v1.8 Sudden Avalanche

Breaking changes

• Python 3.6 is no longer supported as of Lhotse v1.8. If you need to use Python 3.6, please revert to Lhotse 1.7 and earlier.

Highlights

• New experimental module of lhotse: workflows, now integrates optional third party packages that assist corpus creators in automated data curation. With release 1.8, we support OpenAI Whisper for automatic transcription and segmentation, and torchaudio Wav2Vec2/Hubert ASR bundles for forced alignment.

What's Changed

• Fix read and write in piped CLI by @desh2608 in #807
• Default behavior of CutSet.mix by @ZuoyunZheng in #809
• Adding more info about resampling options by @RuABraun in #815
• Add pad_silence option to extend_by by @desh2608 in #816
• Message when calling len() on LazyFilter by @desh2608 in #817
• Refactor cut and retain git blame history by @desh2608 in #820
• Audio backend refactoring and a workaround for FLAC reading from/writing to in-memory buffers by @pzelasko in #814
• Experimental Lhotse feature: corpus creation tools (workflows), starting with OpenAI Whisper support by @pzelasko in #824
• Drop support for Python 3.6 by @pzelasko in #829
• [workflow] Word-level forced alignment with pretrained models from Torchaudio by @pzelasko in #827

New Contributors

• @ZuoyunZheng made their first contribution in #809

Full Changelog: v1.7...v1.8

v1.7 - Rejuvenation Potion

What's Changed

• add test data to bvcc by @oplatek in #797
• Add reverb with fast RIR generator by @desh2608 in #799
• Support snip_edges=True in online_inference of Kaldi feature extractors by @pzelasko in #802
• Remove warning about Lhotse not being stable from README.md by @pzelasko in #804
• Update the documentation related to optional packages by @pzelasko in #805

Full Changelog: v1.6...v1.7

1.6 - Frozen Palm Tree

What's Changed

• Feature/fix 754 voxceleb download by @mikuchar in #776
• Support Kaldi data dierectories without segments file. by @MartinKocour in #789
• Add normalization for text of mulit_cn recipe:thchs_30, tal_csasr, tal_asr, aishell, aishell2,etc by @shanguanma in #760
• Improve support for custom Recordings by @pzelasko in #791
• Add Cut.has(field) method to query Cuts for custom attributes by @pzelasko in #792
• Add normalization for aishell2 recipe by @shanguanma in #790

New Contributors

• @mikuchar made their first contribution in #776
• @MartinKocour made their first contribution in #789

Full Changelog: 1.5...v1.6

1.5 - Little Leaf

What's Changed

• Describe more information about cuts by @pzelasko in #772
• Change vctk.py to adapt the vctk dataset downloaded from edinburgh url by @luomingshuang in #775
• Fix restoring sampler state with world_size>1 by @pzelasko in #773
• Revert #738 to use aidatatang as the prefix for aidatatang_200zh. by @csukuangfj in #782
• use tolerance when checking duration mismatch by @shaynemei in #781

New Contributors

• @shaynemei made their first contribution in #781

Full Changelog: v1.4...v1.5

v1.4 - Candescent Crust

What's Changed

• Fix lambda warnings from lazy manifests + leverage dill if installed for pickling lambdas by @pzelasko in #748
• multi_cn recipes: aishell2, magicdata, primewords, stcmds, tal_asr, tal_csasr, thchs_30 by @shanguanma in #738
• Deprecate strict, proportional_sampling, and bucket_method arguments by @pzelasko in #756
• Fix lhotse cut simple CLI by @pzelasko in #759
• Fix issues with eager CutSet creation from lazy manifests by @pzelasko in #763
• DailyTalk recipe by @pzelasko in #767
• add aishell2 dev test by @yuekaizhang in #766
• Enable GlobalMVN computation with on-the-fly feature extraction by @pzelasko in #769
• Add support for Python 3.10 and PyTorch 1.12 by @pzelasko in #764

