### Install pre-commit Package

Source: https://github.com/k2-fsa/icefall/blob/master/contributing.md

Install the pre-commit package using pip. This is a one-time setup step.

```bash
pip install pre-commit
```

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

### Whisper Model Installation and Training Setup

Source: https://github.com/k2-fsa/icefall/blob/master/egs/aishell/ASR/RESULTS.md

Installs Whisper model requirements and prepares data for fine-tuning.

```bash
pip install -r whisper/requirements.txt

./prepare.sh --stage 30 --stop_stage 30
```

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

### Install Documentation Dependencies

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/contributing/doc.md

Navigate to the docs directory and install the required Python packages for building documentation.

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

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

### Install Dependencies and Build Documentation

Source: https://github.com/k2-fsa/icefall/blob/master/docs/README.md

Navigate to the docs directory, install requirements, clean previous builds, and generate HTML documentation.

```bash
cd /path/to/icefall/docs
pip install -r requirements.txt
make clean
make html
```

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

### Install kaldifeat

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/for-dummies/model-export.md

Install the kaldifeat library with specific version and PyTorch compatibility. Use the provided URL for CPU-specific builds.

```bash
pip install kaldifeat==1.25.3.dev20231221+cpu.torch2.0.0 -f https://csukupeangfj.github.io/kaldifeat/cpu.html
```

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

### Install Dependencies and Download Models

Source: https://github.com/k2-fsa/icefall/blob/master/egs/speech_llm/ASR_LLM/RESULTS.md

Installs necessary Python packages and downloads pre-trained Whisper and Qwen models required for fine-tuning.

```bash
pip install -r whisper_llm_zh/requirements.txt

pip install huggingface_hub['cli']
mkdir -p models/whisper models/qwen

# For aishell fine-tuned whisper model
huggingface-cli download --local-dir models/whisper    yuekai/icefall_asr_aishell_whisper exp_large_v2/whisper-large-v2-aishell1-epoch-10-avg-6.pt
# For multi-hans fine-tuned whisper model
# huggingface-cli download --local-dir models/whisper    yuekai/icefall_asr_multi-hans-zh_whisper v1.1/whisper-large-v2-multi-hans-zh-epoch-3-avg-10.pt

# huggingface-cli download  --local-dir models/qwen     Qwen/Qwen2-7B-Instruct
huggingface-cli download  --local-dir models/qwen     Qwen/Qwen2-1.5B-Instruct
```

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

### Install kaldifeat

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/for-dummies/model-export.md

Install the kaldifeat library with a specific version and PyTorch compatibility. This command is used for installing the necessary tools for feature extraction.

```bash
pip install kaldifeat==1.25.3.dev20231221+cpu.torch2.0.0 -f https://csukuangfj.github.io/kaldifeat/cpu.html
```

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

### TensorBoard Upload Output Example

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/yesno/tdnn.md

This is an example of the output you might see after uploading TensorBoard logs. It includes a link to view your experiment.

```default
TensorFlow installation not found - running with reduced feature set.
Uploading started and will continue reading any new data as it's added to the logdir.

To stop uploading, press Ctrl-C.

New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/yKUbhb5wRmOSXYkId1z9eg/

[2021-08-23T23:49:41] Started scanning logdir.
[2021-08-23T23:49:42] Total uploaded: 135 scalars, 0 tensors, 0 binary objects
Listening for new data in logdir...
```

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

### Install k2 (CPU)

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/for-dummies/environment-setup.md

Installs the k2 library for CPU. A separate command is provided for users in China who might face issues accessing Hugging Face.

```bash
# Please refer to https://k2-fsa.github.io/k2/installation/from_wheels.html#linux-cpu-example

pip install k2==1.24.4.dev20231220+cpu.torch2.0.0 -f https://k2-fsa.github.io/k2/cpu.html

# For users from China
# 中国国内用户，如果访问不了 huggingface, 请使用
# pip install k2==1.24.4.dev20231220+cpu.torch2.0.0 -f https://k2-fsa.github.io/k2/cpu-cn.html
```

