Voice data

Mumble audio channel is used to transmit the actual audio packets over the network. Unlike the TCP control channel, the audio channel uses a custom encoding for the audio packets. The audio channel is transport independent and features such as encryption are implemented by the transport layer. Integers above 8-bits are encoded using the Variable length integer encoding.

Packet format

The mumble audio channel packets are variable length packets that begin with an 8-bit header field which describes the packet type and target. The most significant 3 bits define the packet type while the remaining 5 bits define the target. The header is followed by the packet payload. The maximum size for the whole audio data packet is 1020 bytes. This allows applications to use 1024 byte buffers for receiving UDP datagrams with the 4-byte encryption header overhead.

Audio packet structure
Audio packet structure
7 6 5 4 3 2 1 0
type target
The audio packet type. The packets transmitted over the audio channel are either ping packets used to diagnose the transport layer connectivity or audio packets encoded with different codecs. Different types are listed in Audio packet types table.
Audio packet types
Type Bitfield Description
0 000xxxxx CELT Alpha encoded voice data
1 001xxxxx Ping packet
2 010xxxxx Speex encoded voice data
3 011xxxxx CELT Beta encoded voice data
4 100xxxxx OPUS encoded voice data
5-7   Unused

The target portion defines the recipient for the audio data. The two constant targets are Normal talking (0) and Server Loopback (31). The range 1-30 is reserved for whisper targets. These targets are specified separately in the control channel using the VoiceTarget packets. The targets are listed in Audio targets table.

When a client registers a VoiceTarget on the server, it gives the target an ID. This voice target ID can be used as a target in the voice packets to send audio to specific users or channels. When receiving whisper-audio the server uses target 1 to specify the audio results from a whisper to a channel and target 2 to specify that the audio results from a direct whisper to the user.

Audio targets
Target Description
0 Normal talking

Whisper target

  • VoiceTarget ID when sending whisper from client.
  • 1 when receiving whisper to channel.
  • 2 when receiving direct whisper to user.
31 Server loopback

Ping packet

Audio channel ping packets are used as part of the connectivity checks on the audio transport layer. These packets contain only varint encoded timestamp as data. See UDP connectivity checks section below for the logic involved in the connectivity checks.

Audio transport ping packet
Field Type Description
Header byte 00100000b (0x20)
Data varint Timestamp
Common audio packet header. For ping packets this should have the value of 0x20.
Timestamp. The packet should be echoed back so the timestamp format can be decided by the original sender - the only limitation is that it must fit in a 64-bit integer for the varint encoding.

Encoded audio data packet

Encoded audio packets contain the actual user audio data for the voice communication. Incoming audio data packets contain the common header byte followed by varint encoded session ID of the source user and varint encoded sequence number of the packet. Outgoing audio data packets contain only the header byte and the sequence number of the packet. The server matches these to the correct session using the transport layer information.

The remainder of the packet is made up of multiple encoded audio segments and optional positional audio information. The audio segment format depends on the codec of the whole audio packets. The audio segments contain codec implementation specific information on where the audio segments end so the possible positional audio data can be read from the end.

Incoming encoded audio packet
Field Type Description
Header byte Codec type/Audio target
Session ID varint Session ID of the source user.
Sequence Number varint Sequence number of the first audio data segment.
Payload byte[] Audio payload
Position Info float[3] Positional audio information
Outgoing encoded audio packet
Field Type Description
Header byte Codec type/Audio target
Sequence Number varint Sequence number of the first audio data segment.
Payload byte[] Audio payload
Position Info float[3] Positional audio information
The common audio packet header
Session ID
Session ID of the user to whom the audio packet belongs.
Sequence Number

Audio data sequence number. The sequence number is used to maintain the packet order when the audio data is transported over unreliable transports such as UDP.

The sequence number might increase by more than one between subsequent audio packets in case the audio packets contain multiple audio segments. This allows the packet loss concealment algorithms to figure out how many audio frames were lost between two received packets.

Audio payload. Format depends on the audio codec defined in the Header. The payload must be self-delimiting to determine whether the position info exists at the end of the packet.
Position Info
The XYZ coordinates of the audio source. In addition to sending the position information, the user must be using a positional plugin defined in the UserState message. The plugins might define different contexts which prevent voice communication between users in other contexts.

Speex and CELT audio frames

Encoded Speex and CELT audio is transported as individual encoded frames. Each frame is prefixed with a single byte length and terminator header.

CELT encoded audio data
Field Type Description
Header byte length/continuation header
Data byte[] Encoded voice frame

The length of the Data field. The most significant bit (0x80) acts as the continuation bit and is set for all but the last frame in the payload. The remaining 7 bits of the header contain the actual length of the Data frame.

Note the length may be zero, which is used to signal the end of a voice transmission. In this case the audio data is a single zero-byte which can be interpreted normally as length of 0 with no continuation bit set.

Single encoded audio frame. The encoding depends on the codec type header of the whole audio packet

Opus audio frames

Encoded Opus audio is transported as a single Opus audio frame. The frame is prefixed with a variable byte header.

