C.  Session Configuration Tools

   In Sections 6.1-2 of the main text, we show session descriptions for
   minimal native and mpeg4-generic RTP MIDI streams.  Minimal streams
   lack the flexibility to support some applications.  In this appendix,
   we describe how to customize stream behavior through the use of the
   payload format parameters.

   The appendix begins with 6 sections, each devoted to parameters that
   affect a particular aspect of stream behavior:

     o  Appendix C.1 describes the stream subsetting system (cm_unused
        and cm_used).

     o  Appendix C.2 describes the journalling system (ch_anchor,
        ch_default, ch_never, j_sec, j_update).

     o  Appendix C.3 describes MIDI command timestamp semantics
        (linerate, mperiod, octpos, tsmode).

     o  Appendix C.4 describes the temporal duration ("media time") of
        an RTP MIDI packet (guardtime, rtp_maxptime, rtp_ptime).

     o  Appendix C.5 concerns stream description (musicport).

     o  Appendix C.6 describes MIDI rendering (chanmask, cid, inline,
        multimode, render, rinit, subrender, smf_cid, smf_info,
        smf_inline, smf_url, url).

   The parameters listed above may optionally appear in session
   descriptions of RTP MIDI streams.  If these parameters are used in an
   SDP session description, the parameters appear on an fmtp attribute
   line.  This attribute line applies to the payload type associated
   with the fmtp line.

   The parameters listed above add extra functionality ("features") to
   minimal RTP MIDI streams.  In Appendix C.7, we show how to use these
   features to support two classes of applications: content-streaming
   using RTSP (Appendix C.7.1) and network musical performance using SIP
   (Appendix C.7.2).

   The participants in a multimedia session MUST share a common view of
   all of the RTP MIDI streams that appear in an RTP session, as defined
   by a single media (m=) line.  In some RTP MIDI applications, the
   "common view" restriction makes it difficult to use sendrecv streams
   (all parties send and receive), as each party has its own
   requirements.  For example, a two-party network musical performance
   application may wish to customize the renderer on each host to match
   the CPU performance of the host [NMP].

   We solve this problem by using two RTP MIDI streams -- one sendonly,
   one recvonly -- in lieu of one sendrecv stream.  The data flows in
   the two streams travel in opposite directions, to control receivers
   configured to use different renderers.  In the third example in
   Appendix C.5, we show how the musicport parameter may be used to
   define virtual sendrecv streams.

   As a general rule, the RTP MIDI protocol does not handle parameter
   changes during a session well, because the parameters describe
   heavyweight or stateful configuration that is not easily changed once
   a session has begun.  Thus, parties SHOULD NOT expect that parameter
   change requests during a session will be accepted by other parties.
   However, implementors SHOULD support in-session parameter changes
   that are easy to handle (for example, the guardtime parameter defined
   in Appendix C.4) and SHOULD be capable of accepting requests for
   changes of those parameters, as received by its session management
   protocol (for example, re-offers in SIP [RFC3264]).

   Appendix D defines the Augmented Backus-Naur Form (ABNF, [RFC4234])
   syntax for the payload parameters.  Section 11 provides information
   to the Internet Assigned Numbers Authority (IANA) on the media types
   and parameters defined in this document.

   Appendix C.6.5 defines the media type "audio/asc", a stored object
   for initializing mpeg4-generic renderers.  As described in Appendix
   C.6, the audio/asc media type is assigned to the "rinit" parameter to
   specify an initialization data object for the default mpeg4-generic
   renderer.  Note that RTP stream semantics are not defined for
   "audio/asc".  Therefore, the "asc" subtype MUST NOT appear on the
   rtpmap line of a session description.

C.1.  Configuration Tools: Stream Subsetting

   As defined in Section 3.2 in the main text, the MIDI list of an RTP
   MIDI packet may encode any MIDI command that may legally appear on a
   MIDI 1.0 DIN cable.

   In this appendix, we define two parameters (cm_unused and cm_used)
   that modify this default condition, by excluding certain types of
   MIDI commands from the MIDI list of all packets in a stream.  For
   example, if a multimedia session partitions a MIDI name space into
   two RTP MIDI streams, the parameters may be used to define which
   commands appear in each stream.

   In this appendix, we define a simple language for specifying MIDI
   command types.  If a command type is assigned to cm_unused, the
   commands coded by the string MUST NOT appear in the MIDI list.  If a
   command type is assigned to cm_used, the commands coded by the string
   MAY appear in the MIDI list.

   The parameter list may code multiple assignments to cm_used and
   cm_unused.  Assignments have a cumulative effect and are applied in
   the order of appearance in the parameter list.  A later assignment of
   a command type to the same parameter expands the scope of the earlier
   assignment.  A later assignment of a command type to the opposite

   parameter cancels (partially or completely) the effect of an earlier
   assignment.

   To initialize the stream subsetting system, "implicit" assignments to
   cm_unused and cm_used are processed before processing the actual
   assignments that appear in the parameter list.  The System Common
   undefined commands (0xF4, 0xF5) and the System Real-Time Undefined
   commands (0xF9, 0xFD) are implicitly assigned to cm_unused.  All
   other command types are implicitly assigned to cm_used.

   Note that the implicit assignments code the default behavior of an
   RTP MIDI stream as defined in Section 3.2 in the main text (namely,
   that all commands that may legally appear on a MIDI 1.0 DIN cable may
   appear in the stream).  Also note that assignments of the System
   Common undefined commands (0xF4, 0xF5) apply to the use of these
   commands in the MIDI source command stream, not the special use of
   0xF4 and 0xF5 in SysEx segment encoding defined in Section 3.2 in the
   main text.

   As a rule, parameter assignments obey the following syntax (see
   Appendix D for ABNF):

     <parameter> = [channel list]<command-type list>[field list]

   The command-type list is mandatory; the channel and field lists are
   optional.

   The command-type list specifies the MIDI command types for which the
   parameter applies.  The command-type list is a concatenated sequence
   of one or more of the letters (ABCFGHJKMNPQTVWXYZ).  The letters code
   the following command types:

      o  A: Poly Aftertouch (0xA)
      o  B: System Reset (0xFF)
      o  C: Control Change (0xB)
      o  F: System Time Code (0xF1)
      o  G: System Tune Request (0xF6)
      o  H: System Song Select (0xF3)
      o  J: System Common Undefined (0xF4)
      o  K: System Common Undefined (0xF5)
      o  N: NoteOff (0x8), NoteOn (0x9)
      o  P: Program Change (0xC)
      o  Q: System Sequencer (0xF2, 0xF8, 0xF9, 0xFA, 0xFB, 0xFC)
      o  T: Channel Aftertouch (0xD)
      o  V: System Active Sense (0xFE)
      o  W: Pitch Wheel (0xE)

      o  X: SysEx (0xF0)
      o  Y: System Real-Time Undefined (0xF9)
      o  Z: System Real-Time Undefined (0xFD)

   In addition to the letters above, the letter M may also appear in the
   command-type list.  The letter M refers to the MIDI parameter system
   (see definition in Appendix A.1 and in [MIDI]).  An assignment of M
   to cm_unused codes that no RPN or NRPN transactions may appear in the
   MIDI list.

