Network Working Group Ross Finlayson Internet-Draft LIVE.COM Expire in six months 2000.03.10 Category: Informational A More Loss-Tolerant RTP Payload Format for MP3 Audio 1. Status of this Memo This document is an Internet-Draft and is in full conformance with all provisions of Section 10 of RFC2026. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that other groups may also distribute working documents as Internet-Drafts. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet- Drafts as reference material or to cite them other than as "work in progress." The list of current Internet-Drafts can be accessed at http://www.ietf.org/ietf/1id-abstracts.txt The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html. 2. Abstract While the RTP payload format defined in RFC 2250 is generally applicable to all forms of MPEG audio or video, it is less suitable for MPEG 1 or 2, layer III audio (commonly known as "MP3"). The reason for this is that an MP3 frame is not a true "Application Data Unit" - it contains a back-pointer to data in earlier frames, and so cannot be decoded independently of these earlier frames. Because RFC 2250 defines that packet boundaries coincide with frame boundaries, it handles packet loss inefficiently when carrying MP3 data. The loss of an MP3 frame will render some data in previous (or future) frames useless, even if they are received without loss. In this document we define a new RTP payload format for MP3 audio. The new format is essentially the same as that defined in RFC 2250, except for a data-preserving rearrangement of the original MPEG frames, so that packet boundaries now coincide with true MP3 "Application Data Units", which can (optionally) be rearranged in an interleaving pattern. This new format is therefore more data efficient in the face of packet loss. 3. The Structure of MP3 Frames In this section we give a brief overview of the structure of a MP3 frame. (For more detailed description, see the official MPEG 1 audio [2] and MPEG 2 audio [3] specifications.) Each MPEG audio frame begins with a 4-byte header. Information defined by this header includes: - Whether the audio is MPEG 1 or MPEG 2. - Whether the audio is layer I, II, or III. (The remainder of this document assumes layer III, i.e., "MP3") - Whether the audio is mono or stereo. - Whether or not there is a 2-byte CRC field following the header. - (indirectly) The size of the frame. The following structures appear after the header: - (optionally) A 2-byte CRC field - A "side info" structure. This has the following length: - 32 bytes for MPEG 1 stereo - 17 bytes for MPEG 1 mono, or for MPEG 2 stereo - 9 bytes for MPEG 2 mono - Encoded audio data (filling out the rest of the frame) For the purpose of this document, the "side info" structure is the most important, because it defines the location and size of the "Application Data Unit" (ADU) that an MP3 decoder will process. In particular, the "side info" structure defines: - "main_data_begin": This is a back-pointer (in bytes) to the start of the ADU. The back-pointer is counted from the beginning of the frame, and counts only encoded audio data (i.e., ignoring any header, CRC, or "side info" fields). - Several "part2_3_length" fields. These fields - each of which counts bits - are added together to form the length (in bits) of the ADU. Like the back-pointer, this length counts only encoded audio data - not header, CRC, or "side info" fields. An MP3 decoder processes each ADU independently. The ADUs will generally vary in length, but their average length will, of course, be that of the of the MP3 frames (minus the length of the header, CRC, and "side info" fields). (In MPEG literature, this ADU is sometimes referred to as a "bit reservoir".) 4. A New Payload Format As noted in [4], a payload format should be designed so that packet boundaries coincide with "codec frame boundaries" - i.e., with ADUs. The new payload format for MP3 is exactly the same as that defined by RFC 2250 for MPEG audio [1], EXCEPT: 1/ Instead of containing MP3 frames, each packet contains "ADU frames", where an "ADU frame" is defined as: - The 4-byte MPEG header (the same as the original MP3 frame, except that the first byte is replaced with an "interleaving index", as described in section 7) - The optional 2-byte CRC field (the same as the original MP3 frame) - The "side info" structure (the same as the original MP3 frame) - The complete sequence of encoded data for the ADU (padded at the end by zero-bits to fill