Opus encoded audio data
Field Type Description
Header varint length/terminator header
Data byte[] Encoded voice frame

The length of the Data field. 16-bit variable length integer encoded length and terminator bit value. The varint encoding is the same as with 64-bit values, but only 16-bit unencoded values are allowed.

The maximum voice frame size is 8191 (0x1FFF) bytes requiring the 13 least significant bits of the header. The 14th bit (mask: 0x2000) is the terminator bit which signals whether the packet is the last one in the voice transmission.

Note: In CELT the “continuation bit” in the header defines whether there are more audio frames in the current packet. Opus always contains only one frame in the packet. In CELT the voice transmission end is signaled with a zero-byte CELT packet while in Opus we have a dedicated termination bit in the header.

The encoded Opus data.


Mumble supports three distinct codecs; Older Mumble versions use Speex for low bitrate audio and CELT for higher quality audio while new Mumble versions prefer Opus for all audio. When multiple clients with different capabilities communicate together the server is responsible for resolving the codec to use. The clients should respect the server resolution if they are capable.

If the server resolves a codec a client doesn’t support, that client is free to use any codec it prefers. Usually this means the client will not be able to decode incoming audio, but it can still send encoded audio out.

The CELT bitstream was never frozen which makes most CELT versions incompatible with each other. The two CELT bitstreams supported by Mumble are: CELT 0.7.0 (CELT Alpha) and CELT 0.11.0 (CELT Beta). While CELT 0.7.0 should technically be supported by most Mumble implementations, some servers might be configured to force Opus codec for the users. Mumble has had Opus support since 1.2.4 (June 2013) so it should be safe to assume most clients in use support this now.


Normal talking can be heard by the users of the current channel and all linked channels as long as the speaker has Talk permission on these channels. If the speaker wishes to broadcast the voice to specific users or channels, he may use whispering. This is achieved by registering a voice target using the VoiceTarget message and specifying the target ID as the target in the first byte of the UDP packet.

UDP connectivity checks

Since UDP is a connectionless protocol, it is heavily affected by network topology such as NAT configuration. It should not be used for audio transmission before the connectivity has been determined.

The client starts the connectivity checks by sending a Ping packet to the server. When the server receives this packet it will respond by echoing it back to the address it received it from. Once the client receives the response from the server it can start using the UDP transport for audio data. When the server receives incoming audio data over the UDP transport it can switch the outgoing audio over to UDP transport as well.

If the client stops receiving replies to the UDP pings at some point, it should start tunneling the voice communication through the TCP tunnel as described in the Tunneling audio over TCP below. When the server receives a tunneled packet over the TCP connection it must also stop using the UDP for communication. The client should still continue sending audio ping packets over the UDP transport in case the UDP connection is restored and the communication can be switched back to it.

Tunneling audio over TCP

If the UDP channel isn’t available the voice packets can be transmitted through the TCP transport used for the control channel. These messages use the normal TCP prefixing, as shown in figure Mumble packet: 16-bit message type followed by 32-bit message length. However unlike other TCP messages, the audio packets are not encoded as protocol buffer messages but instead the raw audio packet described in Packet format should be written to the TCP socket verbatim.

When the packets are received it is safe to parse the type and length fields normally. If the type matches that of the audio tunnel the rest of the message should be processed as an UDP packet without attempting a protocol buffer decoding.

Implementation note

When implementing the protocol it is easier to ignore the UDP transfer layer at first and just tunnel the UDP data through the TCP tunnel. The TCP layer must be implemented for authentication in any case. Making sure that the voice transmission works before implementing the UDP protocol simplifies debugging greatly.


All the packets are encrypted once during transfer. The actual encryption depends on the used transport layer. If the packets are tunneled through TCP they are encrypted using the TLS that encrypts the whole control channel connection and if they are sent directly using UDP they must be encrypted using the OCB-AES128 encryption.

Variable length integer encoding

The variable length integer encoding (varint) is used to encode long, 64-bit, integers so that short values do not need the full 8 bytes to be transferred. The basic idea behind the encoding is prefixing the value with a length prefix and then removing the leading zeroes from the value. The positive numbers are always right justified. That is to say that the least significant bit in the encoded presentation matches the least significant bit in the decoded presentation. The varint prefixes table contains the definitions of the different length prefixes. The encoded x bits are part of the decoded number while the _ signifies a unused bit. Encoding should be done by searching the first decoded description that fits the number that should be decoded, truncating it to the required bytes and combining it with the defined encoding prefix.

See the quint64 shift operators in https://github.com/mumble-voip/mumble/blob/master/src/PacketDataStream.h for a reference implementation.

Varint prefixes
Encoded Decoded
0xxxxxxx 7-bit positive number
10xxxxxx + 1 byte 14-bit positive number
110xxxxx + 2 bytes 21-bit positive number
1110xxxx + 3 bytes 28-bit positive number
111100__ + int (32-bit) 32-bit positive number
111101__ + long (64-bit) 64-bit number
111110__ + varint Negative recursive varint
111111xx Byte-inverted negative two bit number (~xx)