Recipes
Directory for reripes of the SpeeChain toolkit
The SpeeChain toolkit organizes its recipes by task, each located within a dedicated sub-folder in /speechain/recipes/.
Every task sub-folder hosts second-level sub-folders that pertain to individual datasets.
Table of Contents
- Available Datasets
- ASR (Automatic Speech Recognition)
- TTS (Text-To-Speech Synthesis)
- Offline TTS-to-ASR Chain
- Experimental File System
Available Datasets
Each task comes with a dedicated README.md file found within its respective folder, providing further details and instructions. Follow the hyperlinks below to navigate to the README.md file for your target task.
Automatic Speech Recognition (ASR)
Refer to this sample structure for an overview of how the ASR folder is organized:
/asr
/librispeech # ASR Recipes for the LibriSpeech dataset
/train-clean-100 # Labeled data: train-clean-100
/data_cfg # Data loading configuration files
/exp_cfg # Experimental configuration files
/train-clean-460 # Labeled data: train-clean-460 (train-clean-100 + train-clean-360)
...
/train-960 # Labeled data: train-960 (train-clean-460 + train-other-500)
...
/libritts_librispeech # ASR Recipes for the joint dataset of LibriSpeech and 16khz-downsampled LibriTTS
/train-960 # Labeled data: LibriSpeech_train-960 & 16khz-LibriTTS_train-960
...
👆Back to the table of contents
Text-To-Speech Synthesis (TTS)
Refer to this sample structure for an overview of how the TTS folder is organized:
/tts
/libritts # TTS Recipes for the LibriTTS dataset
/train-clean-100 # Labeled data: train-clean-100
/data_cfg # Data loading configuration files that are shared by different models
/exp_cfg # Experimental configuration files
/train-clean-460 # Labeled data: train-clean-460 (train-clean-100 + train-clean-360)
...
/train-960 # Labeled data: train-960 (train-clean-460 + train-other-500)
...
/ljspeech # TTS Recipes for the LJSpeech dataset, LJSpeech doesn't have the official subset division.
...
/vctk # TTS Recipes for the VCTK dataset, VCTK doesn't have the official subset division.
👆Back to the table of contents
Offline TTS-to-ASR Chain
Refer to this sample structure for an overview of how the offline TTS-to-ASR Chain folder is organized:
/offline_tts2asr
/librispeech # LibriSpeech is used as labeled data to train ASR models
/train-clean-100 # ASR Labeled data: LibriSpeech_train-clean-100
/data_cfg # Data loading configuration files
/exp_cfg # Experimental configuration files
/train-960 # ASR Labeled data: LibriSpeech_train-960
...
👆Back to the table of contents
Experimental File System
The file structure of an experiment folder is as follows:
/data_cfg
/train_cfg
/exp_cfg
/exp # 'exp' folder of each model
/{exp_name} # Name of a specific experiment
/{test_cfg_name} # Name of a testing configuration file
/{test_model_name} # Name of the model you want to test the performance
/{test_set_name} # Name of a test set
/figures # Folder that contains all the distribution figures of each metric on the test set
{test_metric_name}.png # Histogram distribution figure of the {test_metric_name} values of all the testing instances
...
test.log # Log file that contains the testing process of a specific test set
overall_results.md # .md file that contains the model overall performance on the test set
instance_reports.md # .md file that contains the detailed performance reports of each testing instance
topn_(max/min)_{xxx}.md # .md file that contains the top-n bad cases selected by the metric 'xxx'
idx2{xxx} # suffix-free .txt files that contain individual metrics of each testing instance, a file corresponds to a metric 'xxx'
... # other test sets
... # other test models
... # other test configurations
models/ # This sub-folder contains all the model files
N_{xxx}_average.pth # Average model obtained by the metric 'xxx' on 'N' best models
{xxx}_best.pth # File soft link to the best model obtained by the metric 'xxx'
{xxx}_best_2.pth # File soft link to the second best model obtained by the metric 'xxx'
...
{xxx}_best_n.pth # File soft link to the n-th best model obtained by the metric 'xxx'
epoch_{X}.pth # Actual model file of the X-th epoch
epoch_{Y}.pth # Actual model file of the Y-th epoch
...
figures/ # This sub-folder contains all the snapshotting figures. The .txt files in this folder contain the data used to plot summary.png which can be used to plot your own figures.
train/ # Snapshotting figures made during training
consumed_memory/ # Folder containing the GPU memory consumption records
Rank0.txt # Numerical records of the memory consumption for the GPU rank.0 through epochs
... # Other GPUs if have
RankN.txt # Numerical records of the memory consumption for the GPU rank.N through epochs
summary.png # Curve graph of all the records above through epochs
consumed_time/ # Folder containing the time consumption records
data_load_time.txt # Numerical records of the data loading time through epochs
model_forward_time.txt # Numerical records of the model forward time through epochs
loss_backward_time_{optim_name}.txt # Numerical records of the loss backward time for the optimizer named {optim_name} through epochs
optim_time_{optim_name}.txt # Numerical records of the parameter optimization time for the optimizer named {optim_name} through epochs
summary.png # Curve graph of all the records above through epochs
criteria/ # Folder containing the training criterion value records
{train_criterion_name} # Numerical records of the training criterion named {train_criterion_name} through epochs
... # Other training criteria if have
summary.png # Curve graph of all the records above through epochs
optim_lr/ # Folder containing the learning rate records of each optimizer
{optim_name}.txt # Numerical recores of the learning rates of the optimizer named {optim_name} through epochs
... # Other optimizers if have
summary.png # Curve graph of all the records above through epochs
valid/ # Snapshotting figures made during validation
consumed_memory/
Rank0.txt
...
RankN.txt
summary.png
consumed_time/
data_load_time.txt
model_forward_time.txt
summary.png
criteria/ # Folder containing the validation criterion value records
{valid_criterion_name} # Numerical records of the validation criterion named {valid_criterion_name} through epochs
... # Other validation criteria if have
summary.png # Curve graph of all the records above through epochs
tensorboard/ # This sub-folder contains the writer events for tensorboard visualization
checkpoint.pth # Checkpoint of the training process so far, used for resuming the training process
train_data_cfg.yaml # Data loading configuration for the training-validation part of the experiment, used for resuming the training process
test_data_cfg.yaml # Data loading configuration for the testing part of the experiment, used for resuming the testing process
exp_cfg.yaml # Experiment environment configuration for the experiment, used for resuming the training process
train_cfg.yaml # Model and optimizer configuration for the experiment, used for resuming the training process
train.log # Log file that contains the training process of the given training sets and validation sets