Audio Onset Detection based on Fusion

h1. Wiki

h2. Vamp environment for feature extraction

_Vamp_ is an audio processing plugin system for plugins that extract descriptive information from audio data — typically referred to as audio analysis plugins or audio feature extraction plugins. The _vamp plugins_ are distributed in shared library files with extension _.dll_, _.so_, or _.dylib_ depending on the platform.  

Several common audio file formats are supported, including MP3, Ogg, and a number of PCM formats such as WAV and AIFF. AAC is supported on OS/X only, and only if not DRM protected. WMA is not supported.

See http://vamp-plugins.org/ for more introduction about Vamp environment and the plugins.

h2. Vampy wrapper plugin (for running plugins written in Python)

Vamp plugins are typically written in C++. To use Vamp plugins written in Python you need to install the "Vampy wrapper plugin":http://www.vamp-plugins.org/vampy.html. 

Then, any python script following some simple conventions will be presented to a vamp host as it was an ordinary Vamp plugin.

Download Vampy: http://www.vamp-plugins.org/vampy.html

Tutorial: http://isophonics.net/content/getting-started-vampy

Put the plugins in your Vamp plugin folder or any other designated directory. You may set an environment variable to tell

your system where to look for Vamp plugins. On Linux/Unix or OS/X: $ export VAMPY_PATH="/your/vamp/plugin/directory"

Optionally, you may need to tell your system where to find Python for Vampy to work correctly:

To do this, set the VAMPY_PYLIB environment variable to the location of the python shared library (.dylib / .so).

For more information see "this readme file in the Vampy package":https://code.soundsoftware.ac.uk/projects/vampy/repository/entry/README

h2. Host applications

A vamp plugin cannot be used on its own, but only with a conforming _host_ application which loads the plugin from its shared library and calls functions within the plugin code to configure it, supply it with data, and run it.

* "Sonic Visualiser":http://www.sonicvisualiser.org/ -- An interactive audio analysis and visualisation tool using Vamp plugins

* "Sonic Annotator":https://code.soundsoftware.ac.uk/projects/sonic-annotator -- A command line tool facilitating facilitating feature extraction, parameter configuration and output manipulation for a large bunch of audio files.

* "Audacity":http://audacityteam.org/ -- An audio editor which enables recording, labelling and editing the audio.

__Note for developers:__ If you are developing new plugins in Python it is customary to use a command line host such as Sonic Annotator or the "Vamp Simple Host":http://vamp-plugins.org/develop.html available in the Vamp Plugin SDK. This will allow you to see error messages printed to the terminal, which would otherwise be invisible. Vampy plugins can be used in any GUI host like Sonic Visualiser, but errors reported by plugin scripts will be more difficult to see.

h2. Plugin installation

The vamp plugins need to be put in a directory where the host can locate, i.e., the _VAMP_PATH_. See http://vamp-plugins.org/download.html#install for installation instructions. 

Alternatively, you can set your own VAMP_PATH environment variable to list the locations a host should look in for Vamp plugins. For example, the following command will set @/homes/mitian/hg/onset-fusion@ as an alternative VAMP_PATH. Plugins placed in this directory will also be locatable for the hosts besides those under the system specified vamp directories.

<pre>

mitian$ export VAMP_PATH=$VAMP_PATH:/homes/mitian/hg/onset-fusion

</pre>

Once the plugins are installed in the Vamp path, they will show themselves under the _Transform_ tab in Sonic Visualiser, or the _Analyze_ tab in Audacity, or in the terminal when you list them using Sonic Annotator (@ sonic-annotator -l@). 

h2. Parameters and outputs of this plugin set

h3. Parameters

<pre>

    **Adjustable parameters**

	Fusion Type                         Fusion type [0: early fusion, 1: linear fusion, 2: decision fusion]

	Detection Sensitivity               Sensitivity level for peak detection in the onset likelihood function.

                                            The higher the sensitivity, the more onsets will (rightly or wrongly) be detected.

	Delta Threshold                     Constant offset used for adaptive theresholding using the median of the detection function.

	Backtracking Threshold              Backtracking threshold used determine the stopping condition for backtracking the detected onset location to an earlier 'perceived' location.