   Note that if cm_unused is assigned the letter M, Control Change (0xB)
   commands for the controller numbers in the standard controller
   assignment might still appear in the MIDI list.  For an explanation,
   see Appendix A.3.4 for a discussion of the "general-purpose" use of
   parameter system controller numbers.

   In the text below, rules that apply to "MIDI voice channel commands"
   also apply to the letter M.

   The letters in the command-type list MUST be uppercase and MUST
   appear in alphabetical order.  Letters other than
   (ABCFGHJKMNPQTVWXYZ) that appear in the list MUST be ignored.

   For MIDI voice channel commands, the channel list specifies the MIDI
   channels for which the parameter applies.  If no channel list is
   provided, the parameter applies to all MIDI channels (0-15).  The
   channel list takes the form of a list of channel numbers (0 through
   15) and dash-separated channel number ranges (i.e., 0-5, 8-12, etc).
   Dots (i.e., "." characters) separate elements in the channel list.

   Recall that System commands do not have a MIDI channel associated
   with them.  Thus, for most command-type letters that code System
   commands (B, F, G, H, J, K, Q, V, Y, and Z), the channel list is
   ignored.

   For the command-type letter X, the appearance of certain numbers in
   the channel list codes special semantics.

     o  The digit 0 codes that SysEx "cancel" sublists (Section 3.2 in
        the main text) MUST NOT appear in the MIDI list.

     o  The digit 1 codes that cancel sublists MAY appear in the MIDI
        list (the default condition).

     o  The digit 2 codes that commands other than System Real-time MIDI
        commands MUST NOT appear between SysEx command segments in the
        MIDI list (the default condition).

     o  The digit 3 codes that any MIDI command type may appear between
        SysEx command segments in the MIDI list, with the exception of
        the segmented encoding of a second SysEx command (verbatim SysEx
        commands are OK).

   For command-type X, the channel list MUST NOT contain both digits 0
   and 1, and it MUST NOT contain both digits 2 and 3.  For command-type
   X, channel list numbers other than the numbers defined above are
   ignored.  If X does not have a channel list, the semantics marked
   "the default condition" in the list above apply.

   The syntax for field lists in a parameter assignment follows the
   syntax for channel lists.  If no field list is provided, the
   parameter applies to all controller or note numbers.

   For command-type C (Control Change), the field list codes the
   controller numbers (0-255) for which the parameter applies.

   For command-type M (Parameter System), the field list codes the
   Registered Parameter Numbers (RPNs) and Non-Registered Parameter
   Numbers (NRPNs) for which the parameter applies.  The number range
   0-16383 specifies RPNs, the number range 16384-32767 specifies NRPNs
   (16384 corresponds to NRPN 0, 32767 corresponds to NRPN 16383).

   For command-types N (NoteOn and NoteOff) and A (Poly Aftertouch), the
   field list codes the note numbers for which the parameter applies.

   For command-types J and K (System Common Undefined), the field list
   consists of a single digit, which specifies the number of data octets
   that follow the command octet.

   For command-type X (SysEx), the field list codes the number of data
   octets that may appear in a SysEx command.  Thus, the field list
   0-255 specifies SysEx commands with 255 or fewer data octets, the
   field list 256-4294967295 specifies SysEx commands with more than 255
   data octets but excludes commands with 255 or fewer data octets, and
   the field list 0 excludes all commands.

   A secondary parameter assignment syntax customizes command-type X
   (see Appendix D for complete ABNF):

     <parameter> = "__" <h-list> ["_" <h-list>] "__"

   The assignment defines the class of SysEx commands that obeys the
   semantics of the assigned parameter.  The command class is specified
   by listing the permitted values of the first N data octets that
   follow the SysEx 0xF0 command octet.  Any SysEx command whose first N
   data octets match the list is a member of the class.

   Each <h-list> defines a data octet of the command, as a dot-separated
   (".") list of one or more hexadecimal constants (such as "7F") or
   dash-separated hexadecimal ranges (such as "01-1F").  Underscores
   ("_") separate each <h-list>.  Double-underscores ("__") delineate
   the data octet list.

   Using this syntax, each assignment specifies a single SysEx command
   class.  Session descriptions may use several assignments to cm_used
   and cm_unused to specify complex behaviors.

   The example session description below illustrates the use of the
   stream subsetting parameters:

   v=0
   o=lazzaro 2520644554 2838152170 IN IP6 first.example.net
   s=Example
   t=0 0
   m=audio 5004 RTP/AVP 96
   c=IN IP6 2001:DB80::7F2E:172A:1E24
   a=rtpmap:96 rtp-midi/44100
   a=fmtp:96 cm_unused=ACGHJKNMPTVWXYZ; cm_used=__7F_00-7F_01_01__

   The session description configures the stream for use in clock
   applications.  All voice channels are unused, as are all System
   Commands except those used for MIDI Time Code (command-type F, and
   the Full Frame SysEx command that is matched by the string assigned
   to cm_used), the System Sequencer commands (command-type Q), and
   System Reset (command-type B).

C.2.  Configuration Tools: The Journalling System

   In this appendix, we define the payload format parameters that
   configure stream journalling and the recovery journal system.

   The j_sec parameter (Appendix C.2.1) sets the journalling method for
   the stream.  The j_update parameter (Appendix C.2.2) sets the
   recovery journal sending policy for the stream.  Appendix C.2.2 also
   defines the sending policies of the recovery journal system.

   Appendix C.2.3 defines several parameters that modify the recovery
   journal semantics.  These parameters change the default recovery
   journal semantics as defined in Section 5 and Appendices A-B.

   The journalling method for a stream is set at the start of a session
   and MUST NOT be changed thereafter.  This requirement forbids changes
   to the j_sec parameter once a session has begun.

   A related requirement, defined in the appendix sections below,
   forbids the acceptance of parameter values that would violate the
   recovery journal mandate.  In many cases, a change in one of the
   parameters defined in this appendix during an ongoing session would
   result in a violation of the recovery journal mandate for an
   implementation; in this case, the parameter change MUST NOT be
   accepted.

C.2.1.  The j_sec Parameter

   Section 2.2 defines the default journalling method for a stream.
   Streams that use unreliable transport (such as UDP) default to using
   the recovery journal.  Streams that use reliable transport (such as
   TCP) default to not using a journal.