New Contributors

• @yuekaizhang made their first contribution in #766

Full Changelog: v1.3...1.4

v1.3 - Curiously Inviting Icicles

What's Changed

• Fix plotting MixedCut audio tracks by @pzelasko in #723
• [continued] Fixes Bucketing sampler equal duration method that drops cuts by @m-wiesner in #724
• feature extraction will read RecordingSet from a file, not just json. by @RuABraun in #728
• Use lilcom_chunky as default in CLI by @pzelasko in #729
• Set CLI torch number of threads to 1 by @pzelasko in #732
• Update wenet_speech.py by @fanlu in #731
• Fix heroico regex strings by @jtrmal in #734
• Update mgb2.py by @AmirHussein96 in #725
• Remove file handle caching from LilcomChunkyReader by @pzelasko in #737
• Make h5py an optional dependency by @pzelasko in #741
• Assert CutSet.mix() argument cuts is not a lazy manifest by @pzelasko in #742
• CutSet: more methods are lazy + two simplified common use-cases attach_tensor and load_audio by @pzelasko in #744
• Collections: support reading from/writing to "-" (including webdataset) by @pzelasko in #745
• fix CommonVoice prepare by @mohsen-goodarzi in #743

New Contributors

• @RuABraun made their first contribution in #728
• @mohsen-goodarzi made their first contribution in #743

Full Changelog: v1.2...v1.3

v1.2 - Winter in the South

New Recipes

• Adding lhotse recipe to prepare eval2000 data by @GoVivace in #679
• adding Earnings-21 dataset from rev-dot-com by @jtrmal in #709
• Adding the second revdotcom's earnings corpus by @jtrmal in #713
• MGB2 recipe by @AmirHussein96 in #718

What's Changed

• Fix import namespaces by @pzelasko in #698
• .repeat(..., preserve_id=...) option for repeating manifests by @pzelasko in #699
• Kaldi impex: remove invalid test by @jtrmal in #700
• Minor fix in base url for AliMeeting download by @desh2608 in #702
• [aidatatang_200zh] Avoid being converted to ASCII when preparing manifest by @luomingshuang in #703
• [ali_meeting] Fix some path errors for ali_meeting.py by @luomingshuang in #705
• [ali_meeting] Avoid being converted to ASCII by @luomingshuang in #704
• Test for webdataset data de-duplication across ranks by @pzelasko in #706
• Fixing data duplication with WebDataset in multi-node multi-worker training by @pzelasko in #707
• Fix epoch setting for WebDataset shard shuffling by @pzelasko in #708
• Full shard shuffling with webdataset by @pzelasko in #711
• Raise an error when BucketingSampler is used with a lazy CutSet by @pzelasko in #710
• Normalize output path names for recipes by @desh2608 in #712
• [webdataset] Add shard of origin to Cut.shard_origin custom field by @pzelasko in #714
• Update examples of combining datasets with RoundRobinSampler and add stop_early option. by @pzelasko in #716
• pre-commit, isort + CI checks + running it on all code by @pzelasko in #720

New Contributors

• @GoVivace made their first contribution in #679
• @AmirHussein96 made their first contribution in #718

Full Changelog: v1.1...v1.2

v1.1 - Minor Mana Potion

What's Changed

• auto_increment_epoch in IterableDatasetWrapper, strict=True default for bucketing by @pzelasko in #661
• Support storing Recording objects in cuts custom fields by @pzelasko in #662
• nara_wpe based WPE dereverberation as data augmentation by @pzelasko in #663
• Update rirs noises path by @Tomiinek in #665
• Fix reading duration when importing piped input data from kaldi by @csukuangfj in #667
• Fixes to split-lazy by @wgb14 in #664
• Read matrix shape information without reading the whole matrix. by @csukuangfj in #668
• Set a num_digits for split-lazy by @wgb14 in #669
• DynamicBucketingSampler supports very small data by @pzelasko in #670
• ~20x faster speed perturbation by @pzelasko in #672
• Fix assertions, rename variables by @pzelasko in #677
• Add max_cuts option to DynamicBucketingSampler by @pzelasko in #681
• Fix recording_id in MixedCut.compute_and_store_features(..., mix_eagerly=True) by @pzelasko in #682
• Support for restoring state of dynamic samplers by @pzelasko in #684
• Add .cut_into_windows method to individual cuts by @pzelasko in #685
• kaldi pipeline -- small whitespace fixes and making pycheck happier by @jtrmal in #686
• Some minor fixes. by @csukuangfj in #688
• kaldi import/export -- adding basic tests for kaldi import/export by @jtrmal in #687
• Fix bug in cut.truncate, format selection in cut.save_audio, less memory use in AudioMixer by @pzelasko in #690
• Pad channels to same length when loading audio by @desh2608 in #689
• Fix restoring state in dynamic samplers with DataLoader num_workers>0 by @pzelasko in #692
• Fix samplers dropping cuts when world_size > 1 by @pzelasko in #695
• FIX: kaldi import fets.scp -- use the correct id when underscore mapping by @jtrmal in #694
• Audio loading fault tolerant feature extraction in compute_and_store_features by @pzelasko in #683

