### Run HuBERT Pre-training (First Iteration)

Source: https://github.com/pytorch/audio/blob/main/examples/self_supervised_learning/README.md

Execute the SLURM command to start the first iteration of HuBERT pre-training. Ensure you have generated K-Means labels beforehand and are in the 'examples' directory. Adjust dataset path, experiment directory, and training parameters as needed.

```bash
cd examples

srun \
--gpus-per-node=8 \
--ntasks-per-node=8 \
-N 4 \
--cpus-per-task=10 \
python -m self_supervised_learning.train_hubert \
--dataset-path hubert/exp/data/mfcc/ \
--exp-dir self_supervised_learning/exp_iter1 \
--feature-type mfcc \
--num-class 100 \
--max-updates 250000 \
--learning-rate 0.0005 \
--gpus 8 \
--num-nodes 4
```

--------------------------------

### Check TorchAudio Installation

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

Import the torchaudio library and print its version to confirm a successful installation.

```python
import torchaudio

print(torchaudio.__version__)
```

--------------------------------

### Install Face Detection from GitHub

Source: https://github.com/pytorch/audio/blob/main/examples/avsr/data_prep/tools/README.md

Use this method to install the library directly from its GitHub repository. Ensure Git LFS is installed for handling large files. This installs the package in editable mode.

```Shell
git clone https://github.com/hhj1897/face_detection.git
cd face_detection
git lfs pull
pip install -e .
cd ..
```

--------------------------------

### Install Additional Dependencies

Source: https://github.com/pytorch/audio/blob/main/examples/asr/librispeech_conformer_rnnt_biasing/README.md

Installs necessary Python packages for the project. Ensure PyTorch and TorchAudio are installed with GPU support.

```bash
pip install pytorch-lightning sentencepiece
```

--------------------------------

### Install Additional Dependencies

Source: https://github.com/pytorch/audio/blob/main/examples/asr/librispeech_conformer_rnnt/README.md

Installs necessary Python packages for training and evaluation. Ensure PyTorch and TorchAudio are installed nightly or from source with GPU support.

```bash
pip install pytorch-lightning sentencepiece tensorboard
```

--------------------------------

### Install Dependencies

Source: https://github.com/pytorch/audio/blob/main/examples/avsr/data_prep/README.md

Installs all required dependency packages for the project. Ensure you have the `requirements.txt` file in your current directory.

```Shell
pip install -r requirements.txt
```

--------------------------------

### Install Dependencies

Source: https://github.com/pytorch/audio/blob/main/examples/avsr/README.md

Installs PyTorch and other necessary packages for the project. Ensure PyTorch is installed from source if needed.

```Shell
pip install torch torchvision torchaudio pytorch-lightning sentencepiece
```

--------------------------------

### Install Build Tools for TorchAudio

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

Install the 'ninja' build system, which is required for building TorchAudio.

```bash
pip install ninja
```

--------------------------------

### Install pip in User Environment

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

Download and run the get-pip.py script to install pip in the user's home directory. This avoids permission issues when installing packages.

```bash
wget https://bootstrap.pypa.io/get-pip.py
python get-pip.py --user
```

--------------------------------

### Install TensorBoard Support

Source: https://github.com/pytorch/audio/blob/main/examples/pipeline_tacotron2/README.md

Installs packages required for using TensorBoard to visualize training metrics and results.

```bash
pip install tensorboard pillow
```

--------------------------------

### Build Documentation with Tutorials

Source: https://github.com/pytorch/audio/blob/main/CONTRIBUTING.md

Installs additional requirements and builds the documentation including tutorials, which can be time-consuming.

```bash
pip install -r requirements-tutorials.txt
BUILD_GALLERY=1 make html
```

--------------------------------

### Install sentencepiece and download pretrained files

Source: https://github.com/pytorch/audio/blob/main/examples/asr/librispeech_cuda_ctc_decoder/README.md

Installs the sentencepiece library and downloads necessary pretrained model files for the speech recognition inference script.

```bash
pip install sentencepiece

# download pretrained files
wget -nc https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01/resolve/main/data/lang_bpe_500/bpe.model
wget -nc https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01/resolve/main/exp/cpu_jit.pt
```

--------------------------------

### Build Documentation

Source: https://github.com/pytorch/audio/blob/main/CONTRIBUTING.md

Commands to build the HTML documentation locally after installing requirements.

```bash
cd docs
make html
```

--------------------------------

### Build and Install TorchAudio

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

Clone the TorchAudio repository, navigate into the directory, and install it in editable mode using pip, enabling CUDA support. The '-v' flag provides verbose output during installation.

```bash
git clone https://github.com/pytorch/audio
cd audio
USE_CUDA=1 pip install -v -e . --no-use-pep517
```