out a byte boundary) 2/ The (static) payload type 14 that was defined for MPEG audio [5] MUST NOT be used. Instead, a different, dynamic payload type MUST be used - i.e., one in the range [96,127]. Apart from using a different payload type, the use of the RTP header (and the MPEG audio-specific header for fragmentation) is exactly the same as defined in RFC 2250. Note that no information is lost by converting a sequence of MP3 frames to a corresponding sequence of "ADU frames", so a receiving RTP implementation can either feed the ADU frames directly to an appropriately modified MP3 decoder, or convert them back into a sequence of MP3 frames, as described in appendix A.2 below. 5. Handling Multiple MPEG Audio Layers The RTP payload format described here is intended only for MPEG 1 or 2, layer III audio ("MP3"). In contrast, layer I and layer II frames are self-contained, without a back-pointer to earlier frames. However, it is possible (although unusual) for a sequence of audio frames to consist of a mixture of layer III frames and layer I or II frames. When such a sequence is transmitted, only layer III frames are converted to ADUs; layer I or II frames are sent 'as is'. Similarly, the receiver of a sequence of frames - using this payload format - leaves layer I and II frames untouched, but converts layer III frames from "ADU frames" to regular MP3 frames. (Recall that each frame's layer is identified from its 4-byte MPEG header.) 6. Frame Packetizing and Depacketizing The transmission of a sequence of MP3 frames takes the following steps: MP3 frames -1-> ADU frames -2-> interleaved ADU frames -3-> RTP packets Step 1, the conversion of a sequence of MP3 frames to a corresponding sequence of ADU frames, takes place as described in sections 3 and 4. (Note also the pseudo-code in appendix A.1.) Step 2 is the reordering of the sequence of ADU frames in an (optional) interleaving pattern, prior to packetization, as described in section 7 below. (Note also the pseudo-code in appendix B.1.) Interleaving helps reduce the effect of packet loss, by distributing consecutive ADU frames over more than one packet. (Note that because of the back-pointer in MP3 frames, interleaving can be applied - in general - only to ADU frames. Thus, interleaving was not possible for RFC 2250.) Step 3 is the packetizing of a sequence of (interleaved) ADU frames into RTP packets - exactly as specified in RFC 2250. When computing the RTP timestamps for each packet, this step ignores any reordering that took place in step 2, and treats the sequence of packetized ADU frames as if they were generated sequentially. I.e., the RTP timestamps on outgoing packets are always monotonically increasing, regardless of interleaving. Similarly, a sequence of received RTP packets is handled as follows: RTP packets -4-> RTP packets ordered by RTP sequence number -5-> interleaved ADU frames -6-> ADU frames -7-> MP3 frames Step 4 is the usual sorting of incoming RTP packets using the RTP sequence number. Step 5 is the depacketizing of ADU frames from RTP packets - i.e., the reverse of step 3. Note that the size of each ADU frame within a packet must be determined by examining the "part2_3_length" fields in each frame's "sideinfo" structure (see section 3) - not by scanning for a MPEG 'syncword' (because the first byte of this word is reused for interleaving, as described in section 7 below). Step 6 is the rearranging of the sequence of ADU frames back to its original order (except for ADU frames missing due to packet loss), as described in section 7 below. (Note also the pseudo-code in appendix B.2.) Step 7 is the conversion of the sequence of ADU frames into a corresponding sequence of MP3 frames - i.e., the reverse of step 1. (Note also the pseudo-code in appendix A.2.) With an appropriately modified MP3 decoder, an implementation may omit this step; instead, it could feed ADU frames directly to the (modified) MP3 decoder. 