About Sonic Annotator

Sonic Annotator is a batch tool for feature extraction and annotation of audio files. The audio to be processed can be on the local filesystem or available or over http or ftp. It will run available Vamp plugins on a wide range of audio file types, and can write the results in a selection of formats.

A Quick Tutorial

To use Sonic Annotator, you need to tell it three things: what audio files to extract features from; what features to extract; and how and where to write the results. You can also optionally tell it to summarise the features.

1. What audio files to extract features from

Sonic Annotator accepts a list of audio files on the command line. Any argument that is not understood as a supported command-line option will be taken to be the name of an audio file. Any number of files may be listed.

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.

File paths do not have to be local; you can also provide remote HTTP or FTP URLs for Sonic Annotator to retrieve.

Sonic Annotator also accepts the names of playlist files (with .m3u extension) and will process every file found in the playlist.

A limitation of the current version of Sonic Annotator on Windows is that it requires forward slash as the path separator ("/") instead of backslash ("\") to avoid writing incorrect URLs into the output in RDF writer mode. For example, C:/audio/testfile.wav.

Finally, you can provide a local directory path instead of a file, together with the -r (recursive) option, for Sonic Annotator to process every audio file found in that directory or any of its subdirectories.

2. What features to extract

Sonic Annotator applies "transforms" to its input audio files, where a transform (in this terminology) consists of a Vamp plugin together with a certain set of parameters and a specified execution context including step and block size, sample rate, etc.

To use Sonic Annotator normally, you need to create a file that describes the properties of the transform you want to apply, and then tell Sonic Annotator about it by supplying the transform's filename on the command line with the -t option. There is also a quick way of applying the default configuration of a plugin: see "Default transforms" below.

Transforms are usually described in RDF, following the transform part of the Vamp plugin ontology. A Transform may use any Vamp plugin that is currently installed and available on the system.

You can obtain a list of available plugin outputs by running Sonic Annotator with the -l or --list option:

  $ sonic-annotator -l

And you can obtain a skeleton transform description for one of these plugins with the -s or --skeleton option:

  $ sonic-annotator -s vamp:vamp-example-plugins:fixedtempo:tempo
  @prefix xsd:      <http://www.w3.org/2001/XMLSchema> .
  @prefix vamp:     <http://purl.org/ontology/vamp/> .
  @prefix :         <#> .

  :transform a vamp:Transform ;
      vamp:plugin <http://vamp-plugins.org/rdf/plugins/vamp-example-plugins#fixedtempo> ;
      vamp:step_size "64"^^xsd:int ; 

      vamp:block_size "256"^^xsd:int ; 
      vamp:parameter_binding [
          vamp:parameter [ vamp:identifier "maxbpm" ] ;
          vamp:value "190"^^xsd:float ;
      ] ;
      vamp:parameter_binding [
          vamp:parameter [ vamp:identifier "maxdflen" ] ;
          vamp:value "10"^^xsd:float ;
      ] ;
      vamp:parameter_binding [
          vamp:parameter [ vamp:identifier "minbpm" ] ;
          vamp:value "50"^^xsd:float ;
      ] ;
      vamp:output <http://vamp-plugins.org/rdf/plugins/vamp-example-plugins#fixedtempo_output_tempo> .

The output of this example is an RDF/Turtle document describing the default settings for the Tempo output of the Fixed Tempo Estimator plugin in the Vamp plugin SDK.

(The exact format of the RDF printed may differ -- e.g. if the plugin's RDF description is not installed and so its "home" URI is not known -- but the result should be functionally equivalent to this.)

You can then run this transform by saving the RDF to a file and specifying that file with -t or --transform. You will also need to supply at least one writer option (see "How and where to write the results" below for more about those).

  $ sonic-annotator -s vamp:vamp-example-plugins:fixedtempo:tempo > test.n3
  $ sonic-annotator -t test.n3 audio.wav -w csv --csv-stdout
  (... logging output on stderr, then ...)
  "audio.wav",0.002902494,5.196916099,68.7916,"68.8 bpm" 

The single line of output above consists of the audio file name, the timestamp and duration for a single feature, the value of that feature (the estimated tempo of the given region of time from that file, in bpm -- the plugin in question performs a single tempo estimation and nothing else) and the feature's label.

To run more than one transform on the same set of audio files, just put more than one set of transform RDF descriptions in the same file, or give the -t option more than once with separate transform description files. Remember that if you want to specify more than one transform in the same file, they will need to have distinct URIs (that is, the :transform part of the example above, which may be any arbitrary name, must be distinct for each described transform). You can also list lots of transform filenames in a list file and use the -T or --transforms option to tell Sonic Annotator to load all of those.