New Contributors

• @Tomiinek made their first contribution in #665
• @wgb14 made their first contribution in #664

Full Changelog: v1.0...v1.1

South Peak

Recipes

New corpora

• VoxCeleb and RIR recipes by @desh2608 in #475
• Add WenetSpeech recipe by @pkufool in #487
• ICSI meeting corpus by @desh2608 in #526
• ASpIRE data preparation recipe by @desh2608 in #528
• The People's Speech recipe by @pzelasko in #529
• bvcc/VoiceMOS challange data recipe by @oplatek in #578
• Add recipe for aidatatang_200zh. by @csukuangfj in #593
• SPGISpeech recipe by @desh2608 in #600
• [recipe] AliMeeting by @desh2608 in #608

Existing recipes improvements

• A change for tedlium.py by @luomingshuang in #479
• improve adept preparation: adds text interpretation by @oplatek in #474
• Reduce memory usage when writing GigaSpeech manifests by @pzelasko in #494
• adding previous utterance for libritts supervisions by @oplatek in #510
• bugfix for prepare_librispeech command by @rosrad in #516
• Improving Fisher recipe by @pzelasko in #539
• Some Fisher fixes and manifest validation fixes by @pzelasko in #541
• ICSI Recipe - Minor Doc correction by @LasseWolter in #544
• Minor updates for AMI and ICSI by @desh2608 in #545
• [AMI] Added missing IHM channels by @LasseWolter in #555
• Minor changes to LibriCSS and AISHELL-4 recipes by @desh2608 in #580
• Updated ICSI download structure/args for clarity by @LasseWolter in #583
• Fix(icsi-recipe): Use new directory structure in prepare_icsi by @LasseWolter in #592
• Babel recipe fix by @m-wiesner in #647

New features

Support for custom attributes in Cuts

This feature allows attaching arbitrary type of data to cuts: alignments, multiple feature sets, etc.

• Array and TemporalArray manifests (generalization of Features for arbitrary data) by @pzelasko in #458
• Support custom MonoCut attrs (with special support for loading Arrays) by @pzelasko in #459
• Collation of custom cut fields into PyTorch tensors by @pzelasko in #476
• custom_collate_field pads to the longest read array instead of using CutSet.pad() by @pzelasko in #482
• Add with_path_prefix to Array and TemporalArray by @pzelasko in #499
• Collation promotes custom field int sequences to int64 by @pzelasko in #507
• More flexibility in mixing cuts with custom attributes by @pzelasko in #642

Data augmentation

• Reverberation using room impulse response by @desh2608 in #477
• Add early reverb option for RIR-based augmentation by @desh2608 in #524
• Options for handling multi-channel RIR by @desh2608 in #621

"Lazy cuts" (less memory / faster execution)

• Option to create a CutSet lazily when everything is sorted on recording IDs by @pzelasko in #493
• Enable adding/combining of lazy manifests by @pzelasko in #495
• Resumable batch feature extraction with reduced memory usage by @pzelasko in #508
• DynamicBucketingSampler: on-the-fly bucketing with restricted memory usage by @pzelasko in #517
• DynamicCutSampler (like DynamicBucketingSampler but without bucketing) by @pzelasko in #579
• Custom binary feature storage format by @pzelasko in #522
• lhotse split-lazy and CutSet.split_lazy() for memory-efficient splits by @pzelasko in #558
• Optional dependency orjson for up to 50% JSONL reading by @pzelasko in #563
• CutSet multiplexing by @pzelasko in #565
• stop_early arg for CutSet.mux by @pzelasko in #585
• Option to set shuffle buffer size in dynamic samplers by @pzelasko in #587
• WebDataset integration for optimized sequential I/O by @pzelasko in #582
• Add WebDataset export CLI and a fault_tolerant option by @pzelasko in #599
• Add WebdatasetWriter for iterative cut writing by @pzelasko in #602
• More lazily evaluated methods: map, filter, repeat, shuffle by @pzelasko in #626