   The parameter j_sec may be used to override this default.  This memo
   defines two symbolic values for j_sec: "none", to indicate that all
   stream payloads MUST NOT contain a journal section, and "recj", to
   indicate that all stream payloads MUST contain a journal section that
   uses the recovery journal format.

   For example, the j_sec parameter might be set to "none" for a UDP
   stream that travels between two hosts on a local network that is
   known to provide reliable datagram delivery.

   The session description below configures a UDP stream that does not
   use the recovery journal:

   v=0
   o=lazzaro 2520644554 2838152170 IN IP4 first.example.net
   s=Example
   t=0 0
   m=audio 5004 RTP/AVP 96
   c=IN IP4 192.0.2.94
   a=rtpmap:96 rtp-midi/44100
   a=fmtp:96 j_sec=none

   Other IETF standards-track documents may define alternative journal
   formats.  These documents MUST define new symbolic values for the
   j_sec parameter to signal the use of the format.

   Parties MUST NOT accept a j_sec value that violates the recovery
   journal mandate (see Section 4 for details).  If a session
   description uses a j_sec value unknown to the recipient, the
   recipient MUST NOT accept the description.

   Special j_sec issues arise when sessions are managed by session
   management tools (like RTSP, [RFC2326]) that use SDP for "declarative
   usage" purposes (see the preamble of Section 6 for details).  For
   these session management tools, SDP does not code transport details
   (such as UDP or TCP) for the session.  Instead, server and client
   negotiate transport details via other means (for RTSP, the SETUP
   method).

   In this scenario, the use of the j_sec parameter may be ill-advised,
   as the creator of the session description may not yet know the
   transport type for the session.  In this case, the session
   description SHOULD configure the journalling system using the
   parameters defined in the remainder of Appendix C.2, but it SHOULD
   NOT use j_sec to set the journalling status.  Recall that if j_sec
   does not appear in the session description, the default method for
   choosing the journalling method is in effect (no journal for reliable
   transport, recovery journal for unreliable transport).

   However, in declarative usage situations where the creator of the
   session description knows that journalling is always required or
   never required, the session description SHOULD use the j_sec
   parameter.

C.2.2.  The j_update Parameter

   In Section 4, we use the term "sending policy" to describe the method
   a sender uses to choose the checkpoint packet identity for each
   recovery journal in a stream.  In the sub-sections that follow, we
   normatively define three sending policies: anchor, closed-loop, and
   open-loop.

   As stated in Section 4, the default sending policy for a stream is
   the closed-loop policy.  The j_update parameter may be used to
   override this default.

   We define three symbolic values for j_update: "anchor", to indicate
   that the stream uses the anchor sending policy, "open-loop", to
   indicate that the stream uses the open-loop sending policy, and
   "closed-loop", to indicate that the stream uses the closed-loop
   sending policy.  See Appendix C.2.3 for examples session descriptions
   that use the j_update parameter.

   Parties MUST NOT accept a j_update value that violates the recovery
   journal mandate (Section 4).

   Other IETF standards-track documents may define additional sending
   policies for the recovery journal system.  These documents MUST
   define new symbolic values for the j_update parameter to signal the

   use of the new policy.  If a session description uses a j_update
   value unknown to the recipient, the recipient MUST NOT accept the
   description.

C.2.2.1.  The anchor Sending Policy

   In the anchor policy, the sender uses the first packet in the stream
   as the checkpoint packet for all packets in the stream.  The anchor
   policy satisfies the recovery journal mandate (Section 4), as the
   checkpoint history always covers the entire stream.

   The anchor policy does not require the use of the RTP control
   protocol (RTCP, [RFC3550]) or other feedback from receiver to sender.
   Senders do not need to take special actions to ensure that received
   streams start up free of artifacts, as the recovery journal always
   covers the entire history of the stream.  Receivers are relieved of
   the responsibility of tracking the changing identity of the
   checkpoint packet, because the checkpoint packet never changes.

   The main drawback of the anchor policy is bandwidth efficiency.
   Because the checkpoint history covers the entire stream, the size of
   the recovery journals produced by this policy usually exceeds the
   journal size of alternative policies.  For single-channel MIDI data
   streams, the bandwidth overhead of the anchor policy is often
   acceptable (see Appendix A.4 of [NMP]).  For dense streams, the
   closed-loop or open-loop policies may be more appropriate.

C.2.2.2.  The closed-loop Sending Policy

   The closed-loop policy is the default policy of the recovery journal
   system.  For each packet in the stream, the policy lets senders
   choose the smallest possible checkpoint history that satisfies the
   recovery journal mandate.  As smaller checkpoint histories generally
   yield smaller recovery journals, the closed-loop policy reduces the
   bandwidth of a stream, relative to the anchor policy.

   The closed-loop policy relies on feedback from receiver to sender.
   The policy assumes that a receiver periodically informs the sender of
   the highest sequence number it has seen so far in the stream, coded
   in the 32-bit extension format defined in [RFC3550].  For RTCP,
   receivers transmit this information in the Extended Highest Sequence
   Number Received (EHSNR) field of Receiver Reports.  RTCP Sender or
   Receiver Reports MUST be sent by any participant in a session with
   closed loop sending policy, unless another feedback mechanism has
   been agreed upon.

   The sender may safely use receiver sequence number feedback to guide
   checkpoint history management, because Section 4 requires that
   receivers repair indefinite artifacts whenever a packet loss event
   occur.

   We now normatively define the closed-loop policy.  At the moment a
   sender prepares an RTP packet for transmission, the sender is aware
   of R >= 0 receivers for the stream.  Senders may become aware of a
   receiver via RTCP traffic from the receiver, via RTP packets from a
   paired stream sent by the receiver to the sender, via messages from a
   session management tool, or by other means.  As receivers join and
   leave a session, the value of R changes.

   Each known receiver k (1 <= k <= R) is associated with a 32-bit
   extended packet sequence number M(k), where the extension reflects
   the sequence number rollover count of the sender.

   If the sender has received at least one feedback report from receiver
   k, M(k) is the most recent report of the highest RTP packet sequence
   number seen by the receiver, normalized to reflect the rollover count
   of the sender.

   If the sender has not received a feedback report from the receiver,
   M(k) is the extended sequence number of the last packet the sender
   transmitted before it became aware of the receiver.  If the sender
   became aware of this receiver before it sent the first packet in the
   stream, M(k) is the extended sequence number of the first packet in
   the stream.

   Given this definition of M(), we now state the closed-loop policy.
   When preparing a new packet for transmission, a sender MUST choose a
   checkpoint packet with extended sequence number N, such that M(k) >=
   (N - 1) for all k, 1 <= k <= R, where R >= 1.  The policy does not
   restrict sender behavior in the R == 0 (no known receivers) case.