--------------------------------

### Install TorchAudio Dependencies

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

Install the necessary FFmpeg development libraries required for building TorchAudio.

```bash
sudo apt install libavformat-dev libavcodec-dev libavutil-dev libavdevice-dev libavfilter-dev
```

--------------------------------

### Install Face Detection from Compressed Files

Source: https://github.com/pytorch/audio/blob/main/examples/avsr/data_prep/tools/README.md

Use this method if you encounter over-quota issues with GitHub repositories. Download the zip file, extract it, and then install the package in editable mode.

```Shell
wget https://www.doc.ic.ac.uk/~pm4115/tracker/face_detection.zip -O ./face_detection.zip
unzip -o ./face_detection.zip -d ./ 
cd face_detection
pip install -e .
cd ..
```

--------------------------------

### Install Ninja Build Tool

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.linux.rst

Installs the Ninja build system, which is a required dependency for building torchaudio.

```bash
conda install ninja
```

--------------------------------

### Install Testing Dependencies

Source: https://github.com/pytorch/audio/blob/main/CONTRIBUTING.md

Installs the necessary Python packages for running tests in Torchaudio.

```bash
pip install typing pytest scipy numpy parameterized
```

--------------------------------

### Build Documentation Requirements

Source: https://github.com/pytorch/audio/blob/main/CONTRIBUTING.md

Installs the Python package requirements for building the documentation locally.

```bash
cd docs
pip install -r requirements.txt
```

--------------------------------

### Check Installed JetPack Version

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

Verify the installed version of the nvidia-jetpack package.

```bash
# JetPack
$ apt list --installed | grep nvidia-jetpack
```

--------------------------------

### Install PyTorch Nightly Build

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.linux.rst

Installs the nightly build of PyTorch for CPU. Ensure you have selected the desired PyTorch version from their official website.

```bash
pip install --pre --index-url https://download.pytorch.org/whl/nightly/cpu
```

--------------------------------

### Check Installed CUDA Toolkit Version

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

Verify the installed version of the CUDA toolkit.

```bash
# CUDA
$ apt list --installed | grep cuda-toolkit
```

--------------------------------

### Build Libtorchaudio and Applications

Source: https://github.com/pytorch/audio/blob/main/examples/libtorchaudio/README.md

These commands are used to build libtorchaudio and its associated applications. Ensure you have a working PyTorch installation and follow the steps to clone submodules, create a build directory, and configure CMake with the correct prefix path.

```bash
git submodule update
mkdir build
cd build
cmake -GNinja \
      -DCMAKE_PREFIX_PATH="$(python -c 'import torch;print(torch.utils.cmake_prefix_path)')" \
      -DBUILD_RNNT=ON \
      ..
cmake --build .
```

--------------------------------

### Install Required Packages for Tacotron2

Source: https://github.com/pytorch/audio/blob/main/examples/pipeline_tacotron2/README.md

Installs essential libraries for running the Tacotron2 pipeline, including librosa for audio processing, tqdm for progress bars, inflect for number-to-word conversion, and joblib for parallel processing.

```bash
pip install librosa tqdm inflect joblib
```

--------------------------------

### Install DeepPhonemizer for Phoneme Input

Source: https://github.com/pytorch/audio/blob/main/examples/pipeline_tacotron2/README.md

Installs the DeepPhonemizer package, version 0.0.15, which is used as the phonemizer for converting text to phonemes in the Tacotron2 training pipeline.

```bash
pip install deep-phonemizer==0.0.15
```

--------------------------------

### Add NVIDIA Jetson Repository and Install JetPack

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

Add the NVIDIA Jetson repository to your system's sources and update the package list. Then, upgrade existing packages and install the JetPack meta-package. A reboot is required after this step.

```bash
sudo bash -c 'echo "deb https://repo.download.nvidia.com/jetson/common r34.1 main" >> /etc/apt/sources.list.d/nvidia-l4t-apt-source.list'
sudo bash -c 'echo "deb https://repo.download.nvidia.com/jetson/t234 r34.1 main" >> /etc/apt/sources.list.d/nvidia-l4t-apt-source.list'

sudo apt update
sudo apt dist-upgrade

# REBOOT

sudo apt install nvidia-jetpack
```

--------------------------------

### Generate Audio with Torchaudio Pretrained Models

Source: https://github.com/pytorch/audio/blob/main/examples/pipeline_tacotron2/README.md