New and refreshed APIs

• Deprecate compute_and_store_recording, add save_audio by @desh2608 in #486
• CutSet.decompose() with CLI and streaming read/write enhancements by @pzelasko in #496
• Opensmile wrapper by @marcinwitkowski in #504
• fill_supervision method for cuts and cut sets by @pzelasko in #505
• merge_supervisions method for cuts and cut sets by @pzelasko in #503
• Add CLI for lhotse cut trim-to-supervision <in> <out> by @pzelasko in #514
• CLI for validating recordings+supervisions and extra supervision check by @pzelasko in #515
• rng arg for CutSet.truncate() by @pzelasko in #557
• Add cut extend method by @desh2608 in #571
• Output more information from CutSet.describe. by @csukuangfj in #606
• Cut into windows with hop by @marcinwitkowski in #651

Improved Kaldi import/export

• KaldiWriter for exporting features to feats.scp by @pzelasko in #473
• Support multi-channel wavs in Kaldi export by @pzelasko in #480
• Make speaker_id prefix of utt_id for Kaldi by @desh2608 in #530
• Fix bug in export to kaldi function by @HuangZiliAndy in #531
• Add method to create SupervisionSet from RTTM files by @desh2608 in #566
• Replace kaldiio with kaldi_native_io. by @csukuangfj in #584

PyTorch API

• Add SpecAugment state dict by @janvainer in #472
• Enable threaded batch IO by @pzelasko in #520
• GlobalMVN and Wav2LogFilterBank torchscriptable by @janvainer in #521
• Streaming Kaldi feature extractors by @pzelasko in #523
• Skipping problematic audios during dataloading - continued by @pzelasko in #533
• Make samplers picklable by @pzelasko in #542
• Rename SingleCutSampler -> SimpleCutSampler by @pzelasko in #546
• Sampler reports for how much padding is typically used by @pzelasko in #560
• Modified SpecAugment by @luomingshuang in #598
• optimization for specaugment by @luomingshuang in #604
• Return audio_lens in multi channel audio collater by @desh2608 in #616
• Add sampling statistics report to dynamic samplers by @pzelasko in #628
• New parameter: OnTheFlyFeatures(..., return_audio=True) by @pzelasko in #629
• RoundRobinSampler: samples mini-batches in turn from each sub-sampler by @pzelasko in #649
• Sampler diagnostics updates: preserved across epochs, fixes for various samplers, extended unit tests by @pzelasko in #639
• Stricter batch constraint exceeding checks by @pzelasko in #653

Documentation

• Refresh tutorials in examples + quality of life improvements in code by @pzelasko in #617
• Tutorial for Lhotse's WebDataset integration by @pzelasko in #619
• Add links in tutorials, fix some issues with lazy manifes...

Himalayan Salt Bath

This is mostly a bug, documentation, and installation fix release.

New features

Add dynamic LRU cache for audio and feature I/O (#443)

Lhotse will now cache up to 512 audio/feature arrays that were read from disk to speed-up some common usage patterns.
This behavior can be easily disabled with lhotse.set_caching_enabled(False).

Recipe improvements

• Update timit recipe by @luomingshuang in #446
• Fixes to babel recipe by @jtrmal in #447
• Fixed bug in preparing AMI supervisions by @desh2608 in #449

General improvements

• add export multichannel to kaldi by @jtrmal in #441
• Fix documentation builds by @pzelasko in #448
• fix typo of np.stack by @glynpu in #450
• Rich exception info for I/O in Recording, MonoCut and MixedCut by @pzelasko in #454
• Fix SpecAugment docs by @pzelasko in #463
• Fix misleading manifest format in the comments by @stachu86 in #460
• Ensure the users who installed PyTorch also install torchaudio themselves by @pzelasko in #466
• Add documentation and tests for Kaldi feature extraction layers by @pzelasko in #467

New Contributors

• @glynpu made their first contribution in #450
• @stachu86 made their first contribution in #460