   Under the closed-loop policy as defined above, a sender may transmit
   packets whose checkpoint history is shorter than the session history
   (as defined in Appendix A.1).  In this event, a new receiver that
   joins the stream may experience indefinite artifacts.

   For example, if a Control Change (0xB) command for Channel Volume
   (controller number 7) was sent early in a stream, and later a new
   receiver joins the session, the closed-loop policy may permit all
   packets sent to the new receiver to use a checkpoint history that
   does not include the Channel Volume Control Change command.  As a
   result, the new receiver experiences an indefinite artifact, and
   plays all notes on a channel too loudly or too softly.

   To address this issue, the closed-loop policy states that whenever a
   sender becomes aware of a new receiver, the sender MUST determine if
   the receiver would be subject to indefinite artifacts under the
   closed-loop policy.  If so, the sender MUST ensure that the receiver
   starts the session free of indefinite artifacts.

   For example, to solve the Channel Volume issue described above, the
   sender may code the current state of the Channel Volume controller
   numbers in the recovery journal Chapter C, until it receives the
   first RTCP RR report that signals that a packet containing this
   Chapter C has been received.

   In satisfying this requirement, senders MAY infer the initial MIDI
   state of the receiver from the session description.  For example, the
   stream example in Section 6.2 has the initial state defined in [MIDI]
   for General MIDI.

   In a unicast RTP session, a receiver may safely assume that the
   sender is aware of its presence of a receiver from the first packet
   sent in the RTP stream.  However, in other types of RTP sessions
   (multicast, conference focus, RTP translator/mixer), a receiver is
   often not able to determine if the sender is initially aware of its
   presence as a receiver.

   To address this issue, the closed-loop policy states that if a
   receiver participates in a session where it may have access to a
   stream whose sender is not aware of the receiver, the receiver MUST
   take actions to ensure that its rendered MIDI performance does not
   contain indefinite artifacts.  These protections will be necessarily
   incomplete.  For example, a receiver may monitor the Checkpoint
   Packet Seqnum for uncovered loss events, and "err on the side of
   caution" with respect to handling stuck notes due to lost MIDI
   NoteOff commands, but the receiver is not able to compensate for the
   lack of Channel Volume initialization data in the recovery journal.

   The receiver MUST NOT discontinue these protective actions until it
   is certain that the sender is aware of its presence.  If a receiver
   is not able to ascertain sender awareness, the receiver MUST continue
   these protective actions for the duration of the session.

   Note that in a multicast session where all parties are expected to
   send and receive, the reception of RTCP receiver reports from the
   sender about the RTP stream a receiver is multicasting is evidence of
   the sender's awareness that the RTP stream multicast by the sender is
   being monitored by the receiver.  Receivers may also obtain sender
   awareness evidence from session management tools, or by other means.
   In practice, ongoing observation of the Checkpoint Packet Seqnum to
   determine if the sender is taking actions to prevent loss events for

   a receiver is a good indication of sender awareness, as is the sudden
   appearance of recovery journal chapters with numerous Control Change
   controller data that was not foreshadowed by recent commands coded in
   the MIDI list shortly after sending an RTCP RR.

   The final set of normative closed-loop policy requirements concern
   how senders and receivers handle unplanned disruptions of RTCP
   feedback from a receiver to a sender.  By "unplanned", we refer to
   disruptions that are not due to the signalled termination of an RTP
   stream, via an RTCP BYE or via session management tools.

   As defined earlier in this section, the closed-loop policy states
   that a sender MUST choose a checkpoint packet with extended sequence
   number N, such that M(k) >= (N - 1) for all k, 1 <= k <= R, where R
   >= 1.  If the sender has received at least one feedback report from
   receiver k, M(k) is the most recent report of the highest RTP packet
   sequence number seen by the receiver, normalized to reflect the
   rollover count of the sender.

   If this receiver k stops sending feedback to the sender, the M(k)
   value used by the sender reflects the last feedback report from the
   receiver.  As time progresses without feedback from receiver k, this
   fixed M(k) value forces the sender to increase the size of the
   checkpoint history, and thus increases the bandwidth of the stream.

   At some point, the sender may need to take action in order to limit
   the bandwidth of the stream.  In most envisioned uses of RTP MIDI,
   long before this point is reached, the SSRC time-out mechanism
   defined in [RFC3550] will remove the uncooperative receiver from the
   session (note that the closed-loop policy does not suggest or require
   any special sender behavior upon an SSRC time-out, other than the
   sender actions related to changing R, described earlier in this
   section).

   However, in rare situations, the bandwidth of the stream (due to a
   lack of feedback reports from the sender) may become too large to
   continue sending the stream to the receiver before the SSRC time-out
   occurs for the receiver.  In this case, the closed-loop policy states
   that the sender should invoke the SSRC time-out for the receiver
   early.

   We now discuss receiver responsibilities in the case of unplanned
   disruptions of RTCP feedback from receiver to sender.

   In the unicast case, if a sender invokes the SSRC time-out mechanism
   for a receiver, the receiver stops receiving packets from the sender.
   The sender behavior imposed by the guardtime parameter (Appendix

   C.4.2) lets the receiver conclude that an SSRC time-out has occurred
   in a reasonable time period.

   In this case of a time-out, a receiver MUST keep sending RTCP
   feedback, in order to re-establish the RTP flow from the sender.
   Unless the receiver expects a prompt recovery of the RTP flow, the
   receiver MUST take actions to ensure that the rendered MIDI
   performance does not exhibit "very long transient artifacts" (for
   example, by silencing NoteOns to prevent stuck notes) while awaiting
   reconnection of the flow.

   In the multicast case, if a sender invokes the SSRC time-out
   mechanism for a receiver, the receiver may continue to receive
   packets, but the sender will no longer be using the M(k) feedback
   from the receiver to choose each checkpoint packet.  If the receiver
   does not have additional information that precludes an SSRC time-out
   (such as RTCP Receiver Reports from the sender about an RTP stream
   the receiver is multicasting back to the sender), the receiver MUST
   monitor the Checkpoint Packet Seqnum to detect an SSRC time-out.  If
   an SSRC time-out is detected, the receiver MUST follow the
   instructions for SSRC time-outs described for the unicast case above.

   Finally, we note that the closed-loop policy is suitable for use in
   RTP/RTCP sessions that use multicast transport.  However, aspects of
   the closed-loop policy do not scale well to sessions with large
   numbers of participants.  The sender state scales linearly with the
   number of receivers, as the sender needs to track the identity and
   M(k) value for each receiver k.  The average recovery journal size is
   not independent of the number of receivers, as the RTCP reporting
   interval backoff slows down the rate of a full update of M(k) values.
   The backoff algorithm may also increase the amount of ancillary state
   used by implementations of the normative sender and receiver
   behaviors defined in Section 4.