	Tolerance window                    Tolerance window to combine events from different detection functions for decision fusion.

	Linear combination coefficient      Combination coefficient for linear fusion.

	Median filter window                Median filter window size.

	Cutoff frequency                    Cutoff frequency for the low pass filter.

	**bool parameters**                 [0: off, 1: on]

	Polynomial fitting                  Use polynomial fitting to evaluate detection function peaks.

	Median filtering                    Use median filter based adaptive thresholding.

	Low-pass filtering                  Use low pass filtering to smooth onset detection functions.

	Adaptive whitening                  Use adaptive whitening. It evens out the temporal and frequency variation in the signal.

</pre>

h3. Outputs

# *Note Onsets* – The detected note onset times, returned as a single feature with timestamp.

# *Smoothed Detection Function* – The smoothed onset detection function.

# *Note Onsets* – The detected note onset times, returned as a single feature with timestamp.

NOTE: The last two outputs are not available in fusion based onset detectors. Please check individual plugins for specific output types.

h2. Example usages of the onset detectors

To use these onset detector plugins in Sonic Visualiser and Audacity is quite straightforward. Load your audio file first, then select a feature extractor/plugin from the transform list. You will also find a set of adjustable parameters in the pop-out box associated with each transform. The extracted features, whether it is onset locations or detection functions, will be visualised in the pane on top of the waveform, synchronised with time.

For comprehensive user manuals, see

http://www.sonicvisualiser.org/doc/reference/2.1/en/index.html

http://wiki.audacityteam.org/wiki/Category:Tutorial

Rather than for visualisation purpose, Sonic Annotator lets you to extract and store audio features efficiently. Here we give a quick tutorial of how to detect onset for a given audio file and save the results to local folder using Sonic Annotator.

<pre>

mitian$ sonic-annotator -l | grep vampy-onsets

vamp:vampy:vampy-onsets-bersf:vampy-onsets-bersf-onsets

vamp:vampy:vampy-onsets-cdsf:vampy-onsets-cdsf-onsets

vamp:vampy:vampy-onsets-ber:vampy-onsets-ber-df

vamp:vampy:vampy-onsets-ber:vampy-onsets-ber-onsets

vamp:vampy:vampy-onsets-ber:vampy-onsets-ber-sdf

vamp:vampy:vampy-onsets-cd:vampy-onsets-cd-df

vamp:vampy:vampy-onsets-cd:vampy-onsets-cd-onsets

vamp:vampy:vampy-onsets-cd:vampy-onsets-cd-sdf

vamp:vampy:vampy-onsets-hfc:vampy-onsets-hfc-df

vamp:vampy:vampy-onsets-hfc:vampy-onsets-hfc-onsets

vamp:vampy:vampy-onsets-hfc:vampy-onsets-hfc-sdf

vamp:vampy:vampy-onsets-pd:vampy-onsets-pd-df

vamp:vampy:vampy-onsets-pd:vampy-onsets-pd-onsets

vamp:vampy:vampy-onsets-pd:vampy-onsets-pd-sdf

vamp:vampy:vampy-onsets-sd:vampy-onsets-sd-df

vamp:vampy:vampy-onsets-sd:vampy-onsets-sd-onsets

vamp:vampy:vampy-onsets-sd:vampy-onsets-sd-sdf

vamp:vampy:vampy-onsets-sf:vampy-onsets-sf-df

vamp:vampy:vampy-onsets-sf:vampy-onsets-sf-onsets

vamp:vampy:vampy-onsets-sf:vampy-onsets-sf-sdf

mitian$

mitian$ sonic-annotator -d vamp:vampy:vampy-onsets-cdsf:vampy-onsets-cdsf-onsets -w csv test.wav --csv-basedir /homes/mitian/hg/workspace/onsetdetection

Extracting and writing features... Done

Vampy::~PyExtensionManager: Extension module cleaned.

mitian$

mitian$

</pre>

Each line of the output csv file consists of the timestamp (and duration where applicable) for a single feature, the value of that feature and the feature's label (where applicable), separated by comma.

Sonic Anntator also allows for batch processing in one go. 

