sometimes it is desirable to pack more than one frame per packet (or
>> other basic unit of storage). The proper way to do it is to call
>> speex_encode $ N$  times before writing the stream with speex_bits_write.
>>
>> In cases where the number of frames is not determined by an out-of-band
>> mechanism, it is possible to include a terminator code. That terminator
>> consists of the code 15 (decimal) encoded with 5 bits, as shown in Table
>> 4. Note that as of version 1.0.2, calling speex_bits_write automatically
>> inserts the terminator so as to fill the last byte. This doesn't
>> involves any overhead and makes sure Speex can always detect when there
>> is no more frame in a packet."
>>
>> The end of the packet is able detected because it looks like the start
>> of another frame, but the mode id is the terminator code.  This is
>> needed because a full valid speex frame can be only 5 bits.
>>
>> What happens when libspeex writes frames is this:
>> If all the combined frames end on a byte boundary, no terminator code is
>> used because the packet data will end at the end of a frame.  If the
>> combined frames do not end on a byte boundary, a 0-bit followed by
>> enough 1's to pad to the next byte are appended.  The result is that if
>> there are 5 or more bits, the terminator code will be detected.
> 
> iam not truely happy about all that but i agree that its overall alot
> more convenient and simpler to keep these bit packed pieces of bull shit
> together but in that case it has to be considered that this is a whole
> frame and all variables have to be set accordingly.

It is definitely a lot simpler.  I agree it's ugly, but it all works out
ok if we're consistent in the encoder, decoder, and all muxers and
demuxers supporting speex.  I'll send a new patch sometime soon with
more documentation about how speex frames are to be handled throughout
FFmpeg.

-Justin