C.2.2.3.  The open-loop Sending Policy

   The open-loop policy is suitable for sessions that are not able to
   implement the receiver-to-sender feedback required by the closed-loop
   policy, and that are also not able to use the anchor policy because
   of bandwidth constraints.

   The open-loop policy does not place constraints on how a sender
   chooses the checkpoint packet for each packet in the stream.  In the
   absence of such constraints, a receiver may find that the recovery
   journal in the packet that ends a loss event has a checkpoint history
   that does not cover the entire loss event.  We refer to loss events
   of this type as uncovered loss events.

   To ensure that uncovered loss events do not compromise the recovery
   journal mandate, the open-loop policy assigns specific recovery tasks
   to senders, receivers, and the creators of session descriptions.  The
   underlying premise of the open-loop policy is that the indefinite
   artifacts produced during uncovered loss events fall into two
   classes.

   One class of artifacts is recoverable indefinite artifacts.
   Receivers are able to repair recoverable artifacts that occur during
   an uncovered loss event without intervention from the sender, at the
   potential cost of unpleasant transient artifacts.

   For example, after an uncovered loss event, receivers are able to
   repair indefinite artifacts due to NoteOff (0x8) commands that may
   have occurred during the loss event, by executing NoteOff commands
   for all active NoteOns commands.  This action causes a transient
   artifact (a sudden silent period in the performance), but ensures
   that no stuck notes sound indefinitely.  We refer to MIDI commands
   that are amenable to repair in this fashion as recoverable MIDI
   commands.

   A second class of artifacts is unrecoverable indefinite artifacts.
   If this class of artifact occurs during an uncovered loss event, the
   receiver is not able to repair the stream.

   For example, after an uncovered loss event, receivers are not able to
   repair indefinite artifacts due to Control Change (0xB) Channel
   Volume (controller number 7) commands that have occurred during the
   loss event.  A repair is impossible because the receiver has no way
   of determining the data value of a lost Channel Volume command.  We
   refer to MIDI commands that are fragile in this way as unrecoverable
   MIDI commands.

   The open-loop policy does not specify how to partition the MIDI
   command set into recoverable and unrecoverable commands.  Instead, it
   assumes that the creators of the session descriptions are able to
   come to agreement on a suitable recoverable/unrecoverable MIDI
   command partition for an application.

   Given these definitions, we now state the normative requirements for
   the open-loop policy.

   In the open-loop policy, the creators of the session description MUST
   use the ch_anchor parameter (defined in Appendix C.2.3) to protect
   all unrecoverable MIDI command types from indefinite artifacts, or
   alternatively MUST use the cm_unused parameter (defined in Appendix

   C.1) to exclude the command types from the stream.  These options act
   to shield command types from artifacts during an uncovered loss
   event.

   In the open-loop policy, receivers MUST examine the Checkpoint Packet
   Seqnum field of the recovery journal header after every loss event,
   to check if the loss event is an uncovered loss event.  Section 5
   shows how to perform this check.  If an uncovered loss event has
   occurred, a receiver MUST perform indefinite artifact recovery for
   all MIDI command types that are not shielded by ch_anchor and
   cm_unused parameter assignments in the session description.

   The open-loop policy does not place specific constraints on the
   sender.  However, the open-loop policy works best if the sender
   manages the size of the checkpoint history to ensure that uncovered
   losses occur infrequently, by taking into account the delay and loss
   characteristics of the network.  Also, as each checkpoint packet
   change incurs the risk of an uncovered loss, senders should only move
   the checkpoint if it reduces the size of the journal.

C.2.3.  Recovery Journal Chapter Inclusion Parameters

   The recovery journal chapter definitions (Appendices A-B) specify
   under what conditions a chapter MUST appear in the recovery journal.
   In most cases, the definition states that if a certain command
   appears in the checkpoint history, a certain chapter type MUST appear
   in the recovery journal to protect the command.

   In this section, we describe the chapter inclusion parameters.  These
   parameters modify the conditions under which a chapter appears the
   journal.  These parameters are essential to the use of the open-loop
   policy (Appendix C.2.2.3) and may also be used to simplify
   implementations of the closed-loop (Appendix C.2.2.2) and anchor
   (Appendix C.2.2.1) policies.

   Each parameter represents a type of chapter inclusion semantics.  An
   assignment to a parameter declares which chapters (or chapter
   subsets) obey the inclusion semantics.  We describe the assignment
   syntax for these parameters later in this section.

   A party MUST NOT accept chapter inclusion parameter values that
   violate the recovery journal mandate (Section 4).  All assignments of
   the subsetting parameters (cm_used and cm_unused) MUST precede the
   first assignment of a chapter inclusion parameter in the parameter
   list.

   Below, we normatively define the semantics of the chapter inclusion
   parameters.  For clarity, we define the action of parameters on
   complete chapters.  If a parameter is assigned a subset of a chapter,
   the definition applies only to the chapter subset.

     o  ch_never.  A chapter assigned to the ch_never parameter MUST NOT
        appear in the recovery journal (Appendix A.4.1-2 defines
        exceptions to this rule for Chapter M).  To signal the exclusion
        of a chapter from the journal, an assignment to ch_never MUST be
        made, even if the commands coded by the chapter are assigned to
        cm_unused.  This rule simplifies the handling of commands types
        that may be coded in several chapters.

     o  ch_default.  A chapter assigned to the ch_default parameter MUST
        follow the default semantics for the chapter, as defined in
        Appendices A-B.

     o  ch_anchor.  A chapter assigned to the ch_anchor MUST obey a
        modified version of the default chapter semantics.  In the
        modified semantics, all references to the checkpoint history are
        replaced with references to the session history, and all
        references to the checkpoint packet are replaced with references
        to the first packet sent in the stream.

   Parameter assignments obey the following syntax (see Appendix D for
   ABNF):

     <parameter> = [channel list]<chapter list>[field list]

   The chapter list is mandatory; the channel and field lists are
   optional.  Multiple assignments to parameters have a cumulative
   effect and are applied in the order of parameter appearance in a
   media description.

   To determine the semantics of a list of chapter inclusion parameter
   assignments, we begin by assuming an implicit assignment of all
   channel and system chapters to the ch_default parameter, with the
   default values for the channel list and field list for each chapter
   that are defined below.

   We then interpret the semantics of the actual parameter assignments,
   using the rules below.

   A later assignment of a chapter to the same parameter expands the
   scope of the earlier assignment.  In most cases, a later assignment
   of a chapter to a different parameter cancels (partially or
   completely) the effect of an earlier assignment.