<pre>

mitian$ sonic-annotator -d vamp:vampy:vampy-onsets-cdsf:vampy-onsets-cdsf-onsets -w csv /homes/mitian/documents/audio/*.wav --csv-basedir /homes/mitian/hg/workspace/onsetdetection

Extracting and writing features... Done

Vampy::~PyExtensionManager: Extension module cleaned.

</pre>

h2. Play around with your vamp plugin feature extractors

We may want use parameter settings other the default. With Sonic Visualiser or Audacity, you can easily access the GUI to modify them. Whiles with Sonic Annotator, you would need to interact with 

Firstly, we can check the current parameter configurations in the skeleton transform description using Sonic Annotator's @-s@ command.

<pre>

mitian$ sonic-annotator -s vamp:vampy:vampy-onsets-cdsf:vampy-onsets-cdsf-onsets

@prefix xsd:      <http://www.w3.org/2001/XMLSchema> .

  @prefix vamp:     <http://purl.org/ontology/vamp/> .

  @prefix :         <#> .

:transform_plugin a vamp:Plugin ;

    vamp:identifier "vampy-onsets-cdsf" .

:transform_library a vamp:PluginLibrary ;

    vamp:identifier "vampy" ;

    vamp:available_plugin :transform_plugin .

:transform a vamp:Transform ;

    vamp:plugin :transform_plugin ;

    vamp:step_size "512"^^xsd:int ;

    vamp:block_size "1024"^^xsd:int ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "bt-threshold" ] ;

        vamp:value "0.9"^^xsd:float ;

    ] ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "cut-off" ] ;

        vamp:value "0.34"^^xsd:float ;

    ] ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "dthreshold" ] ;

        vamp:value "0"^^xsd:float ;

    ] ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "filtfilt" ] ;

        vamp:value "1"^^xsd:float ;

    ] ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "fusion-type" ] ;

        vamp:value "2"^^xsd:float ;

    ] ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "lin-threshold" ] ;

        vamp:value "0.5"^^xsd:float ;

    ] ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "med-threshold" ] ;

        vamp:value "7"^^xsd:float ;

    ] ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "medfilt" ] ;

        vamp:value "1"^^xsd:float ;

    ] ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "polyfit" ] ;

        vamp:value "1"^^xsd:float ;

    ] ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "threshold" ] ;

        vamp:value "50"^^xsd:float ;

    ] ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "tol-threshold" ] ;

        vamp:value "0.03"^^xsd:float ;

    ] ;

    vamp:parameter_binding [

        vamp:parameter [ vamp:identifier "whitening" ] ;

        vamp:value "0"^^xsd:float ;

    ] ;

    vamp:output [ vamp:identifier "vampy-onsets-cdsf-onsets" ] .

mitian$

</pre>

Transforms are usually described in _RDF_, following the transform part of the [["Vamp plugin ontology":http://www.omras2.org/VampOntology]]. Now, how can be change the parameter settings in the transform specification file? 

The solution is to save this RDF file and modify any content in it as we wish, then to run this transform by specifying the saved file with the @-t@ option. Let's see an example.

<pre>

mitian$ sonic-annotator -s vamp:vampy:vampy-onsets-cdsf:vampy-onsets-cdsf-onsets >param_config.n3

</pre>

Now the transform file has been saved to the .n3 (RDF) file. Say, we want to use a higher threshold and change it from 50 to 70. Open an text editor, simple modify the corresponding lines in the saved n3 file

<pre>

        vamp:parameter [ vamp:identifier "threshold" ] ;

        vamp:value "50"^^xsd:float ;

</pre>

into 

<pre>

        vamp:parameter [ vamp:identifier "threshold" ] ;

        vamp:value "70"^^xsd:float ;

</pre>

and save it.

Now we can run the same plugin using the modified parameter settings.

<pre>

mitian$ sonic-annotator -t param_config.n3 -w csv test.wav --csv-basedir /homes/mitian/hg/workspace/onsetdetection

Extracting and writing features... Done

Vampy::~PyExtensionManager: Extension module cleaned.

</pre>

This example only involves a couple of operations for Sonic Annotator. You may also want to type

@sonic-annotator -h@

for a helpbook. Also your working directory should be where Sonic Annotator is installed. 

h2. That's all guys! Happy hacking with Vamp plugins :)