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

### Install sherpa-onnx

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/TTS/ljspeech/vits.md

Install the sherpa-onnx package, which provides tools for running ONNX models.

```bash
pip install sherpa-onnx
```

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

### TensorBoard Upload Output Example

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/aishell/tdnn_lstm_ctc.md

This is an example of the output you might see when uploading TensorBoard logs. It includes experiment creation and upload progress.

```default
TensorFlow installation not found - running with reduced feature set.
Upload started and will continue reading any new data as it's added to the logdir.

To stop uploading, press Ctrl-C.

New experiment created. View your TensorBoard at: https://tensorboard.dev/experiment/LJI9MWUORLOw3jkdhxwk8A/

[2021-09-13T11:59:23] Started scanning logdir.
[2021-09-13T11:59:24] Total uploaded: 4454 scalars, 0 tensors, 0 binary objects
Listening for new data in logdir...

```

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

### Install Git LFS

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/aishell/tdnn_lstm_ctc.md

Install git-lfs using apt-get if you have sudo permissions. This is required for downloading language models from Hugging Face.

```bash
sudo apt-get install git-lfs
git-lfs install
```

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

### Install k2 from Wheels

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/installation/index.md

Installs the k2 library using pre-compiled wheels, specifying the CUDA and PyTorch versions. A fallback URL for users in China is provided.

```bash
(test-icefall) kuangfangjun:~$ pip install k2==1.24.3.dev20230725+cuda11.6.torch1.13.0 -f https://k2-fsa.github.io/k2/cuda.html
# For users from China
# 中国国内用户，如果访问不了 huggingface, 请使用
# pip install k2==1.24.3.dev20230725+cuda11.6.torch1.13.0 -f https://k2-fsa.github.io/k2/cuda-cn.html

Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
Looking in links: https://k2-fsa.github.io/k2/cuda.html
Collecting k2==1.24.3.dev20230725+cuda11.6.torch1.13.0
  Downloading https://huggingface.co/csukuangfj/k2/resolve/main/ubuntu-cuda/k2-1.24.3.dev20230725%2Bcuda11.6.torch1.13.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (104.3 MB)
     ________________________________________ 104.3/104.3 MB 5.1 MB/s eta 0:00:00
Requirement already satisfied: torch==1.13.0 in /star-fj/fangjun/test-icefall/lib/python3.8/site-packages (from k2==1.24.3.dev20230725+cuda11.6.torch1.13.0) (1.13.0+cu116)
Collecting graphviz
  Using cached https://pypi.tuna.tsinghua.edu.cn/packages/de/5e/fcbb22c68208d39edff467809d06c9d81d7d27426460ebc598e55130c1aa/graphviz-0.20.1-py3-none-any.whl (47 kB)
Requirement already satisfied: typing-extensions in /star-fj/fangjun/test-icefall/lib/python3.8/site-packages (from torch==1.13.0->k2==1.24.3.dev20230725+cuda11.6.torch1.13.0) (4.7.1)
Installing collected packages: graphviz, k2
Successfully installed graphviz-0.20.1 k2-1.24.3.dev20230725+cuda11.6.torch1.13.0
```

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

### Simulated Streaming Decoding Example (Epoch-based)

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Streaming-ASR/librispeech/pruned_transducer_stateless.md

Example command for performing simulated streaming decoding using epoch-based checkpoints. Requires setting `--simulate-streaming` and `--causal-convolution` to 1.

```bash
for m in greedy_search fast_beam_search modified_beam_search; do
  for epoch in 25 20; do
    for avg in 7 5 3 1; do
      ./pruned_transducer_stateless4/decode.py \
        --epoch $epoch \
        --avg $avg \
        --simulate-streaming 1 \
        --causal-convolution 1 \
        --decode-chunk-size 16 \
        --left-context 64 \
        --exp-dir pruned_transducer_stateless4/exp \
        --max-duration 600 \
        --decoding-method $m
    done
  done
done
```

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

### Serve Documentation Locally

Source: https://github.com/k2-fsa/icefall/blob/master/docs/README.md

Change to the build HTML directory and start a local HTTP server to view the documentation in a browser.

```bash
cd build/html
python3 -m http.server 8000
```