This example demonstrates generating audio using specific torchaudio pretrained checkpoints for both Tacotron2 and WaveRNN. It specifies the checkpoint names and requires DeepPhonemizer for phoneme preprocessing.

```bash
python inference.py \
    --checkpoint-path tacotron2_english_phonemes_1500_epochs_wavernn_ljspeech \
    --wavernn-checkpoint-path wavernn_10k_epochs_8bits_ljspeech \
    --vocoder wavernn \
    --n-fft 2048 \
    --mel-fmin 40 \
    --mel-fmax 11025 \
    --input-text "Hello world!" \
    --text-preprocessor english_phonemes \
    --phonimizer DeepPhonemizer \
    --phoimizer-checkpoint ./en_us_cmudict_forward.pt \
    --cmudict-root ./ \
    --output-path "./outputs.wav"
```

--------------------------------

### torchaudio.datasets.LIBRISPEECH Dataset Example

Source: https://context7.com/pytorch/audio/llms.txt

Example of using the LIBRISPEECH dataset. It automatically downloads and caches the data. The dataset returns tuples of (waveform, sample_rate, utterance, speaker_id, chapter_id, utterance_id).

```python
import torchaudio
from torch.utils.data import DataLoader

# Downloads to ./data/LibriSpeech/train-clean-100
dataset = torchaudio.datasets.LIBRISPEECH(
    root="./data",
    url="train-clean-100",
    download=True,
)

print(f"Utterances: {len(dataset)}")  # 28539

waveform, sample_rate, transcript, speaker_id, chapter_id, utt_id = dataset[0]
print(waveform.shape, sample_rate)  # e.g., torch.Size([1, 93680]) 16000
print(transcript[:50])

loader = DataLoader(dataset, batch_size=None, num_workers=4)
```

--------------------------------

### torchaudio.datasets.SPEECHCOMMANDS Dataset Example

Source: https://context7.com/pytorch/audio/llms.txt

Example of using the SPEECHCOMMANDS dataset for keyword spotting. It returns (waveform, sample_rate, label, speaker_id, utterance_number) and supports V1 (30 classes) and V2 (35 classes).

```python
import torchaudio

dataset = torchaudio.datasets.SPEECHCOMMANDS(
    root="./data",
    url="speech_commands_v0.02",
    subset="training",
    download=True,
)

waveform, sr, label, speaker_id, utt_num = dataset[0]
print(f"Label: {label}, shape: {waveform.shape}, sr: {sr}")

# Unique labels
labels = set(item[2] for item in dataset)
print(f"Classes ({len(labels)}): {sorted(labels)}")
```

--------------------------------

### Verify pip Installation Location

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

After installing pip in the user environment, check that the 'pip' command points to the correct executable in the user's local bin directory.

```bash
$ which pip
/home/USER/.local/bin/pip
```

--------------------------------

### Download and Install PyTorch for Jetson

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

Download the specific PyTorch wheel file provided by NVIDIA for your JetPack version and install it using pip. Ensure you use the correct package name corresponding to your system's architecture and Python version.

```bash
$ package=torch-1.13.0a0+340c4120.nv22.06-cp38-cp38-linux_aarch64.whl
$ wget "https://developer.download.nvidia.com/compute/redist/jp/v50/pytorch/${package}"
$ pip install --no-cache "${package}"
```

--------------------------------

### Sample Training Command

Source: https://github.com/pytorch/audio/blob/main/examples/asr/librispeech_conformer_rnnt_biasing/README.md

Example SLURM command for training a Conformer RNN-T model with TCPGen. Requires specifying experiment directory, LibriSpeech path, SentencePiece model, and biasing list. Adjust droprate and maxsize for biasing list configuration.

```bash
srun --cpus-per-task=16 --gpus-per-node=1 -N 1 --ntasks-per-node=1 python train.py --exp-dir <Path_to_exp> --librispeech-path <Path_to_librispeech_data> --sp-model-path ./spm_unigram_600_100suffix.model --biasing --biasing-list ./blists/rareword_f15.txt --droprate 0.1 --maxsize 200 --epochs 90
```

--------------------------------

### Verify PyTorch Installation and CUDA Accessibility

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

Run a Python script to check the installed PyTorch version, confirm CUDA availability, and test CUDA device accessibility by creating a tensor on the GPU.

```python
import torch
print(torch.__version__)
print(torch.cuda.is_available())
print(torch.empty((1, 2), device=torch.device("cuda")))
```

--------------------------------

### SLURM Job Submission Script

Source: https://github.com/pytorch/audio/blob/main/examples/source_separation/README.md

Example script for submitting a training job to a SLURM cluster.

```bash
#!/bin/bash

#SBATCH --job-name=source_separation

#SBATCH --output=/checkpoint/%u/jobs/%x/%j.out

#SBATCH --error=/checkpoint/%u/jobs/%x/%j.err

#SBATCH --nodes=1

#SBATCH --ntasks-per-node=2

#SBATCH --cpus-per-task=8

#SBATCH --mem-per-cpu=16G

#SBATCH --gpus-per-node=2

#srun env
srun wrapper.sh $@