7. ADU Frame Interleaving In MPEG audio frames (MPEG 1 or 2; all layers) the high-order byte of the 4-byte MPEG header ('syncword') is always 0xFF. When reordering a sequence of ADU frames for transmission, we reuse this byte as an "Interleaving Sequence Number" (ISN). (This is replaced with 0xFF once again after reordering upon reception.) The structure of the ISN is (a,b), where: - a == high-order 3 bits: Interleave Cycle Count - b == low-order 5 bits: Interleave Index (within Cycle) Example: An interleave cycle of size 8, with a stride of 4, would have a pattern of: 0,4,2,6,1,5,3,7 - giving the following sequence of ISNs: (0,0) (0,4) (0,2) (0,6) (0,1) (0,5) (0,3) (0,7) (1,0) (1,4) (1,2) etc. So, in this example, a sequence of ADU frames f0 f1 f2 f3 f4 f5 f6 f7 f8 f9 (etc.) would get reordered, in step 2, into: (0,0)f0 (0,4)f4 (0,2)f2 (0,6)f6 (0,1)f1 (0,5)f5 (0,3)f3 (0,7)f7 (1,0)f8 (1,4)f12 (1,2)f10 (etc.) and the reverse reordering (along with replacement of the 0xFF) would occur upon reception. The reason for breaking the ISN into "Interleave Cycle Count" and "Interleave Index" (rather than just treating it as a single 8-bit counter) is to give receivers a way of knowing when an ADU frame should be 'released' to the ADU->MP3 conversion process (step 7 above), rather than waiting for more interleaved ADU frames to arrive. E.g., in the example above, when the receiver sees a frame with ISN (1,), it knows that it can release all previously-seen frames with ISN (0,), even if some other (0,) frames remain missing due to packet loss. A 5-bit Interleave Index allows interleave cycles of size up to 32. The choice of an interleaving order can be made independently of RTP packetization. Thus, a simple implementation could choose an interleaving order first, reorder the ADU frames accordingly (step 2), then simply pack them sequentially into RTP packets (step 3). However, the size of ADU frames - and thus the number of ADU frames that will fit in each RTP packet - will typically vary in size, so a more optimal implementation would combine steps 2 and 3, by choosing an interleaving order that better reflected the number of ADU frames packed within each RTP packet. Each receiving implementation of this payload format MUST recognize the ISN and be able to perform deinterleaving of incoming ADU frames (step 6). However, a sending implementation of this payload format MAY choose not to perform interleaving - i.e., by omitting step 2. In this case, the high-order byte in each 4-byte MPEG header would remain at 0xFF. Receiving implementations would thus see a sequence of identical ISNs (all 0xFF), treat this as if the Interleave Cycle Count had wrapped-around, and simply release the sequence of incoming ADU frames sequentially to the ADU->MP3 conversion process (step 7), without reordering. 8. SDP Payload Format Description Pending any future standardization of this payload format, SDP "rtpmap" attributes [6] use the name "X-MP3" to denote this format. 9. Security Considerations The security considerations for this payload format are identical to those noted for RFC 2250 [1]. 10. Acknowledgements The suggestion of adding an interleaving option (using the first byte of the MPEG 'syncword' as an interleaving index) is due to Dave Singer and Stefan Gewinner, who also made additional suggestions to help improve the clarity of this document. 11. References [1] Hoffman, D., Fernando, G., Goyal, V., and Civanlar, M. "RTP Payload Format for MPEG1/MPEG2 Video", RFC 2250, January 1998. [2] ISO/IEC International Standard 11172-3; "Coding of moving pictures and associated audio for digital storage media up to about 1,5 Mbits/s - Part 3: Audio", 1993. [3] ISO/IEC International Standard 13818-3; "Generic coding of moving pictures and associated audio information - Part 3: Audio", 1998. [4] Handley, M. "Guidelines for Writers of RTP Payload Format Specifications" Work-in-Progress, Internet-Draft "draft-ietf-avt-rtp-format-guidelines-04.txt", October 1999. [5] Schulzrinne, H. "RTP Profile for Audio and Video Conferences with Minimal Control" RFC 1890, January 1996. [6] Handley, M., Jacobson, V., "SDP: Session Description Protocol", RFC 2327, April 1998. 9. Author's Address Ross Finlayson, Live Networks, Inc. (LIVE.COM) email: finlayson@live.com WWW: http://www.live.com/ Appendix A. Translating Between "MP3 Frames" and "ADU Frames" The following 'pseudo code' describes how a sender using this payload format can translate a sequence of regular "MP3 Frames" to "ADU Frames", and how a receiver can perform the reverse translation: from "ADU Frames" to "MP3 Frames". We first define the following abstract data structures: - "Segment": A record that represents either a "MP3 Frame" or an "ADU Frame". It consists of the following fields: - "header": the 4-byte MPEG header - "headerSize": a constant (== 4) - "sideInfo": the 'side info' structure, *including* the optional 2-byte CRC field, if present - "sideInfoSize": the size (in bytes) of the above structure - "frameData": the remaining data in this frame - "frameDataSize": the size (in bytes) of the above data - "backpointer": the size (in bytes) of the backpointer for this frame - "aduSize": the size (in bytes) of the ADU associated with this frame. (If the frame is already an "ADU Frame", then aduSize == frameDataSize) - "mp3FrameSize": the total size (in bytes) that this frame would have if it were a regular "MP3 Frame". (If it is already a "MP3 Frame", then mp3FrameSize == headerSize + sideInfoSize + frameDataSize) Note that this size can be derived completely from "header". - "SegmentQueue": A FIFO queue of "Segment"s, with operations - void enqueue(Segment) - Segment dequeue() - Boolean isEmpty() - Segment head() - Segment tail() - Segment previous(Segment): returns the segment prior to a given one - Segment next(Segment): returns the segment after a given one - unsigned totalDataSize(): returns the sum of the "frameDataSize" fields of each entry in the queue A.1 Converting a sequence of "MP3 Frames" to a sequence of "ADU Frames": SegmentQueue pendingMP3Frames; // initially empty while (1) { // Enqueue new MP3 Frames, until we have enough data to generate // the ADU for a frame: do { int totalDataSizeBefore = pendingMP3Frames.totalDataSize(); Segment newFrame = 'the next MP3 Frame'; pendingMP3Frames.enqueue(newFrame); int totalDataSizeAfter = pendingMP3Frames.totalDataSize(); } while (totalDataSizeBefore < newFrame.backpointer || totalDataSizeAfter < newFrame.aduSize); // We now have enough data to generate the ADU for the most // recently enqueued frame (i.e., the tail of the queue). // (The earlier frames in the queue - if any - must be discarded, // as we don't have enough data to generate their ADUs.) Segment tailFrame = pendingMP3Frames.tail(); // Output the header and side info: output(tailFrame.header); output(tailFrame.sideInfo); // Go back to the frame that contains the start of our ADU data: int offset = 0; Segment curFrame = tailFrame; int prevBytes = tailFrame.backpointer; while (prevBytes > 0) { curFrame = pendingMP3Frames.previous(curFrame); int dataHere = curFrame.frameDataSize; if (dataHere < prevBytes) { prevBytes -= dataHere; } else { offset = dataHere - prevBytes; break; } } // Dequeue any frames that we no longer need: while (pendingMP3Frames.head() != curFrame) { pendingMP3Frames.dequeue(); } // Output, from the remaining frames, the ADU data that we want: int bytesToUse = tailFrame.aduSize; while (bytesToUse > 0) { int dataHere = curFrame.frameDataSize - offset; int bytesUsedHere = dataHere < bytesToUse ? dataHere : bytesToUse; output("bytesUsedHere" bytes from curFrame.frameData, starting from "offset"); bytesToUse -= bytesUsedHere; offset = 0; curFrame = pendingMP3Frames.next(curFrame); } } A.2 Converting a sequence of "ADU Frames" to a sequence of "MP3 Frames": SegmentQueue pendingADUFrames; // initially empty while (1) { while (needToGetAnADU()) { Segment newADU = 'the next ADU Frame'; pendingADUFrames.enqueue(newADU); insertDummyADUsIfNecessary(); } generateFrameFromHeadADU(); } Boolean needToGetAnADU() { // Checks whether we need to enqueue one or more new ADUs before // we have enough data to generate a frame for the head ADU. Boolean needToEnqueue = True; if (!pendingADUFrames.isEmpty()) { Segment curADU = pendingADUFrames.head(); int endOfHeadFrame = curADU.mp3FrameSize - curADU.headerSize - curADU.sideInfoSize; int frameOffset = 0; while (1) { int endOfData = frameOffset - curADU.backpointer + curADU.aduSize; if (endOfData >= endOfHeadFrame) { // We have enough data to generate a frame. needToEnqueue = False; break; } frameOffset += curADU.mp3FrameSize - curADU.headerSize - curADU.sideInfoSize; if (curADU == pendingADUFrames.tail()) break; curADU = pendingADUFrames.next(curADU); } } return needToEnqueue; } void generateFrameFromHeadADU() { Segment curADU = pendingADUFrames.head(); // Output the header and side info: output(curADU.header); output(curADU.sideInfo); // Begin by zeroing out the rest of the frame, in case the ADU // data doesn't fill it in completely: int endOfHeadFrame = curADU.mp3FrameSize - curADU.headerSize - curADU.sideInfoSize; output("endOfHeadFrame" zero bytes); // Fill in the frame with appropriate ADU data from this and // subsequent ADUs: int frameOffset = 0; int toOffset = 0; while (toOffset < endOfHeadFrame) { int startOfData = frameOffset - curADU.backpointer; if (startOfData > endOfHeadFrame) { break; // no more ADUs are needed } int endOfData = startOfData + curADU.aduSize; if (endOfData > endOfHeadFrame) { endOfData = endOfHeadFrame; } int fromOffset; if (startOfData <= toOffset) { fromOffset = toOffset - startOfData; startOfData = toOffset; if (endOfData < startOfData) { endOfData = startOfData; } } else { fromOffset = 0; // leave some zero bytes beforehand: toOffset = startOfData; } int bytesUsedHere = endOfData - startOfData; output(starting at offset "toOffset, "bytesUsedHere" bytes from "&curADU.frameData[fromOffset]"); toOffset += bytesUsedHere; frameOffset += curADU.mp3FrameSize - curADU.headerSize - curADU.sideInfoSize; curADU = pendingADUFrames.next(curADU); } pendingADUFrames.dequeue(); } void insertDummyADUsIfNecessary() { // The tail segment (ADU) is assumed to have been recently // enqueued. If its backpointer would overlap the data // of the previous ADU, then we need to insert one or more empty, // 'dummy' ADUs ahead of it. (This situation should occur only if // an intermediate ADU was missing - e.g., due to packet loss.) while (1) { Segment tailADU = pendingADUFrames.tail(); int prevADUend; // relative to the start of the tail ADU if (pendingADUFrames.head() != tailADU) { // there is a previous ADU Segment prevADU = pendingADUFrames.previous(tailADU); prevADUend = prevADU.mp3FrameSize + prevADU.backpointer - prevADU.headerSize - curADU.sideInfoSize; if (prevADU.aduSize > prevADUend) { // this shouldn't happen if the previous ADU // was well-formed prevADUend = 0; } else { prevADUend -= prevADU.aduSize; } } else { prevADUend = 0; } if (tailADU.backpointer > prevADUend) { // Insert a 'dummy' ADU in front of the tail. // This ADU can have the same "header" (and thus // "mp3FrameSize") as the tail ADU, but should have // an "aduSize" of zero. The simplest way to do // this is to copy the "sideInfo" from the tail ADU, // and zero out the "main_data_begin" and all of the // "part2_3_length" fields. } else { break; // no more dummy ADUs need to be inserted } } } Appendix B: Interleaving and Deinterleaving The following 'pseudo code' describes how a sender can reorder a sequence of "ADU Frames" according to an interleaving pattern (step 2), and how a receiver can perform the reverse reordering (step 6). B.1 Interleaving a sequence of "ADU Frames": We first define the following abstract data structures: - "interleaveCycleSize": an integer in the range [1,32] - "interleaveCycle": an array, of size "interleaveCycleSize", containing some permutation of the integers from the set [0 .. interleaveCycleSize-1] e.g., if "interleaveCycleSize" == 8, "interleaveCycle" might contain: 0,4,2,6,1,5,3,7 - "inverseInterleaveCycle": an array containing the inverse of the permutation in "interleaveCycle" - i.e., such that interleaveCycle[inverseInterleaveCycle[i]] == i - "icc": the current Interleave Cycle Count (initially 0) - "ii": the current Interleave Index (initially 0) - "aduFrameBuffer": an array, of size "interleaveCycleSize", of ADU Frames that are awaiting packetization while (1) { int currentISN = inverseInterleaveCycle[ii]; aduFrameBuffer[currentISN] = the next ADU frame; replace the high-order byte of this frame's MPEG header with (icc,ii); if (++ii == interleaveCycleSize) { // we've finished this cycle, so pass all // pending frames to the packetizing step for (int i = 0; i < interleaveCycleSize; ++i) { pass aduFrameBuffer[i] to the packetizing step; } ii = 0; icc = (icc+1)%8; } } B.2 Deinterleaving a sequence of (interleaved) "ADU Frames": We first define the following abstract data structures: - "icc": the Interleave Cycle Count from the current incoming ADU frame - "ii": the Interleave Index from the current incoming ADU frame - "iccLastSeen": the most recently seen Interleave Cycle Count (initially, some integer *not* in the range [0,7]) - "iiLastSeen": the most recently seen Interleave Index (initially, some integer *not* in the range [0,31]) - "aduFrameBuffer": an array, of size 32, of (pointers to) ADU Frames that have just been depacketized (initially, all entries are NULL) while (1) { aduFrame = the next ADU frame from the depacketizing step; (icc,ii) = "the high-order byte of aduFrame's MPEG header"; "the high-order byte of aduFrame's MPEG header" = 0xFF; if (icc != iccLastSeen || ii == iiLastSeen) { // We've started a new interleave cycle // (or interleaving was not used). Release all // pending ADU frames to the ADU->MP3 conversion step: for (int i = 0; i < 32; ++i) { if (aduFrameBuffer[i] != NULL) { release aduFrameBuffer[i]; aduFrameBuffer[i] = NULL; } } } iccLastSeen = icc; iiLastSeen = ii; aduFrameBuffer[ii] = aduFrame; }