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

### Example Output of Alignment Script

Source: https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/transducer_stateless/README.md

This is an example of the output produced by the alignment script, showing cut IDs and lists of words with their corresponding start times in seconds.

```text
5694-64029-0022-1998-0
[('THE', '0.20'), ('LEADEN', '0.36'), ('HAIL', '0.72'), ('STORM', '1.00'), ('SWEPT', '1.48'), ('THEM', '1.88'), ('OFF', '2.00'), ('THE', '2.24'), ('FIELD', '2.36'), ('THEY', '3.20'), ('FELL', '3.36'), ('BACK', '3.64'), ('AND', '3.92'), ('RE', '4.04'), ('FORMED', '4.20')]

3081-166546-0040-308-0
[('IN', '0.32'), ('OLDEN', '0.60'), ('DAYS', '1.00'), ('THEY', '1.40'), ('WOULD', '1.56'), ('HAVE', '1.76'), ('SAID', '1.92'), ('STRUCK', '2.60'), ('BY', '3.16'), ('A', '3.36'), ('BOLT', '3.44'), ('FROM', '3.84'), ('HEAVEN', '4.04')]

2035-147960-0016-1283-0
[('A', '0.44'), ('SNAKE', '0.52'), ('OF', '0.84'), ('HIS', '0.96'), ('SIZE', '1.12'), ('IN', '1.60'), ('FIGHTING', '1.72'), ('TRIM', '2.12'), ('WOULD', '2.56'), ('BE', '2.76'), ('MORE', '2.88'), ('THAN', '3.08'), ('ANY', '3.28'), ('BOY', '3.56'), ('COULD', '3.88'), ('HANDLE', '4.04')]

2428-83699-0020-1734-0
[('WHEN', '0.28'), ('THE', '0.48'), ('TRAP', '0.60'), ('DID', '0.88'), ('APPEAR', '1.08'), ('IT', '1.80'), ('LOOKED', '1.96'), ('TO',
'2.24'), ('ME', '2.36'), ('UNCOMMONLY', '2.52'), ('LIKE', '3.16'), ('AN', '3.40'), ('OPEN', '3.56'), ('SPRING', '3.92'), ('CART', '4.28')]

8297-275154-0026-2108-0
[('LET', '0.44'), ('ME', '0.72'), ('REST', '0.92'), ('A', '1.32'), ('LITTLE', '1.40'), ('HE', '1.80'), ('PLEADED', '2.00'), ('IF', '3.04'), ("I'M", '3.28'), ('NOT', '3.52'), ('IN', '3.76'), ('THE', '3.88'), ('WAY', '4.00')]

652-129742-0007-1002-0
[('SURROUND', '0.28'), ('WITH', '0.80'), ('A', '0.92'), ('GARNISH', '1.00'), ('OF', '1.44'), ('COOKED', '1.56'), ('AND', '1.88'), ('DICED', '4.16'), ('CARROTS', '4.28'), ('TURNIPS', '4.44'), ('GREEN', '4.60'), ('PEAS', '4.72')]
```

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

### Prepare Data for Yesno Recipe

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/installation/index.md

Set up the environment and navigate to the yesno recipe directory to prepare the dataset.

```bash
(test-icefall) kuangfangjun:icefall$ export PYTHONPATH=/tmp/icefall:$PYTHONPATH

(test-icefall) kuangfangjun:icefall$ cd /tmp/icefall

(test-icefall) kuangfangjun:icefall$ cd egs/yesno/ASR

(test-icefall) kuangfangjun:ASR$ ./prepare.sh
```

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

### HTTP Server Output

Source: https://github.com/k2-fsa/icefall/blob/master/docs/README.md

Example output when starting the Python HTTP server, indicating the address and port it is serving on.

```text
Serving HTTP on 0.0.0.0 port 8000 (http://0.0.0.0:8000/) ...
```

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

### Setup Experiment Directory and Checkpoint Link

Source: https://github.com/k2-fsa/icefall/blob/master/egs/speech_llm/ASR_LLM/RESULTS.md