```

--------------------------------

### Sample Evaluation Command

Source: https://github.com/pytorch/audio/blob/main/examples/asr/librispeech_conformer_rnnt_biasing/README.md

Example SLURM command for evaluating a trained Conformer RNN-T model with TCPGen. Specify the model checkpoint, dataset path, SentencePiece model, experiment directory, and biasing list. Set droprate to 0.0 for evaluation.

```bash
srun --cpus-per-task=16 --gpus-per-node=1 -N 1 --ntasks-per-node=1 python eval.py --checkpoint-path <Path_to_model_checkpoint> --librispeech-path <Path_to_librispeech_data> --sp-model-path ./spm_unigram_600_100suffix.model --expdir <Path_to_exp> --use-cuda --biasing --biasing-list ./blists/all_rare_words.txt --droprate 0.0 --maxsize 1000
```

--------------------------------

### Build Specific Tutorial

Source: https://github.com/pytorch/audio/blob/main/CONTRIBUTING.md

Builds only a specific tutorial by using the GALLERY_PATTERN environment variable. Omitting BUILD_GALLERY assumes it is set to 1.

```bash
BUILD_GALLERY=1 GALLERY_PATTERN=forced_alignment_tutorial.py make html
```

```bash
GALLERY_PATTERN=forced_alignment_tutorial.py make html
```

--------------------------------

### Check Installed cuDNN Version

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.jetson.rst

Verify the installed version of the cuDNN library.

```bash
# cuDNN
$ apt list --installed | grep cudnn
```

--------------------------------

### Sample Training Command

Source: https://github.com/pytorch/audio/blob/main/examples/asr/librispeech_conformer_rnnt/README.md

Launches the training script for the Conformer RNN-T model. Requires a full LibriSpeech dataset, a SentencePiece model, and pre-computed global statistics.

```bash
srun --cpus-per-task=12 --gpus-per-node=8 -N 4 --ntasks-per-node=8 python train.py --exp-dir ./experiments --librispeech-path ./librispeech/ --global-stats-path ./global_stats.json --sp-model-path ./spm_unigram_1023.model --epochs 160
```

--------------------------------

### Run HuBERT Pre-training (1st Iteration)

Source: https://github.com/pytorch/audio/blob/main/examples/hubert/README.md

Command to train a HuBERTPretrainModel for the first iteration using PyTorch Lightning on GPU nodes. This is for 250k steps on 32 GPUs.

```bash
srun --gpus-per-node=8 --ntasks-per-node=8 -N 4 --cpus-per-task=10 python train.py --dataset-path ./exp/data/mfcc/ --exp-dir ./exp_iter1 --feature-type mfcc --num-class 100 --max-updates 250000 --learning-rate 0.0005 --gpus 8 --num-nodes 4
```

--------------------------------

### Load and Use a Dataset with DataLoader

Source: https://github.com/pytorch/audio/blob/main/docs/source/datasets.rst

Demonstrates how to instantiate a dataset (YESNO) and load it using PyTorch's DataLoader. Ensure the dataset is downloaded if necessary.

```python
yesno_data = torchaudio.datasets.YESNO('.', download=True)
data_loader = torch.utils.data.DataLoader(
    yesno_data,
    batch_size=1,
    shuffle=True,
    num_workers=args.nThreads)
```

--------------------------------

### Run Training Script

Source: https://github.com/pytorch/audio/blob/main/examples/source_separation/README.md

Execute the PyTorch-Lightning training script. Use `--help` for a full list of parameters.

```bash
python lightning_train.py \
        [--data-dir DATA_DIR] \
        [--num-gpu NUM_GPU] \
        [--num-workers NUM_WORKERS] \
        ...
```

```bash
python lightning_train.py --help
```

--------------------------------

### Instantiate and Use Audio Processing Pipeline

Source: https://github.com/pytorch/audio/blob/main/docs/source/transforms.rst

Demonstrates how to instantiate a custom audio processing pipeline and move it to a CUDA device for computation. This snippet shows the typical workflow for preparing and running a transform pipeline on audio data.

```python
# Instantiate a pipeline
pipeline = MyPipeline()

# Move the computation graph to CUDA
pipeline.to(device=torch.device("cuda"), dtype=torch.float32)

# Perform the transform
features = pipeline(waveform)
```

--------------------------------

### Run HuBERT Pre-training (2nd Iteration)

Source: https://github.com/pytorch/audio/blob/main/examples/hubert/README.md

Command to train the HuBERT model for the second iteration for 400k steps. This uses the clustering labels generated in the second pre-processing step.

```bash
srun --gpus-per-node=8 --ntasks-per-node=8 -N 4 --cpus-per-task=10 python train.py --dataset-path ./exp/data/hubert_6/ --exp-dir ./exp_iter2 --feature-type hubert --num-class 500 --max-updates 400000 --learning-rate 0.0005 --gpus 8 --num-nodes 4
```