   The chapter list specifies the channel or system chapters for which
   the parameter applies.  The chapter list is a concatenated sequence
   of one or more of the letters corresponding to the chapter types
   (ACDEFMNPQTVWX).  In addition, the list may contain one or more of
   the letters for the sub-chapter types (BGHJKYZ) of System Chapter D.

   The letters in a chapter list MUST be uppercase and MUST appear in
   alphabetical order.  Letters other than (ABCDEFGHJKMNPQTVWXYZ) that
   appear in the chapter list MUST be ignored.

   The channel list specifies the channel journals for which this
   parameter applies; if no channel list is provided, the parameter
   applies to all channel journals.  The channel list takes the form of
   a list of channel numbers (0 through 15) and dash-separated channel
   number ranges (i.e., 0-5, 8-12, etc.).  Dots (i.e., "." characters)
   separate elements in the channel list.

   Several of the systems chapters may be configured to have special
   semantics.  Configuration occurs by specifying a channel list for the
   systems channel, using the coding described below (note that MIDI
   Systems commands do not have a "channel", and thus the original
   purpose of the channel list does not apply to systems chapters).  The
   expression "the digit N" in the text below refers to the inclusion of
   N as a "channel" in the channel list for a systems chapter.

   For the J and K Chapter D sub-chapters (undefined System Common), the
   digit 0 codes that the parameter applies to the LEGAL field of the
   associated command log (Figure B.1.4 of Appendix B.1), the digit 1
   codes that the parameter applies to the VALUE field of the command
   log, and the digit 2 codes that the parameter applies to the COUNT
   field of the command log.

   For the Y and Z Chapter D sub-chapters (undefined System Real-time),
   the digit 0 codes that the parameter applies to the LEGAL field of
   the associated command log (Figure B.1.5 of Appendix B.1) and the
   digit 1 codes that the parameter applies to the COUNT field of the
   command log.

   For Chapter Q (Sequencer State Commands), the digit 0 codes that the
   parameter applies to the default Chapter Q definition, which forbids
   the TIME field.  The digit 1 codes that the parameter applies to the
   optional Chapter Q definition, which supports the TIME field.

   The syntax for field lists follows the syntax for channel lists.  If
   no field list is provided, the parameter applies to all controller or
   note numbers.  For Chapter C, if no field list is provided, the
   controller numbers do not use enhanced Chapter C encoding (Appendix
   A.3.3).

   For Chapter C, the field list may take on values in the range 0 to
   255.  A field value X in the range 0-127 refers to a controller
   number X, and indicates that the controller number does not use
   enhanced Chapter C encoding.  A field value X in the range 128-255
   refers to a controller number "X minus 128" and indicates the
   controller number does use the enhanced Chapter C encoding.

   Assignments made to configure the Chapter C encoding method for a
   controller number MUST be made to the ch_default or ch_anchor
   parameters, as assignments to ch_never act to exclude the number from
   the recovery journal (and thus the indicated encoding method is
   irrelevant).

   A Chapter C field list MUST NOT encode conflicting information about
   the enhanced encoding status of a particular controller number.  For
   example, values 0 and 128 MUST NOT both be coded by a field list.

   For Chapter M, the field list codes the Registered Parameter Numbers
   (RPNs) and Non-Registered Parameter Numbers (NRPNs) for which the
   parameter applies.  The number range 0-16383 specifies RPNs, the
   number range 16384-32767 specifies NRPNs (16384 corresponds to NRPN
   0, 32767 corresponds to NRPN 16383).

   For Chapters N and A, the field list codes the note numbers for which
   the parameter applies.  The note number range specified for Chapter N
   also applies to Chapter E.

   For Chapter E, the digit 0 codes that the parameter applies to
   Chapter E note logs whose V bit is set to 0, and the digit 1 codes
   that the parameter applies to note logs whose V bit is set to 1.

   For Chapter X, the field list codes the number of data octets that
   may appear in a SysEx command that is coded in the chapter.  Thus,
   the field list 0-255 specifies SysEx commands with 255 or fewer data
   octets, the field list 256-4294967295 specifies SysEx commands with
   more than 255 data octets but excludes commands with 255 or fewer
   data octets, and the field list 0 excludes all commands.

   A secondary parameter assignment syntax customizes Chapter X (see
   Appendix D for complete ABNF):

     <parameter> = "__" <h-list> ["_" <h-list>] "__"

   The assignment defines a class of SysEx commands whose Chapter X
   coding obeys the semantics of the assigned parameter.  The command
   class is specified by listing the permitted values of the first N

   data octets that follow the SysEx 0xF0 command octet.  Any SysEx
   command whose first N data octets match the list is a member of the
   class.

   Each <h-list> defines a data octet of the command, as a dot-separated
   (".") list of one or more hexadecimal constants (such as "7F") or
   dash-separated hexadecimal ranges (such as "01-1F").  Underscores
   ("_") separate each <h-list>.  Double-underscores ("__") delineate
   the data octet list.

   Using this syntax, each assignment specifies a single SysEx command
   class.  Session descriptions may use several assignments to the same
   (or different) parameters to specify complex Chapter X behaviors.
   The ordering behavior of multiple assignments follows the guidelines
   for chapter parameter assignments described earlier in this section.

   The example session description below illustrates the use of the
   chapter inclusion parameters:

   v=0
   o=lazzaro 2520644554 2838152170 IN IP6 first.example.net
   s=Example
   t=0 0
   m=audio 5004 RTP/AVP 96
   c=IN IP6 2001:DB80::7F2E:172A:1E24
   a=rtpmap:96 rtp-midi/44100
   a=fmtp:96 j_update=open-loop; cm_unused=ABCFGHJKMQTVWXYZ;
   cm_used=__7E_00-7F_09_01.02.03__;
   cm_used=__7F_00-7F_04_01.02__; cm_used=C7.64;
   ch_never=ABCDEFGHJKMQTVWXYZ; ch_never=4.11-13N;
   ch_anchor=P; ch_anchor=C7.64;
   ch_anchor=__7E_00-7F_09_01.02.03__;
   ch_anchor=__7F_00-7F_04_01.02__

   (The a=fmtp line has been wrapped to fit the page to accommodate
    memo formatting restrictions; it comprises a single line in SDP.)

   The j_update parameter codes that the stream uses the open-loop
   policy.  Most MIDI command-types are assigned to cm_unused and thus
   do not appear in the stream.  As a consequence, the assignments to
   the first ch_never parameter reflect that most chapters are not in
   use.