Default transforms

A quicker way to run a single plugin in its default configuration is to use the -d (default) option:

  $ sonic-annotator -d vamp:vamp-example-plugins:fixedtempo:tempo audio.wav -w csv --csv-stdout
  (... some log output on stderr, then ...)
  "audio.wav",0.002902494,5.196916099,68.7916,"68.8 bpm" 

Although handy for experimentation, the -d option is inadvisable in any "production" situation because the plugin configuration is not guaranteed to be the same each time (for example if an updated version of a plugin changes some of its defaults). It's better to save a well-defined transform to a file and refer to that, even if it is just the transform created by the skeleton generator (the -s option).

Transform configuration

The following description applies to transforms expressed in RDF/Turtle format. It is also possible to describe them in an XML format, which is not documented here.

Plugin identifier and output

The plugin itself, and the output to obtain features from, is specified using the vamp:plugin and vamp:output properties. Normally these will take URI values matching those given in the RDF file distributed along with the plugin. The --skeleton option will generate valid URIs for any plugin and output that is currently installed.

Plugin parameters

These associate a parameter ID, as a plain literal string, with a value, which is also a literal but that always contains a floating-point number. (Vamp plugin parameters are always numeric: boolean parameters, for example, take the value 1 for on and 0 for off.) An example is shown in the skeleton output above.


A small number of plugins support "programs", a means of setting several parameters at once according to a label. For example, an onset detector might have programs for instruments with "hard" or "soft" onsets each of which sets a number of separate parameters. You can set these using the vamp:program property.

Sample rate

The audio sample rate (in Hz) to which input audio will be resampled before being presented to the plugin. The default is to use the sample rate of the first audio file.

Start time and duration

These are newly supported in Sonic Annotator v1.1. To apply a plugin to only part of an input audio file, you can provide vamp:start_time and vamp:duration properties specifying the range to use as input. These take rather fiddly xsd:duration format values:

:transform0 a vamp:Transform;
    vamp:plugin examples:percussiononsets ;
    vamp:output examples:percussiononsets_output_detectionfunction ;
    vamp:start "PT2.0S"^^xsd:duration ;
    vamp:duration "PT2.0S"^^xsd:duration .

Note that the actual start and end of the audio passed to the plugin will depend on Sonic Annotator's internal processing block size.

There will usually be some additional samples included at the end, as the duration is usually not an exact multiple of the internal processing block size and the actual audio data provided is always rounded up to that.

At the start, the start time will be matched exactly if only a single transforms is being run, or if all transforms share the same start time. Otherwise there may be some extra samples at the start, because the actual start time will snap to the next earlier multiple of the block size since the start time of any transform that began sooner.

Step and block size

These are straightforwardly defined using vamp:step_size and vamp:block_size properties.

The block size is the number of audio samples per frame (time or frequency domain) passed to the plugin's process function. If unspecified, this will either take the preferred size requested by the plugin (if any) or 1024.

The step size is the increment in audio samples from one processing frame to the next. For plugins taking time-domain input, this is usually the same as the block size (blocks do not overlap). For those taking frequency-domain input the default is half the step size, a 50% overlap.

Window type

For plugins taking frequency-domain input, you can choose the window shape used for time-domain frames prior to the short-time Fourier transform. To do so, supply a vamp:window_type property taking one of the literal values: "rectangular" "bartlett" "hamming" "hanning" "blackman" "gaussian" "parzen" "nuttall" or "blackman-harris".

Plugin version

Newly supported in Sonic Annotator v1.1. To ensure repeatable results, you can specify a particular version of the plugin using the vamp:plugin_version property. If the plugin actually installed is found to have a different version, Sonic Annotator will refuse to use it.

3. How and where to write the results

Sonic Annotator supports various different output modules (and it is fairly easy for the developer to add new ones). You have to choose at least one output module; use the -w (writer) option to do so. Each module has its own set of parameters which can be adjusted on the
command line, as well as its own default rules about where to write the results.

The following writers are currently supported. (Others exist, but are not properly implemented or not supported.)


Writes the results into comma-separated data files.

One file is created for each transform applied to each input audio file, named after the input audio file and transform name with .csv suffix and ":" replaced by "_" throughout, placed in the same directory as the audio file.

To instruct Sonic Annotator to place the output files in another location, use --csv-basedir with a directory name.

To write a single file with all data in it, use --csv-one-file.

To write all data to standard output instead of to a file, use --csv-stdout.

Sonic Annotator will not write to an output file that already exists. If you want to make it do this, use --csv-force to overwrite or --csv-append to append to it.

The data generated consists of one line for each result feature, containing the feature timestamp, feature duration if present, all of the feature's bin values in order, followed by the feature's label if present. If the --csv-one-file or --csv-stdout option is specified, then an additional column will appear before any of the above, containing the audio file name from which the feature was extracted, if it differs from that of the previous row. To suppress this additional column, use the --csv-omit-filenames option.