--------------------------------

### Run HuBERT Pre-processing (1st Iteration)

Source: https://github.com/pytorch/audio/blob/main/examples/hubert/README.md

Command to generate data file lists and train a KMeans clustering model using MFCC features for the first iteration of HuBERT pre-training.

```bash
srun --cpus-per-task=24 python preprocess.py --root-dir /home/datasets --feat-type mfcc --exp-dir ./exp --num-cluster 100
```

--------------------------------

### Run Wav2Letter++ Training Pipeline

Source: https://github.com/pytorch/audio/blob/main/examples/pipeline_wav2letter/README.md

Invoke the main training script with specified dataset paths, hyperparameters, and training configurations. Ensure dataset paths are correctly set before execution.

```bash
DATASET_ROOT = <Top>/<level>/<folder>
DATASET_FOLDER_IN_ARCHIVE =  'LibriSpeech'

python main.py \
    --reduce-lr-valid \
    --dataset-root "${DATASET_ROOT}" \
    --dataset-folder-in-archive "${DATASET_FOLDER_IN_ARCHIVE}" \
    --dataset-train train-clean-100 train-clean-360 train-other-500 \
    --dataset-valid dev-clean \
    --batch-size 128 \
    --learning-rate .6 \
    --momentum .8 \
    --weight-decay .00001 \
    --clip-grad 0. \
    --gamma .99 \
    --hop-length 160 \
    --win-length 400 \
    --n-bins 13 \
    --normalize \
    --optimizer adadelta \
    --scheduler reduceonplateau \
    --epochs 40
```

--------------------------------

### Generate Audio with WaveRNN Vocoder (Phonemes)

Source: https://github.com/pytorch/audio/blob/main/examples/pipeline_tacotron2/README.md

This command generates audio using WaveRNN with phoneme-based text preprocessing. It requires installing DeepPhonemizer and downloading its pretrained weights.

```bash
python inference.py --checkpoint-path ${model_path} \
    --vocoder wavernn \
    --n-fft 2048 \
    --mel-fmin 40 \
    --mel-fmax 11025 \
    --input-text "Hello world!" \
    --text-preprocessor english_phonemes \
    --phonimizer DeepPhonemizer \
    --phoimizer-checkpoint ./en_us_cmudict_forward.pt \
    --cmudict-root ./ \
    --output-path "./outputs.wav"
```

--------------------------------

### VGGishBundle.VGGishInputProcessor

Source: https://github.com/pytorch/audio/blob/main/docs/source/_templates/autosummary/bundle_class.rst

Documentation for VGGishBundle.VGGishInputProcessor, detailing its callable methods.

```APIDOC
## VGGishBundle.VGGishInputProcessor

### Description
Documentation for VGGishBundle.VGGishInputProcessor.

### Methods

__call__

.. container:: py attribute

   .. automethod:: torchaudio.prototype.models.vggish.VGGishBundle.VGGishInputProcessor.__call__
```

--------------------------------

### Enable MSVC 2022 x64 Toolset

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.windows.rst

Call the vcvarsall.bat script to enable the MSVC x64 toolset for compilation. Ensure the path to the script is correct for your Visual Studio installation.

```batch
call "C:\Program Files (x86)\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvarsall.bat" x64
```

--------------------------------

### View Head of File List

Source: https://github.com/pytorch/audio/blob/main/examples/libtorchaudio/speech_recognition/README.md

Display the first few lines of the generated file list to verify its format and content.

```bash
head flist-de.txt
```

--------------------------------

### Build TorchScript Pipeline from Fairseq

Source: https://github.com/pytorch/audio/blob/main/examples/libtorchaudio/speech_recognition/README.md

Use this script to create TorchScript files for audio loading, encoding, and decoding from a fairseq wav2vec model.

```bash
mkdir -p pipeline-fairseq
python build_pipeline_from_fairseq.py \
    --model-file "wav2vec_small_960.pt" \
    --dict-dir <DIRECTORY_WHERE_dict.ltr.txt_IS_FOUND> \
    --output-path "./pipeline-fairseq/"

```

--------------------------------

### SLURM Wrapper Script for Training

Source: https://github.com/pytorch/audio/blob/main/examples/source_separation/README.md

