Changelog

Next Version

Breaking Changes

New Features

Fixes

v5.0.0

Breaking Changes

• Changed audiomate.corpus.validation.InvalidItemsResult to use it not only for Utterances, but also for Tracks for example.
• Refactoring and addition of splitting functions in the audiomate.corpus.subset.Splitter.

New Features

• Added audiomate.corpus.validation.TrackReadValidator to check for corrupt audio tracks/files.
• Added reader (audiomate.corpus.io.FluentSpeechReader) for the Fluent Speech Commands Dataset.
• Added functions to check for contained tracks and issuers (audiomate.corpus.CorpusView.contains_track(), audiomate.corpus.CorpusView.contains_issuer()).
• Multiple options for controlling the behavior of the audiomate.corpus.io.KaldiWriter.
• Added writer (audiomate.corpus.io.Wav2LetterWriter) for the wav2letter engine.
• Added module with functions to read/write sclite trn files (audiomate.formats.trn).

Fixes

• Improved performance of Tuda-Reader (audiomate.corpus.io.TudaReader).
• Added wrapper for the `audioread.audio_open` function (audiomate.utils.audioread) to cache available backends. This speeds up audioopen operations a lot.
• Performance improvements, especially for importing utterances, merging, subviews.

v4.0.1

Fixes

• Fix audiomate.corpus.io.CommonVoiceReader to use correct file-extension of the audio files.

v4.0.0

Breaking Changes

• For utterances and labels -1 was used for representing that the end is the same as the end of the parent utterance/track. In order to prevent -1 checks in different methods/places float('inf') is now used. This makes it easier to implement stuff like label overlapping.
• audiomate.annotations.LabelList is now backed by an interval-tree instead of a simple list. Therefore the labels have no fixed order anymore. The interval-tree provides functionality for operations like merging, splitting, finding overlaps with much lower code complexity.
• Removed module audiomate.annotations.label_cleaning, since those methods are available on audiomate.annotations.LabelList directly.

New Features

• Added reader (audiomate.corpus.io.RouenReader) and downloader (audiomate.corpus.io.RouenDownloader) for the LITIS Rouen Audio scene dataset.
• Added downloader (audiomate.corpus.io.AEDDownloader) for the Acoustic Event Dataset.
• [#69] Method to get labels within range: audiomate.annotations.LabelList.labels_in_range().
• [#68] Add convenience method to create Label-List with list of label values: audiomate.annotations.LabelList.with_label_values().
• [#61] Added function to split utterances of a corpus into multiple utterances with a maximal duration: audiomate.corpus.CorpusView.split_utterances_to_max_time().
• Add functions to check for overlap between labels: audiomate.annotations.Label.do_overlap() and audiomate.annotations.Label.overlap_duration().
• Add function to merge equal labels that overlap within a label-list: audiomate.annotations.LabelList.merge_overlapping_labels().
• Added reader (audiomate.corpus.io.AudioMNISTReader) and downloader (audiomate.corpus.io.AudioMNISTDownloader) for the AudioMNIST dataset.

Fixes

• [#76][#77][#78] Multiple fixes on KaldiWriter

v3.0.0

Breaking Changes

• Moved label-encoding to its own module (audiomate.encoding). It now provides the processing of full corpora and store it in containers.
• Moved audiomate.feeding.PartitioningFeatureIterator to the audiomate.feeding module.
• Added audiomate.containers.AudioContainer to store audio tracks in a single file. All container classes are now in a separate module audiomate.containers.
• Corpus now contains Tracks not Files anymore. This makes it possible to different kinds of audio sources. Audio from a file is now included using audiomate.tracks.FileTrack. New is the audiomate.tracks.ContainerTrack, which reads data stored in a container.
• The audiomate.corpus.io.DefaultReader and the audiomate.corpus.io.DefaultWriter now load and store tracks, that are stored in a container.
• All functionality regarding labels was moved to its own module audiomate.annotations.
• The class audiomate.tracks.Utterance was moved to the tracks module.

New Features

• Introducing the audiomate.feeding module. It provides different tools for accessing container data. Via a audiomate.feeding.Dataset data can be accessed by indices. With a audiomate.feeding.DataIterator one can easily iterate over data, such as frames.
• Added processing steps for computing Onset-Strength (audiomate.processing.pipeline.OnsetStrength)) and Tempogram (audiomate.processing.pipeline.Tempogram)).
• Introduced audiomate.corpus.validation module, that is used to validate a corpus.
• Added reader (audiomate.corpus.io.SWCReader) for the SWC corpus. But it only works for the prepared corpus.
• Added function (audiomate.corpus.utils.label_cleaning.merge_consecutive_labels_with_same_values()) for merging consecutive labels with the same value
• Added downloader (audiomate.corpus.io.GtzanDownloader) for the GTZAN Music/Speech.
• Added audiomate.corpus.assets.Label.tokenized() to get a list of tokens from a label. It basically splits the value and trims whitespace.
• Added methods on audiomate.corpus.CorpusView, audiomate.corpus.assets.Utterance and audiomate.corpus.assets.LabelList to get a set of occurring tokens.
• Added audiomate.encoding.TokenOrdinalEncoder to encode labels of an utterance by mapping every token of the label to a number.
• Create container base class (audiomate.corpus.assets.Container), that can be used to store arbitrary data per utterance. The audiomate.corpus.assets.FeatureContainer is now an extension of the container, that provides functionality especially for features.
• Added functions to split utterances and label-lists into multiple parts. (audiomate.corpus.assets.Utterance.split(), audiomate.corpus.assets.LabelList.split())
• Added audiomate.processing.pipeline.AddContext to add context to frames, using previous and subsequent frames.
• Added reader (audiomate.corpus.io.MailabsReader) and downloader (audiomate.corpus.io.MailabsDownloader) for the M-AILABS Speech Dataset.

Fixes

• [#58] Keep track of number of samples per frame and between frames. Now the correct values will be stored in a Feature-Container, if the processor implements it correctly.
• [#72] Fix bug, when reading samples from utterance, using a specific duration, while the utterance end is not defined.

v2.0.0

Breaking Changes

• Update various readers to use the correct label-list identifiers as defined in Data Mapping.

New Features

• Added downloader (audiomate.corpus.io.TatoebaDownloader) and reader (audiomate.corpus.io.TatoebaReader) for the Tatoeba platform.
• Added downloader (audiomate.corpus.io.CommonVoiceDownloader) and reader (audiomate.corpus.io.CommonVoiceReader) for the Common Voice Corpus.
• Added processing steps audiomate.processing.pipeline.AvgPool and audiomate.processing.pipeline.VarPool for computing average and variance over a given number of sequential frames.
• Added downloader (audiomate.corpus.io.MusanDownloader) for the Musan Corpus.
• Added constants for common label-list identifiers/keys in audiomate.corpus.

v1.0.0

Breaking Changes

• The (pre)processing module has moved to audiomate.processing. It now supports online processing in chunks. For this purpose a pipeline step can require context. The pipeline automatically buffers data, until enough frames are ready.

New Features

• Added downloader (audiomate.corpus.io.FreeSpokenDigitDownloader) and reader (audiomate.corpus.io.FreeSpokenDigitReader) for the Free-Spoken-Digit-Dataset.

v0.1.0

Initial release