To make the CSV writer emit the end time instead of the duration (for features with duration) use the --csv-end-times option.

To make the writer always emit end time or duration, even when the feature lacks duration, by using the time of the following feature as the end time, use the --csv-fill-ends option.

The default column separator is a comma; you can specify a different one with the --csv-separator option.


Writes the results into a tab-separated label file (.lab).

This is equivalent to using the CSV writer with a tab separator and the options --csv-end-times --csv-omit-filenames.

It supports the --lab-basedir, --lab-one-file, --lab-stdout, --lab-force, --lab-append, and --lab-fill-ends options, which all behave similarly to their CSV writer equivalents.


Writes the results into RDF/Turtle documents following the Audio Features ontology.

One file is created for each input audio file containing the features extracted by all transforms applied to that file, named after the input audio file with .n3 extension, placed in the same directory as the audio file.

To instruct Sonic Annotator to place the output files in another location, use --rdf-basedir with a directory name.

To write a single file with all data (from all input audio files) in it, use --rdf-one-file.

To write one file for each transform applied to each input audio file, named after the input audio file and transform name with .n3 suffix and ":" replaced by "_" throughout, use --rdf-many-files.

To write all data to standard output instead of to a file, use --rdf-stdout.

Sonic Annotator will not write to an output file that already exists. If you want to make it do this, use --rdf-force to overwrite or --rdf-append to append to it.

Sonic Annotator will use plugin description RDF if available to enhance its output (for example identifying note onset times as note onset times, if the plugin's RDF says that is what it produces, rather than writing them as plain events). Best results will be obtained if an RDF document is provided with your plugins (for example, vamp-example-plugins.n3) and you have this installed in the same location as the plugins. To override this enhanced output and write plain events for all features, use --rdf-plain.

The output RDF will include an available_as property linking the results to the original audio signal URI. By default, this will point to the URI of the file or resource containing the audio that Sonic Annotator processed, such as the file:/// location on disk. To override this, for example to process a local copy of a file while generating RDF that describes a copy of it available on a network, you can use the --rdf-signal-uri option to specify an alternative signal URI.


Writes the results into JSON format following JAMS, the JSON Annotated Music Specification. This writer is provisional as of Sonic Annotator v1.1.


Writes the results to MIDI files. All features are written as MIDI notes.

If a feature has at least one value, its first value will be used as the note pitch, the second value (if present) for velocity. If a feature has units of Hz, then its pitch will be converted from frequency to an integer value in MIDI range, otherwise it will be written directly.

Multiple (up to 16) transforms can be written to a single MIDI file, where they will be given separate MIDI channel numbers.

4. Optionally, how to summarise the features

Sonic Annotator can also calculate and write summaries of features, such as mean and median values.

To obtain a summary as well as the feature results, just use the -S option, naming the type of summary you want (min, max, mean, median, mode, sum, variance, sd or count). You can also tell it to produce only the summary, not the individual features, with --summary-only.

Alternatively, you can specify a summary in a transform description. The following example tells Sonic Annotator to write both the times of note onsets estimated by the simple percussion onset detector example plugin, and the variance of the plugin's onset detection function.
(It will only process the audio file and run the plugin once.)

  @prefix rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns>.
  @prefix vamp: <http://purl.org/ontology/vamp/>.
  @prefix examples: <http://vamp-plugins.org/rdf/plugins/vamp-example-plugins>.
  @prefix : <#>.

  :transform1 a vamp:Transform;
     vamp:plugin examples:percussiononsets ;
     vamp:output examples:percussiononsets_output_onsets .

  :transform2 a vamp:Transform;
     vamp:plugin examples:percussiononsets ;
     vamp:output examples:percussiononsets_output_detectionfunction ;
     vamp:summary_type "variance" .

Sonic Annotator can also summarise in segments — if you provide a comma-separated list of times as an argument to the --segments option, it will calculate one summary for each segment bounded by the times you provided. For example,

  $ sonic-annotator -d vamp:vamp-example-plugins:percussiononsets:detectionfunction \
    -S variance --sumary-only --segments 1,2,3 -w csv --csv-stdout audio.wav
  (... some log output on stderr, then ...)
  "audio.wav",0.000000000,1.000000000,variance,1723.99,"(variance, continuous-time average)" 
  ,1.000000000,1.000000000,variance,1981.75,"(variance, continuous-time average)" 
  ,2.000000000,1.000000000,variance,1248.79,"(variance, continuous-time average)" 
  ,3.000000000,7.031020407,variance,1030.06,"(variance, continuous-time average)" 

Here the first row contains a summary covering the time period from 0 to 1 second, the second from 1 to 2 seconds, the third from 2 to 3 seconds and the fourth from 3 seconds to the end of the (short) audio file.