A wrapper script to configure and launch the PyTorch-Lightning training job on a SLURM cluster. Ensure dataset paths and resource allocations are correctly set.

```bash
#!/bin/bash
num_speakers=2
this_dir="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
exp_dir="/checkpoint/${USER}/exp/"
root_dir="/dataset/" # The directory where the directory ``Libri2Mix`` or ``Libri3Mix`` is stored.
num_gpu=2 # The number of GPUs used on one node.
num_node=1 # The number of nodes used on the cluster.
batch_size=6 # The batch size per GPU.


mkdir -p "${exp_dir}"

python -u \
  "${this_dir}/lightning_train.py" \
  --num-speakers "${num_speakers}" \
  --sample-rate 8000 \
  --root-dir "${root_dir}" \
  --exp-dir "${exp_dir}" \
  --num-gpu ${num_gpu} \
  --num-node ${num_node} \
  --batch-size ${batch_size} \

```

--------------------------------

### Wav2Letter++ Training Output Log

Source: https://github.com/pytorch/audio/blob/main/examples/pipeline_wav2letter/README.md

Example JSON output lines from the training process, detailing metrics like character error rate (CER), word error rate (WER), loss, and learning rate per iteration and epoch. This format is suitable for parsing with libraries like pandas.

```json
{"name": "train", "epoch": 40, "batch char error": 925, "batch char total": 22563, "batch char error rate": 0.040996321411159865, "epoch char error": 1135098.0, "epoch char total": 23857713.0, "epoch char error rate": 0.047577821059378154, "batch word error": 791, "batch word total": 4308, "batch word error rate": 0.18361188486536675, "epoch word error": 942906.0, "epoch word total": 4569507.0, "epoch word error rate": 0.20634742435015418, "lr": 0.06, "batch size": 128, "n_channel": 13, "n_time": 1685, "dataset length": 132096.0, "iteration": 1032.0, "loss": 0.07428030669689178, "cumulative loss": 90.47326805442572, "average loss": 0.08766789540157531, "iteration time": 1.9895553588867188, "epoch time": 2036.8874564170837}
```

```json
{"name": "train", "epoch": 40, "batch char error": 1131, "batch char total": 24260, "batch char error rate": 0.0466199505358615, "epoch char error": 1136229.0, "epoch char total": 23881973.0, "epoch char error rate": 0.04757684802675223, "batch word error": 957, "batch word total": 4657, "batch word error rate": 0.2054971011380717, "epoch word error": 943863.0, "epoch word total": 4574164.0, "epoch word error rate": 0.20634655862798099, "lr": 0.06, "batch size": 128, "n_channel": 13, "n_time": 1641, "dataset length": 132224.0, "iteration": 1033.0, "loss": 0.08775319904088974, "cumulative loss": 90.5610212534666, "average loss": 0.08766797798012256, "iteration time": 2.108018159866333, "epoch time": 2038.99547457695}
```

```json
{"name": "train", "epoch": 40, "batch char error": 1099, "batch char total": 23526, "batch char error rate": 0.0467142735696676, "epoch char error": 1137328.0, "epoch char total": 23905499.0, "epoch char error rate": 0.04757599914563591, "batch word error": 936, "batch word total": 4544, "batch word error rate": 0.20598591549295775, "epoch word error": 944799.0, "epoch word total": 4578708.0, "epoch word error rate": 0.20634620071863066, "lr": 0.06, "batch size": 128, "n_channel": 13, "n_time": 1682, "dataset length": 132352.0, "iteration": 1034.0, "loss": 0.0791337713599205, "cumulative loss": 90.64015502482653, "average loss": 0.08765972439538348, "iteration time": 2.0329701900482178, "epoch time": 2041.0284447669983}
```

```json
{"name": "train", "epoch": 40, "batch char error": 1023, "batch char total": 22399, "batch char error rate": 0.045671681771507655, "epoch char error": 1138351.0, "epoch char total": 23927898.0, "epoch char error rate": 0.04757421650660664, "batch word error": 863, "batch word total": 4318, "batch word error rate": 0.1998610467809171, "epoch word error": 945662.0, "epoch word total": 4583026.0, "epoch word error rate": 0.20634009058643787, "lr": 0.06, "batch size": 128, "n_channel": 13, "n_time": 1644, "dataset length": 132480.0, "iteration": 1035.0, "loss": 0.07874362915754318, "cumulative loss": 90.71889865398407, "average loss": 0.08765110981061262, "iteration time": 1.9106628894805908, "epoch time": 2042.9391076564789}
```

```json
{"name": "validation", "epoch": 40, "cumulative loss": 12.095281183719635, "dataset length": 2688.0, "iteration": 21.0, "batch char error": 1867, "batch char total": 14792, "batch char error rate": 0.12621687398593834, "epoch char error": 37119.0, "epoch char total": 280923.0, "epoch char error rate": 0.13213229247872194, "batch word error": 1155, "batch word total": 2841, "batch word error rate": 0.4065469904963041, "epoch word error": 22601.0, "epoch word total": 54008.0, "epoch word error rate": 0.418475040734706, "average loss": 0.575965770653316, "validation time": 24.185853481292725}
```

