Smart Filters
Introduction
The smart filters policy allows automatically linking filters together, in a chain, and tied to a specific target node. This is useful when we want to apply a specific processing chain to a specific device, for example. When a stream is about to be linked to a target node that is associated with a smart filter chain, the policy will automatically link the stream with the first filter in the chain, and the last filter in the chain with the target node. This is done transparently to the client, allowing users to define a specific processing chain for a specific device without having to create setups with virtual sinks (or sources) that must be explicitly targeted by the clients.
Filters, in general, are nodes that are placed in the middle of the graph and are used to modify the data that passes through them. For example, the echo-cancel, the filter-chain, or the loopback nodes are filters. Filters can be implemented either as a single node or as a pair of nodes with opposite directions. For example, the null-audio-sink node can be configured to be a single-node filter. On the other hand, the filter-chain is a pair of nodes with opposite directions, where one node captures the audio from the graph and the other node sends the modified audio back to the graph.
For the purpose of the smart filters policy, WirePlumber will only consider
pairs of nodes as filters, not single-node ones. More specifically, a pair of
nodes will be considered to be a filter by WirePlumber if they have the
node.link-group property set to a common value. This property is always set
on pairs of nodes that are internally linked together and is a good indicator
that the nodes are implementing a filter.
That pair of nodes must always consist of a stream node and a main node. The main node acts as a virtual device, where the data is sent or captured to/from, and the stream node acts as a regular stream, where the data is sent or received to/from the next node in the graph. This is designated by their media class, as shown in the table below:
Input filter (virtual sink) |
Output filter (virtual source) |
|
|---|---|---|
Main node |
|
|
Stream node |
|
|
For instance, if a smart filter is used between an application playback stream and the default audio sink, the graph would look like this:
The same logic is applied if the smart filter is used between an application capture stream and the default audio source, it is just all in the opposite direction. This is how the graph would look like in this case:
When multiple filters have the same direction, they can also be chained together so that the output of one filter is sent to the input of the next filter. The next section describes how these chains can be described with properties so that they are automatically linked by WirePlumber in any way we want.
Filter properties
When a filter node is created, WirePlumber will check for the presence of the following optional node properties on the main node:
filter.smart
Boolean indicating whether smart policy will be used for these filter nodes or not. This is disabled by default, therefore filter nodes will be treated as regular nodes, without applying any kind of extra logic. On the other hand, if this property is set to
true, automatic (smart) filter policy will be used when linking them. The properties below will then also apply, providing further instructions.filter.smart.name
The unique name of the filter. WirePlumber will use the value of the
node.link-groupproperty as the filter name if this property is not set.filter.smart.disabled
Boolean indicating whether the filter should be disabled or not. A disabled filter will never be used under any circumstances. If the property is not set, WirePlumber will consider the filter as enabled (i.e. disabled = false).
filter.smart.on-request
When true, only streams requesting this filter by
filter.applyorfilter.wantare automatically connected to it. The value must match the filter'sfilter.smart.name. The default is false; existing smart filters continue to apply to all streams targeting their device.filter.applytakes precedence, including an empty value. A matchingfilter.suppressdisables either kind of request. Anecho-cancelpreference infilter.wantis ignored for targets withphoneindevice.intended-roles; an explicitfilter.applyis not.Request-only filters are separate entry points, placed before the permanent chain irrespective of their
before/afterposition. A permanent filter must never feed into a request-only filter, because its audio is already shared with streams that did not request processing. Request-only filters do not chain into each other. Permanent filters retain their relative order.filter.smart.targetable
Boolean indicating whether the filter can be directly linked with clients that have it defined as a target (e.g.
pw-play --target <filter-name>) or not. This can be useful when a client wants to be linked with a filter that is in the middle of the chain in order to bypass the filters that are placed before the selected one. If the property is not set, WirePlumber will consider the filter not targetable by default, meaning filters will never by bypassed by clients, and clients will always be linked with the first filter in the chain.filter.smart.target
A JSON object that defines the matching properties of the filter's target node. A filter target can never be another filter node (WirePlumber will ignore it), it must be a device or virtual sink (or source, depending on the direction of the filter). If this property is not set, WirePlumber will use the default sink/source as the target.
filter.smart.before
A JSON array containing the names of the filters that are supposed to be chained after this filter (i.e. this filter here should be chained before those). If not set, WirePlumber will link the filters by order of creation.
filter.smart.after
A JSON array containing the names of the filters that are supposed to be chained before this filter (i.e. this filter here should be chained after those). If not set, WirePlumber will link the filters by order of creation.
Note
These properties must be set on the filter's main node, not the stream node.
As an example, we will describe here how to create 2 loopback filters in PipeWire's configuration, with names loopback-1 and loopback-2, that will be linked with the default audio device, and use loopback-2 filter as the last filter in the chain.
