Piotr Żelasko
Piotr Żelasko
Resolves https://github.com/lhotse-speech/lhotse/discussions/857 I don't think it's really possible to have the same duration buckets used on each GPU node, because every node may see a subset of data and have...
# Motivation Currently Lhotse offers two types of data storage: - random-access (HDF5, lilcom-chunky, individual files, etc.) with manifests that are fast to read and easy to extend with new...
This PR shows how to use Lhotse Shar data format to train models in Icefall. I only added the necessary adaptation to the basic TDNN LSTM recipe as I don't...
# What does this PR do ? Extensions in Lhotse dataloading augmentations: * RIR augmentation support (currently requires lhotse RecordingSet input, see e.g. https://github.com/lhotse-speech/lhotse/blob/master/lhotse/recipes/rir_noise.py ) * Broader input type support...
# What does this PR do ? Add a one line overview of what this PR aims to accomplish. **Collection**: [Note which collection this PR will affect] # Changelog -...
# What does this PR do ? Add a one line overview of what this PR aims to accomplish. **Collection**: [Note which collection this PR will affect] # Changelog -...
# What does this PR do ? Context: together with @galv we found feature normalization and specaug to take approx 30% of the total forward step time in Canary training,...
# What does this PR do ? The key config options to set are: ``` +model.train_ds.input_cfg=/path/to/speech_text_input.yaml \ +model.train_ds.use_multimodal_sampling=true \ +model.train_ds.batch_tokens=65536 \ +model.train_ds.token_equivalent_duration=0.08 \ +model.train_ds.quadratic_factor=4096 \ +model.train_ds.tarred_audio_filepaths=null \ model.train_ds.manifest_filepath=null \ model.train_ds.batch_size=null...
# What does this PR do ? * speed-up in training start when duration bins are not provided (we will assume uniform split to num_buckets - 1 bins over the...
# What does this PR do ? Generic prompt formatter for text modality with several out-of-the-box prompt format definitions. See the class documentation for more details. Also, enables support for...