--------------------------------

### Run HuBERT Pre-processing (2nd Iteration)

Source: https://github.com/pytorch/audio/blob/main/examples/hubert/README.md

Command to train a new KMeans clustering model using the output from the 6th transformer layer of the first pre-training iteration. The number of clusters is increased to 500.

```bash
srun --cpus-per-task=24 python preprocess.py --root-dir /home/datasets --feat-type hubert --exp-dir ./exp --layer-index 6 --checkpoint-path ./exp_iter1/ --num-rank 40 checkpoints_librispeech_hubert_pretrain_base/xxx.ckpt --num-cluster 500 --percent 0.1
```

--------------------------------

### Build TorchScript Pipeline from Hugging Face Transformers

Source: https://github.com/pytorch/audio/blob/main/examples/libtorchaudio/speech_recognition/README.md

Use this script to import and convert a wav2vec2.0 model from Hugging Face Hub into TorchScript format for C++ applications.

```bash
mkdir -p pipeline-hf
python build_pipeline_from_huggingface_transformers.py \
    --model facebook/wav2vec2-large-xlsr-53-german \
    --output-path ./pipeline-hf/

```

--------------------------------

### Run HuBERT Fine-tuning

Source: https://github.com/pytorch/audio/blob/main/examples/hubert/README.md

Command to fine-tune the pre-trained HuBERT model on the LibriLightLimited dataset. Feature extraction layers are frozen initially, then unfrozen along with the CTC layer.

```bash
srun --gpus-per-node=1 -N 1 --ntasks-per-node=1 --cpus-per-task=10 \
  python finetune.py --dataset-path /root/datasets/ --exp-dir ./exp_finetune \
  --checkpoint ./exp_iter2/checkpoints_librispeech_hubert_pretrain_base/epoch=361-step=399999.ckpt \
  --gpus 1 --debug --warmup-updates 2000 --hold-updates 8000 --decay-updates 10000 --max-updates 20000 --learning-rate 5e-5
```

--------------------------------

### Sample Evaluation Command

Source: https://github.com/pytorch/audio/blob/main/examples/asr/librispeech_conformer_rnnt/README.md

Evaluates a trained Conformer RNN-T model on the LibriSpeech test-clean dataset. Ensure the checkpoint path and dataset paths are correctly specified.

```bash
srun python eval.py --checkpoint-path ./experiments/checkpoints/epoch=159.ckpt --librispeech-path ./librispeech/ --sp-model-path ./spm_unigram_1023.model --use-cuda
```

--------------------------------

### Build torchaudio In-Place

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.linux.rst

Builds the torchaudio library in-place using pip. The '-v' flag provides verbose output, and '--no-build-isolation' is required for in-place builds.

```bash
pip install -e . -v --no-build-isolation
```

--------------------------------

### Tacotron2TTSBundle.TextProcessor

Source: https://github.com/pytorch/audio/blob/main/docs/source/_templates/autosummary/bundle_class.rst

Documentation for Tacotron2TTSBundle.TextProcessor, including its properties and callable methods.

```APIDOC
## Tacotron2TTSBundle.TextProcessor

### Description
Documentation for Tacotron2TTSBundle.TextProcessor.

Properties
----------

tokens

.. container:: py attribute

   .. autoproperty:: torchaudio.prototype.models.tacotron2.Tacotron2TTSBundle.TextProcessor.tokens

Methods
-------

__call__

.. container:: py attribute

   .. automethod:: torchaudio.prototype.models.tacotron2.Tacotron2TTSBundle.TextProcessor.__call__
```

--------------------------------

### Wav2Vec2FABundle.Tokenizer

Source: https://github.com/pytorch/audio/blob/main/docs/source/_templates/autosummary/bundle_class.rst

Documentation for Wav2Vec2FABundle.Tokenizer, detailing its callable methods.

```APIDOC
## Wav2Vec2FABundle.Tokenizer

### Description
Documentation for Wav2Vec2FABundle.Tokenizer.

### Methods

__call__

.. container:: py attribute

   .. automethod:: torchaudio.prototype.models.wav2vec2.Wav2Vec2FABundle.Tokenizer.__call__
```

--------------------------------

### VGGishBundle.VGGish

Source: https://github.com/pytorch/audio/blob/main/docs/source/_templates/autosummary/bundle_class.rst

Documentation for VGGishBundle.VGGish, detailing its forward method.

```APIDOC
## VGGishBundle.VGGish

### Description
Documentation for VGGishBundle.VGGish.

Methods
-------

forward

.. container:: py attribute

   .. automethod:: torchaudio.prototype.models.vggish.VGGishBundle.VGGish.forward
```

--------------------------------

### Clone torchaudio Repository

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.linux.rst

Clones the torchaudio repository from GitHub and navigates into the project directory.

```bash
git clone https://github.com/pytorch/audio
cd audio
```