   Chapter N for several MIDI channels is assigned to ch_never.  Chapter
   N for MIDI channels other than 4, 11, 12, and 13 may appear in the
   recovery journal, using the (default) ch_default semantics.  In
   practice, this assignment pattern would reflect knowledge about a
   resilient rendering method in use for the excluded channels.

   The MIDI Program Change command and several MIDI Control Change
   controller numbers are assigned to ch_anchor.  Note that the ordering
   of the ch_anchor chapter C assignment after the ch_never command acts
   to override the ch_never assignment for the listed controller numbers
   (7 and 64).

   The assignment of command-type X to cm_unused excludes most SysEx
   commands from the stream.  Exceptions are made for General MIDI
   System On/Off commands and for the Master Volume and Balance
   commands, via the use of the secondary assignment syntax.  The
   cm_used assignment codes the exception, and the ch_anchor assignment
   codes how these commands are protected in Chapter X.

C.3.  Configuration Tools: Timestamp Semantics

   The MIDI command section of the payload format consists of a list of
   commands, each with an associated timestamp.  The semantics of
   command timestamps may be set during session configuration, using the
   parameters we describe in this section

   The parameter "tsmode" specifies the timestamp semantics for a
   stream.  The parameter takes on one of three token values: "comex",
   "async", or "buffer".

   The default "comex" value specifies that timestamps code the
   execution time for a command (Appendix C.3.1) and supports the
   accurate transcoding Standard MIDI Files (SMFs, [MIDI]).  The "comex"
   value is also RECOMMENDED for new MIDI user-interface controller
   designs.  The "async" value specifies an asynchronous timestamp
   sampling algorithm for time-of-arrival sources (Appendix C.3.2).  The
   "buffer" value specifies a synchronous timestamp sampling algorithm
   (Appendix C.3.3) for time-of-arrival sources.

   Ancillary parameters MAY follow tsmode in a media description.  We
   define these parameters in Appendices C.3.2-3 below.

C.3.1.  The comex Algorithm

   The default "comex" (COMmand EXecution) tsmode value specifies the
   execution time for the command.  With comex, the difference between
   two timestamps indicates the time delay between the execution of the
   commands.  This difference may be zero, coding simultaneous
   execution.

   The comex interpretation of timestamps works well for transcoding a
   Standard MIDI File (SMF, [MIDI]) into an RTP MIDI stream, as SMFs
   code a timestamp for each MIDI command stored in the file.  To
   transcode an SMF that uses metric time markers, use the SMF tempo map

   (encoded in the SMF as meta-events) to convert metric SMF timestamp
   units into seconds-based RTP timestamp units.

   New MIDI controller designs (piano keyboard, drum pads, etc.) that
   support RTP MIDI and that have direct access to sensor data SHOULD
   use comex interpretation for timestamps, so that simultaneous
   gestural events may be accurately coded by RTP MIDI.

   Comex is a poor choice for transcoding MIDI 1.0 DIN cables [MIDI],
   for a reason that we will now explain.  A MIDI DIN cable is an
   asynchronous serial protocol (320 microseconds per MIDI byte).  MIDI
   commands on a DIN cable are not tagged with timestamps.  Instead,
   MIDI DIN receivers infer command timing from the time of arrival of
   the bytes.  Thus, two two-byte MIDI commands that occur at a source
   simultaneously are encoded on a MIDI 1.0 DIN cable with a 640
   microsecond time offset.  A MIDI DIN receiver is unable to tell if
   this time offset existed in the source performance or is an artifact
   of the serial speed of the cable.  However, the RTP MIDI comex
   interpretation of timestamps declares that a timestamp offset between
   two commands reflects the timing of the source performance.

   This semantic mismatch is the reason that comex is a poor choice for
   transcoding MIDI DIN cables.  Note that the choice of the RTP
   timestamp rate (Section 6.1-2 in the main text) cannot fix this
   inaccuracy issue.  In the sections that follow, we describe two
   alternative timestamp interpretations ("async" and "buffer") that are
   a better match to MIDI 1.0 DIN cable timing, and to other MIDI time-
   of-arrival sources.

   The "octpos", "linerate", and "mperiod" ancillary parameters (defined
   below) SHOULD NOT be used with comex.

C.3.2.  The async Algorithm

   The "async" tsmode value specifies the asynchronous sampling of a
   MIDI time-of-arrival source.  In asynchronous sampling, the moment an
   octet is received from a source, it is labelled with a wall-clock
   time value.  The time value has RTP timestamp units.

   The "octpos" ancillary parameter defines how RTP command timestamps
   are derived from octet time values.  If octpos has the token value
   "first", a timestamp codes the time value of the first octet of the
   command.  If octpos has the token value "last", a timestamp codes the
   time value of the last octet of the command.  If the octpos parameter
   does not appear in the media description, the sender does not know
   which octet of the command the timestamp references (for example, the
   sender may be relying on an operating system service that does not
   specify this information).

   The octpos semantics refer to the first or last octet of a command as
   it appears on a time-of-arrival MIDI source, not as it appears in an
   RTP MIDI packet.  This distinction is significant because the RTP
   coding may contain octets that are not present in the source.  For
   example, the status octet of the first MIDI command in a packet may
   have been added to the MIDI stream during transcoding, to comply with
   the RTP MIDI running status requirements (Section 3.2).

   The "linerate" ancillary parameter defines the timespan of one MIDI
   octet on the transmission medium of the MIDI source to be sampled
   (such as a MIDI 1.0 DIN cable).  The parameter has units of
   nanoseconds, and takes on integral values.  For MIDI 1.0 DIN cables,
   the correct linerate value is 320000 (this value is also the default
   value for the parameter).

   We now show a session description example for the async algorithm.
   Consider a sender that is transcoding a MIDI 1.0 DIN cable source
   into RTP.  The sender runs on a computing platform that assigns time
   values to every incoming octet of the source, and the sender uses the
   time values to label the first octet of each command in the RTP
   packet.  This session description describes the transcoding:

   v=0
   o=lazzaro 2520644554 2838152170 IN IP4 first.example.net
   s=Example
   t=0 0
   m=audio 5004 RTP/AVP 96
   c=IN IP4 192.0.2.94
   a=rtpmap:96 rtp-midi/44100
   a=sendonly
   a=fmtp:96 tsmode=async; linerate=320000; octpos=first

C.3.3.  The buffer Algorithm

   The "buffer" tsmode value specifies the synchronous sampling of a
   MIDI time-of-arrival source.

   In synchronous sampling, octets received from a source are placed in
   a holding buffer upon arrival.  At periodic intervals, the RTP sender
   examines the buffer.  The sender removes complete commands from the
   buffer and codes those commands in an RTP packet.  The command
   timestamp codes the moment of buffer examination, expressed in RTP
   timestamp units.  Note that several commands may have the same
   timestamp value.

