Thawing Potion

[pzelasko]

New corpora

• CMU Arctic (Add the CMU Arctic corpus #225)
• L2 Arctic (Add the L2 Arctic corpus #227, Store L2 arctic read speech and "suitcase" corpus separately #251)
• VCTK (Add VCTK corpus #228, Fix VCTK manifest storing #253, Fix storing manifests in 3 recipes #254)
• CallHome English (CallHome English + reading audio chunks from pipe commands with a warning #278)
• CallHome Egyptian (Add tolerance to validate_recordings_and_supervisions (#205) #208)
• Multilingual Librispeech (MLS recipe and reading OPUS files with an optional audioread dependency #282 -- can be quite slow but we plan on improving the speed in further releases)

Features

PyTorch Dataset API

• Lhotse's samplers are now fully deterministic, have len() that returns the number of batches, and return a consistent number of batches in all distributed workers. (Make sampler length equal to the number of returned batches (#212) #213, Fix setting world size and rank in distributed samplers #222, [fix] Restore random bucket selection in the bucketing sampler #223, Add test for bucketing randomness to prevent regressions of this sort #224, Fix cut pairs sampler criteria #255, Samplers always return the same number of batches in distributed mode #267, thanks @janvainer)
• On-the-fly feature extraction in PyTorch datasets (On-the-fly feature extraction support in PyTorch datasets #229)
• visualisations of ASR batches with multiple transforms applied (ASR batch visualisation (features only) #234)

Features and transforms

• Add LIbrosaFbank consistent with various TTS applications (Add LIbrosaFbank consistent with various TTS applications #252, thanks @janvainer)
• SpecAugment (SpecAugment implementation #246)
• option to pad cuts from left/right/both directions (More padding configurations - left/right, duration/num_frames/num_samples #216)
• Randomized smoothing augmentation (Randomized smoothing augmentation #272, Randomized smoothing - apply with prob #273, Fix flaky randomized smoothing test #274)
• Randomized extra padding (Randomized extra padding #281)

I/O and serialization

• [experimental] Downloading audio from HTTP/S3/GCP/Azure URLs upon request ([experimental] Downloading audio from HTTP/S3/GCP/Azure URLs upon request #233)
• Use HDF5 as the default storage backend for features (Use HDF5 as the default storage backend for features #237)
• Add JSONL support (Add JSONL support #262)
• Support for auto-magically determined serializers (CLI + Python API) (Support for auto-magically determined serializers (CLI + Python API) #264)

Removed features

• Removed WavAugment support (use torchaudio.sox_effects instead) (Remove WavAugment support (use torchaudio.sox_effects instead) #232)

General improvements

• Add tolerance to validate_recordings_and_supervisions (Add tolerance to validate_recordings_and_supervisions (#205) #208, thanks @janvainer )
• Fix incorrect truncation in cut mixing for data augmentation (Fix incorrect truncation in cut mixing for data augmentation #214)
• Return lengths from feature and token collations (Return lengths from feature and token collations #211, thanks @janvainer)
• Refactor Standardize to GlobalMVN (Refactor Standardize to GlobalMVN #230, thanks @janvainer)
• Fix rare error in randomized Recording's resampling test (Fix rare error in randomized Recording's resampling test #239)
• Fix concatenate cuts omitting the longest cut when duration_factor > 1 (Fix concatenate cuts omitting the longest cut when duration_factor > 1 #240)
• Fix CutMix not adding enough noise in long cuts (Fix CutMix not adding enough noise in long cuts #241)
• Add max_cuts keyword to global stats computation in GlobalMVN (Add max_cuts keyword to global stats computation in GlobalMVN #245)
• Improved error message for mixing audio (Improved error message for mixing audio #248)
• Add a check for matching sampling rates when mixing cuts (Add a check for matching sampling rates when mixing cuts #247)
• Fix - make VCTK CLI discoverable (Fix - make VCTK CLI discoverable #250)
• Fix trim_to_supervisions and CLI (Fix trim_to_supervisions and CLI #249)
• Fix find segments float rounding issues (Fix find segments float rounding issues #265)
• Update the examples of libirispeech (full) and ami (Update the examples of libirispeech (full) and ami #268, thanks @jimbozhang)
• Fix test for sphere files (Fix test for sphere files. #269, thanks @csukuangfj)
• More informative error message for incorrect channels in load_audio() (More informative error message for incorrect channels in load_audio() #270)

---

This discussion was created from the release Thawing Potion.