--------------------------------

### Generate VoxForge German Dataset File List

Source: https://github.com/pytorch/audio/blob/main/examples/libtorchaudio/speech_recognition/README.md

Create a file list for the VoxForge German dataset by running a Python script with the path to your dataset. The output file `flist-de.txt` will contain audio paths and their transcriptions.

```bash
python parse_voxforge.py <PATH_TO_YOUR_DATASET> > ./flist-de.txt
```

--------------------------------

### Run Inference Script for LibriSpeech

Source: https://github.com/pytorch/audio/blob/main/examples/asr/librispeech_ctc_decoder/README.md

This command shows how to execute the inference script. Specify the path to the LibriSpeech dataset, the desired split, the pre-trained model, and parameters for beam search, language model weight, and word score.

```bash
python inference.py \
    --librispeech_path ./librispeech/ \
    --split test-other \
    --model WAV2VEC2_ASR_BASE_960H \
    --beam-size 1500 \
    --lm-weight 1.74 \
    --word-score 0.52
```

--------------------------------

### Launch Bash from Miniconda

Source: https://github.com/pytorch/audio/blob/main/docs/source/build.windows.rst

Optional step to launch a Bash environment from Miniconda for a Linux/macOS-like user experience. Note that file paths and module compatibility may differ in this environment.

```bash
Miniconda3\Library\bin\bash.exe
```

--------------------------------

### Fork and Create Branch

Source: https://github.com/pytorch/audio/blob/main/CONTRIBUTING.md

Standard Git commands to fork the repository and create a new branch for development.

```bash
git checkout main && git checkout -b my_cool_feature
```

--------------------------------

### Perform WaveRNN Inference

Source: https://github.com/pytorch/audio/blob/main/examples/pipeline_wavernn/README.md

Run the inference script using a pre-trained checkpoint to generate an audio waveform. The output file path can be specified.

```bash
python inference.py \
    --checkpoint-name wavernn_10k_epochs_8bits_ljspeech \
    --output-wav-path ./output.wav

```

--------------------------------

### Wav2Vec2Model

Source: https://github.com/pytorch/audio/blob/main/docs/source/_templates/autosummary/model_class.rst

Documentation for the Wav2Vec2Model class, including its methods and related factory functions.

```APIDOC
## Wav2Vec2Model

### Description
Documentation for the Wav2Vec2Model class.

### Methods
#### forward

.. automethod:: torchaudio.models.Wav2Vec2Model.forward

#### extract_features

.. automethod:: torchaudio.models.Wav2Vec2Model.extract_features

### Factory Functions

.. autosummary::
   :toctree: ../generated
   :nosignatures:

   ~torchaudio.models.wav2vec2_model
   ~torchaudio.models.wav2vec2_base
   ~torchaudio.models.wav2vec2_large
   ~torchaudio.models.wav2vec2_large_lv60k
   ~torchaudio.models.wav2vec2_xlsr_300m
   ~torchaudio.models.wav2vec2_xlsr_1b
   ~torchaudio.models.wav2vec2_xlsr_2b
   ~torchaudio.models.hubert_base
   ~torchaudio.models.hubert_large
   ~torchaudio.models.hubert_xlarge
   ~torchaudio.models.wavlm_model
   ~torchaudio.models.wavlm_base
   ~torchaudio.models.wavlm_large

### Utility Functions

.. currentmodule:: torchaudio.models

.. autosummary::
   :toctree: ../generated
   :nosignatures:

   ~torchaudio.models.wav2vec2.utils.import_fairseq_model
   ~torchaudio.models.wav2vec2.utils.import_huggingface_model

.. autosummary::
   :toctree: ../generated
   :nosignatures:

   ~torchaudio.models.emformer_hubert_model
   ~torchaudio.models.emformer_hubert_base
   ~torchaudio.models.conformer_wav2vec2_model
   ~torchaudio.models.conformer_wav2vec2_base

```

--------------------------------

### torchaudio.load_with_torchcodec

Source: https://github.com/pytorch/audio/blob/main/docs/source/torchaudio.rst

Loads an audio file using TorchCodec.

```APIDOC
## torchaudio.load_with_torchcodec

### Description
Loads an audio file using TorchCodec. This is the recommended way to load audio files.

### Method
Not applicable (Python function)

### Endpoint
Not applicable (Python function)

### Parameters
Refer to the official TorchCodec documentation for detailed parameters.

### Request Example
```python
import torchaudio

speech_waveform, sample_rate = torchaudio.load_with_torchcodec("speech.wav")
```

### Response
- **waveform** (Tensor) - The audio waveform.
- **sample_rate** (int) - The sample rate of the audio.
```