Thin Ribbon of Snow

[pzelasko]

Breaking changes

• Lhotse CutSampler classes now return mini-batch CutSets instead of a list of string cut IDs (Lhotse Dataset classes are adjusted correspondingly) (Overhaul of Dataset API #345)
• Cut refactoring (Cut is now an abstract base class for all cut types; what was previously called Cut is now called MonoCut) (Refactor and document the Cut hierarchy (Cut -> MonoCut, CutUtilsBase -> Cut) #328)
• CLI: lhotse obtain is now lhotse download (Adding GigaSpeech recipe CLI, minor GigaSpeech fixes, renaming "lhotse obtain" to "lhotse download" #329)

Corpora

• TIMIT (prepare timit manifests #324 thanks @luomingshuang)
• Fisher English (Add Fisher English #374 thanks @videodanchik)
• Fisher Spanish (Feature/fisher spanish #376 thanks @videodanchik)
• yesno (Add recipe for the yes_no dataset. #380 thanks @csukuangfj)
• improvements to GigaSpeech recipe (Adding GigaSpeech recipe CLI, minor GigaSpeech fixes, renaming "lhotse obtain" to "lhotse download" #329 Chunked HDF5 feature storage + minor recording fixes + adjust GigaSpeech recipe #334 Fix and validate the manifests in GigaSpeech #337 Downloading GigaSpeech with PySpeechColab #381 thanks @jimbozhang)
• including word alignments in LibriSpeech recipe (Adding LibriSpeech word alignments in supervisions #379)

New features

CutSampler improvements (PyTorch data API)

• ZipSampler for batches constructed from different cut sources (ZipSampler for batches constructed from different cut sources #344 Fix docstring for ZipSampler #347 Fix test involving ZipSampler #363 thanks for fixes @janvainer)
• drop_last option and get_report() method for cut samplers (Samplers: drop_last option and get_report() diagnostic method #357)
• find_pessimistic_batches utility to help fail fast with GPU OOM (Sampling: find_pessimistic_batches utility to help fail fast with GPU OOM. #358)
• streaming variant of shuffling for lazy CutSets in samplers (Sampling: streaming variant of shuffling for lazy CutSets #359)
• a bucketing method with equal cumulative bucket duration for BucketingSampler (Add a bucketing method with equal cumulative bucket duration #365)
• approximate proportional sampling in BucketingSampler (Approximate proportional sampling in BucketingSampler; remaining_duration, remaining_cuts, num_cuts properties for samplers. #372)

I/O improvements

• chunked OPUS file reads (Special OPUS logic for efficient chunk reads and auto-resampling #339)
• chunked sphere file reads (add chunked sph file processing #367 thanks @videodanchik)
• faster OnTheFlyFeatures (padding audio instead of features) (Padding features instead of audio in OnTheFlyFeatures #352)
• ChunkedLilcomHdf5Writer (and reader) for efficient chunk reads of lilcom-compressed arrays (Chunked HDF5 feature storage + minor recording fixes + adjust GigaSpeech recipe #334)
• a global cache for re-using smart_open connection sessions (improves performance for repeated smart_open calls e.g., to S3) (Wrapper around smart_open to cache transport_params for every call #335, thanks @oplatek)

Data augmentation

• tempo perturbation (Add tempo perturbation #375 thanks @janvainer)
• volume perturbation (Feature/volume perturbation #382 thanks @videodanchik)

Others

• CutSet.trim_to_supervision has new arguments for including actual acoustic context next to the supervisions (Improving and fixing trim_to_supervisions, adding more documentation, fixes to filter_supervisions #330 Adding trim_to_supervisions to Cut; new options to include acoustic context when using trim_to_supervisions #331)
• SupervisionSegment is now mutable (and all Lhotse manifests will remain mutable) (Mutable supervisions #333)
• .shuffle() method for Lhotse *Set classes (Adding .shuffle() method for *Set classes #341)
• lhotse fix CLI (CLI: lhotse fix #360)
• lhotse install-sph2pipe for handling LDC corpora compressed with shorten (auto-registers sph2pipe so no further actions are needed) (Utilities to install sph2pipe #370)

General improvements

• refreshed docs (Update supervision related docs #327 Refactor and document the Cut hierarchy (Cut -> MonoCut, CutUtilsBase -> Cut) #328 Improving and fixing trim_to_supervisions, adding more documentation, fixes to filter_supervisions #330)
• improvements to downloading corpora (Avoid archive downloads when the unpacked directory exists  #340)
• experimental dataloader that allows two levels of parallelism (Custom DataLoader with two levels of subprocess workers #343, might be abandoned for other alternatives)
• auto-detection of compatible torchaudio version for pytorch (Detect the compatible torchaudio version when torch is already installed. #348)
• improvements to Kaldi data dir import/export (Support exporting various audio formats to Kaldi data dir #351 Fix import of Kaldi data dirs with pipes in wav.scp #354)
• fixed cut ordering in CutSet.subset(cut_ids=...) (Fix cut ordering in CutSet.subset(cut_ids=...) #353)
• improvements to storing cuts as recordings (Avoid extra work in compute_and_store_recording + sensible default for open_writer() #355)
• refactored lhotse.dataset.sampling file into a directory module (Splitting sampling.py to multiple files #366)
• improvements to CLI (update cmd line for bash scripts #369 Make lhotse/bin/lhotse executable #371 thanks @songmeixu)
• improvements to setup (fix dev requires in setup.py #377 Change how lhotse entry point is registered in setuptools #383 thanks @songmeixu)
• Colab notebook with ESPnet + Lhotse example (Mention Icefall and ESPnet colab notebook in documentation as examples of use #384)
• improvements to Lhotse versioning (Versioning updates + release v0.8 #385)

---

This discussion was created from the release Thin Ribbon of Snow.