Create a directory for ASR LLM experiments and create a symbolic link to a specific checkpoint file.

```bash
mkdir -p whisper_llm_zh/exp_aishell_whisper_qwen2_1.5B
ln -s models/checkpoint/epoch-10-avg-5.pt whisper_llm_zh/exp_aishell_whisper_qwen2_1.5B/epoch-999.pt
```

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

### Download and Setup Icefall Repository

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/installation/index.md

Clones the Icefall repository, installs its requirements, and sets the PYTHONPATH environment variable to include the Icefall directory.

```bash
cd /tmp
git clone https://github.com/k2-fsa/icefall
cd icefall
pip install -r requirements.txt
export PYTHONPATH=/tmp/icefall:$PYTHONPATH
```

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

### Distributed Data Parallel Training

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/aishell/stateless_transducer.md

Configure and start distributed training using multiple GPUs. This example uses GPUs 0 and 3 for training with a world size of 2.

```bash
#!/bin/bash
cd egs/aishell/ASR
export CUDA_VISIBLE_DEVICES="0,3"
./transducer_stateless_modified/train.py --world-size 2
```

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

### Start Fine-tuning with Adapters

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Finetune/adapter/finetune_adapter.md

Initiates a fine-tuning experiment for a Zipformer model using adapters. This command configures parameters like adapter dimension, learning rate, and the pre-trained checkpoint to use.

```bash
$ do_finetune=1
$ use_adapters=1
$ adapter_dim=8

$ ./zipformer_adapter/train.py \
    --world-size 2 \
    --num-epochs 20 \
    --start-epoch 1 \
    --exp-dir zipformer_adapter/exp_giga_finetune_adapters${use_adapters}_adapter_dim${adapter_dim} \
    --use-fp16 1 \
    --base-lr 0.045 \
    --use-adapters $use_adapters --adapter-dim $adapter_dim \
    --bpe-model data/lang_bpe_500/bpe.model \
    --do-finetune $do_finetune \
    --master-port 13022 \
    --finetune-ckpt icefall-asr-librispeech-zipformer-2023-05-15/exp/pretrained.pt \
    --max-duration 1000
```

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

### Build Documentation Locally

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/contributing/doc.md

After writing documentation, navigate to the docs directory and use the make command to build the HTML output.

```bash
cd docs
make html
```

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

### Train Large-Scale Librispeech ASR Model

Source: https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/RESULTS.md

Command to initiate training for a large-scale Librispeech ASR model. This example shows the setup for using 2 A100 GPUs and exporting the CUDA visible devices.

```bash
export CUDA_VISIBLE_DEVICES="0,1"

```

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

### Example Data Cut JSON

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/for-dummies/data-preparation.md

This JSON object represents a single audio cut, including its ID, start and end times, channel information, supervision details (text, language), and feature extraction parameters. It also contains information about the original recording.

```json
{"id": "0_0_0_1_0_1_1_0-1", "start": 0, "duration": 6.11, "channel": 0, "supervisions": [{"id": "0_0_0_1_0_1_1_0", "recording_id": "0_0_0_1_0_1_1_0", "start": 0.0, "duration": 6.11, "channel": 0, "text": "NO NO NO YES NO YES YES NO", "language": "Hebrew"}], "features": {"type": "kaldi-fbank", "num_frames": 611, "num_features": 23, "frame_shift": 0.01, "sampling_rate": 8000, "start": 0, "duration": 6.11, "storage_type": "lilcom_chunky", "storage_path": "data/fbank/yesno_feats_train.lca", "storage_key": "16570,12964,2929", "channels": 0}, "recording": {"id": "0_0_0_1_0_1_1_0", "sources": [{"type": "file", "channels": [0], "source": "/tmp/icefall/egs/yesno/ASR/download/waves_yesno/0_0_0_1_0_1_1_0.wav"}], "sampling_rate": 8000, "num_samples": 48880, "duration": 6.11, "channel_ids": [0]}, "type": "MonoCut"}
```

```json
{"id": "0_0_1_0_0_1_1_0-2", "start": 0, "duration": 6.02, "channel": 0, "supervisions": [{"id": "0_0_1_0_0_1_1_0", "recording_id": "0_0_1_0_0_1_1_0", "start": 0.0, "duration": 6.02, "channel": 0, "text": "NO NO YES NO NO YES YES NO", "language": "Hebrew"}], "features": {"type": "kaldi-fbank", "num_frames": 602, "num_features": 23, "frame_shift": 0.01, "sampling_rate": 8000, "start": 0, "duration": 6.02, "storage_type": "lilcom_chunky", "storage_path": "data/fbank/yesno_feats_train.lca", "storage_key": "32463,12936,2696", "channels": 0}, "recording": {"id": "0_0_1_0_0_1_1_0", "sources": [{"type": "file", "channels": [0], "source": "/tmp/icefall/egs/yesno/ASR/download/waves_yesno/0_0_1_0_0_1_1_0.wav"}], "sampling_rate": 8000, "num_samples": 48160, "duration": 6.02, "channel_ids": [0]}, "type": "MonoCut"}
```