The PipeWire configuration files for the 2 filters should be like this:
~/.config/pipewire/pipewire.conf.d/loopback-1.conf:
context.modules = [ { name = libpipewire-module-loopback args = { node.name = loopback-1-sink node.description = "Loopback 1 Sink" capture.props = { audio.position = [ FL FR ] media.class = Audio/Sink filter.smart = true filter.smart.name = loopback-1 filter.smart.before = [ loopback-2 ] } playback.props = { audio.position = [ FL FR ] node.passive = true stream.dont-remix = true } } } ]~/.config/pipewire/pipewire.conf.d/loopback-2.conf:
context.modules = [ { name = libpipewire-module-loopback args = { node.name = loopback-2-sink node.description = "Loopback 2 Sink" capture.props = { audio.position = [ FL FR ] media.class = Audio/Sink filter.smart = true filter.smart.name = loopback-2 } playback.props = { audio.position = [ FL FR ] node.passive = true stream.dont-remix = true } } } ]
After restarting PipeWire to apply the configuration changes, playing a test wave audio file with paplay to the default device should result in the following graph:
Now, if we remove the filter.smart.before = [ loopback-2 ] property from the
loopback-1 filter, and add a filter.smart.before = [ loopback-1 ] property
in the loopback-2 filter configuration file, WirePlumber should link the
loopback-1 filter as the last filter in the chain, like this:
In addition, the filters can have different targets. For example, we can define the filters like this:
~/.config/pipewire/pipewire.conf.d/loopback-1.conf:
context.modules = [ { name = libpipewire-module-loopback args = { node.name = loopback-1-sink node.description = "Loopback 1 Sink" capture.props = { audio.position = [ FL FR ] media.class = Audio/Sink filter.smart = true filter.smart.name = loopback-1 filter.smart.after = [ loopback-2 ] filter.smart.target = { node.name = "not-default-audio-device" } } playback.props = { audio.position = [ FL FR ] node.passive = true stream.dont-remix = true } } } ]~/.config/pipewire/pipewire.conf.d/loopback-2.conf:
context.modules = [ { name = libpipewire-module-loopback args = { node.name = loopback-2-sink node.description = "Loopback 2 Sink" capture.props = { audio.position = [ FL FR ] media.class = Audio/Sink filter.smart = true filter.smart.name = loopback-2 } playback.props = { audio.position = [ FL FR ] node.passive = true stream.dont-remix = true } } } ]
In this case, playing a test wave audio file with paplay to the
not-default-audio-device device should result in the following graph:
In this configuration, the loopback-1 filter will only be linked if the application stream is targeting the device node called "not-default-audio-device".
Filters metadata
Similar to the default metadata, it is also possible to override the filter properties using the "filters" metadata object. This allow users to change the filters policy at runtime.
For example, assuming the id of the loopback-1 main node is 40, we can
disable the filter by setting its filter.smart.disabled metadata key to
true using the pw-metadata tool like this:
$ pw-metadata -n filters 40 "filter.smart.disabled" true Spa:String:JSON
We can also change the target of a filter at runtime:
$ pw-metadata -n filters 40 "filter.smart.target" "{ node.name = new-target-node-name }" Spa:String:JSON
Every time a key in the filters metadata changes, all filters are unlinked and re-linked properly, following the new policy.
On-demand echo cancellation
The optional hooks.node.echo-cancel component creates an echo-cancel module
when a matching audio stream is being linked, and unloads it after the last
requester disappears or stops requesting the filter. It is not enabled in the
default profile. To try it with the standard linking policy, add a configuration
fragment in wireplumber.conf.d:
wireplumber.profiles = {
main = { hooks.node.echo-cancel = required }
}
For a fixed microphone/speaker pair, as on an embedded phone, configure the
actual node.name values from wpctl status -n:
node.echo-cancel = {
source = "my_alsa_input"
sink = "my_alsa_output"
}
Only streams whose selected target matches the configured device on their side request that module. Without explicit devices, the filter follows default-device selection, including streams explicitly naming the current default. Explicit unrelated targets, audio groups and role-policy targets are left to their existing policies. This component manages one pair, not a separate AEC instance for every application or arbitrary target pair.
The filter main nodes are named wp.echo-cancel.source and
wp.echo-cancel.sink. Their capture/playback streams are passive, do not save
route preferences, and do not fall back when a configured target is missing.
Other applications keep their normal paths through permanent filters. No default
device metadata or application properties are rewritten.
The existing PipeWire WebRTC AEC SPA plugin must be installed. If loading fails, a warning is logged once per nonempty group of requests instead of retrying on every graph rescan. After all requesters have gone, a new request may try again. No PulseAudio module parameters are interpolated into module configuration.
Testing
The request, chain-isolation and lifecycle regression tests run with the normal
meson test -C build suite. The integration exercise at
tests/integration/echo-cancel-on-request.py starts private PipeWire,
pipewire-pulse and WirePlumber processes with synthetic source/sink nodes. It
requires a common installation prefix with the WebRTC plugin and PulseAudio
pacat/parec tools:
WP_TEST_PREFIX=/path/to/prefix dbus-run-session -- \
python3 tests/integration/echo-cancel-on-request.py
It checks real PulseAudio requests, graph links, sharing, suppression and module teardown. Synthetic routing success does not demonstrate acoustic echo reduction on a Librem 5. Device-specific audio quality and interaction with the phone's existing policy must be validated separately before deployment.