   The "mperiod" ancillary parameter defines the nominal periodic
   sampling interval.  The parameter takes on positive integral values
   and has RTP timestamp units.

   The "octpos" ancillary parameter, defined in Appendix C.3.1 for
   asynchronous sampling, plays a different role in synchronous
   sampling.  In synchronous sampling, the parameter specifies the
   timestamp semantics of a command whose octets span several sampling
   periods.

   If octpos has the token value "first", the timestamp reflects the
   arrival period of the first octet of the command.  If octpos has the
   token value "last", the timestamp reflects the arrival period of the
   last octet of the command.  The octpos semantics refer to the first
   or last octet of the command as it appears on a time-of-arrival
   source, not as it appears in the RTP packet.

   If the octpos parameter does not appear in the media description, the
   timestamp MAY reflect the arrival period of any octet of the command;
   senders use this option to signal a lack of knowledge about the
   timing details of the buffering process at sub-command granularity.

   We now show a session description example for the buffer algorithm.
   Consider a sender that is transcoding a MIDI 1.0 DIN cable source
   into RTP.  The sender runs on a computing platform that places source
   data into a buffer upon receipt.  The sender polls the buffer 1000
   times a second, extracts all complete commands from the buffer, and
   places the commands in an RTP packet.  This session description
   describes the transcoding:

   v=0
   o=lazzaro 2520644554 2838152170 IN IP6 first.example.net
   s=Example
   t=0 0
   m=audio 5004 RTP/AVP 96
   c=IN IP6 2001:DB80::7F2E:172A:1E24
   a=rtpmap:96 rtp-midi/44100
   a=sendonly
   a=fmtp:96 tsmode=buffer; linerate=320000; octpos=last; mperiod=44

   The mperiod value of 44 is derived by dividing the clock rate
   specified by the rtpmap attribute (44100 Hz) by the 1000 Hz buffer
   sampling rate and rounding to the nearest integer.  Command
   timestamps might not increment by exact multiples of 44, as the
   actual sampling period might not precisely match the nominal mperiod
   value.

C.4.  Configuration Tools: Packet Timing Tools

   In this appendix, we describe session configuration tools for
   customizing the temporal behavior of MIDI stream packets.

C.4.1.  Packet Duration Tools

   Senders control the granularity of a stream by setting the temporal
   duration ("media time") of the packets in the stream.  Short media
   times (20 ms or less) often imply an interactive session.  Longer
   media times (100 ms or more) usually indicate a content streaming
   session.  The RTP AVP profile [RFC3551] recommends audio packet media
   times in a range from 0 to 200 ms.

   By default, an RTP receiver dynamically senses the media time of
   packets in a stream and chooses the length of its playout buffer to
   match the stream.  A receiver typically sizes its playout buffer to
   fit several audio packets and adjusts the buffer length to reflect
   the network jitter and the sender timing fidelity.

   Alternatively, the packet media time may be statically set during
   session configuration.  Session descriptions MAY use the RTP MIDI
   parameter "rtp_ptime" to set the recommended media time for a packet.
   Session descriptions MAY also use the RTP MIDI parameter
   "rtp_maxptime" to set the maximum media time for a packet permitted
   in a stream.  Both parameters MAY be used together to configure a
   stream.

   The values assigned to the rtp_ptime and rtp_maxptime parameters have
   the units of the RTP timestamp for the stream, as set by the rtpmap
   attribute (see Section 6.1).  Thus, if rtpmap sets the clock rate of
   a stream to 44100 Hz, a maximum packet media time of 10 ms is coded
   by setting rtp_maxptime=441.  As stated in the Appendix C preamble,
   the senders and receivers of a stream MUST agree on common values for
   rtp_ptime and rtp_maxptime if the parameters appear in the media
   description for the stream.

   0 ms is a reasonable media time value for MIDI packets and is often
   used in low-latency interactive applications.  In a packet with a 0
   ms media time, all commands execute at the instant they are coded by
   the packet timestamp.  The session description below configures all
   packets in the stream to have 0 ms media time:

   v=0
   o=lazzaro 2520644554 2838152170 IN IP4 first.example.net
   s=Example
   t=0 0
   m=audio 5004 RTP/AVP 96
   c=IN IP4 192.0.2.94
   a=rtpmap:96 rtp-midi/44100
   a=fmtp:96 rtp_ptime=0; rtp_maxptime=0

   The session attributes ptime and maxptime [RFC4566] MUST NOT be used
   to configure an RTP MIDI stream.  Sessions MUST use rtp_ptime in lieu
   of ptime and MUST use rtp_maxptime in lieu of maxptime.  RTP MIDI
   defines its own parameters for media time configuration because 0 ms
   values for ptime and maxptime are forbidden by [RFC3264] but are
   essential for certain applications of RTP MIDI.

   See the Appendix C.7 examples for additional discussion about using
   rtp_ptime and rtp_maxptime for session configuration.

C.4.2.  The guardtime Parameter

   RTP permits a sender to stop sending audio packets for an arbitrary
   period of time during a session.  When sending resumes, the RTP
   sequence number series continues unbroken, and the RTP timestamp
   value reflects the media time silence gap.

   This RTP feature has its roots in telephony, but it is also well
   matched to interactive MIDI sessions, as players may fall silent for
   several seconds during (or between) songs.

   Certain MIDI applications benefit from a slight enhancement to this
   RTP feature.  In interactive applications, receivers may use on-line
   network models to guide heuristics for handling lost and late RTP
   packets.  These models may work poorly if a sender ceases packet
   transmission for long periods of time.

   Session descriptions may use the parameter "guardtime" to set a
   minimum sending rate for a media session.  The value assigned to
   guardtime codes the maximum separation time between two sequential
   packets, as expressed in RTP timestamp units.

   Typical guardtime values are 500-2000 ms.  This value range is not a
   normative bound, and parties SHOULD be prepared to process values
   outside this range.

   The congestion control requirements for sender implementations
   (described in Section 8 and [RFC3550]) take precedence over the
   guardtime parameter.  Thus, if the guardtime parameter requests a
   minimum sending rate, but sending at this rate would violate the
   congestion control requirements, senders MUST ignore the guardtime
   parameter value.  In this case, senders SHOULD use the lowest minimum
   sending rate that satisfies the congestion control requirements.

   Below, we show a session description that uses the guardtime
   parameter.

   v=0
   o=lazzaro 2520644554 2838152170 IN IP6 first.example.net
   s=Example
   t=0 0
   m=audio 5004 RTP/AVP 96
   c=IN IP6 2001:DB80::7F2E:172A:1E24
   a=rtpmap:96 rtp-midi/44100
   a=fmtp:96 guardtime=44100; rtp_ptime=0; rtp_maxptime=0