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

### Start Training

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/for-dummies/training.md

This command initiates the training process. Ensure you are in the correct directory and have set up the Python path. Training will use CPU if CUDA_VISIBLE_DEVICES is empty.

```bash
cd /tmp/icefall
export PYTHONPATH=/tmp/icefall:$PYTHONPATH
cd egs/yesno/ASR

# We use CPU for training by setting the following environment variable
export CUDA_VISIBLE_DEVICES=""

./tdnn/train.py
```

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

### Install lhotse

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/installation/index.md

Use pip to install the lhotse library. This command installs lhotse and its dependencies.

```bash
pip install lhotse
```

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

### Verify CUDA Installation

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/installation/index.md

Checks the installed CUDA version using the 'nvcc' command. This is useful after installing the CUDA toolkit.

```bash
nvcc --version
```

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

### Install Whisper Requirements

Source: https://github.com/k2-fsa/icefall/blob/master/egs/multi_zh-hans/ASR/RESULTS.md

Installs the necessary Python packages for the Whisper model training. Ensure you have the correct version of pip installed.

```bash
pip install -r whisper/requirements.txt
```

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

### Start Training with 4 GPUs

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/librispeech/zipformer_ctc_blankskip.md

Initiates the training process using 4 GPUs. This command sets up the environment and specifies training parameters.

```bash
export CUDA_VISIBLE_DEVICES="0,1,2,3"
./pruned_transducer_stateless7_ctc_bs/train.py \
  --world-size 4 \
  --num-epochs 30 \
  --start-epoch 1 \
  --full-libri 1 \
  --exp-dir pruned_transducer_stateless7_ctc_bs/exp \
  --max-duration 600 \
  --use-fp16 1
```

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

### Verify Torchaudio Installation

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/installation/index.md

Confirms that torchaudio has been installed successfully by displaying its version.

```bash
(test-icefall) kuangfangjun:~$ python3 -c "import torchaudio; print(torchaudio.__version__)"

0.13.0+cu116
```

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

### Display Training Options

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/timit/tdnn_ligru_ctc.md

View all available command-line options for the training script by running it with the --help flag.

```bash
./tdnn_ligru_ctc/train.py --help
```

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

### Install Lhotse

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/for-dummies/environment-setup.md

Installs the latest version of the lhotse library from its GitHub repository.

```bash
pip install git+https://github.com/lhotse-speech/lhotse
```

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

### Install PyTorch and Torchaudio (CPU)

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/for-dummies/environment-setup.md

Installs specific versions of PyTorch and torchaudio for CPU usage. This order of installation is important. For macOS users, a different command without the '+cpu' suffix is provided.

```bash
# We use torch 2.0.0 and torchaduio 2.0.0 in this tutorial.
# Other versions should also work.

pip install torch==2.0.0+cpu torchaudio==2.0.0+cpu -f https://download.pytorch.org/whl/torch_stable.html

# If you are using macOS, please use the following command to install torch and torchaudio
# pip install torch==2.0.0 torchaudio==2.0.0 -f https://download.pytorch.org/whl/torch_stable.html
```

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

### View TensorBoard Logs Locally

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/librispeech/zipformer_ctc_blankskip.md

Starts a local TensorBoard server to view training logs. Navigate to the tensorboard directory first.

```bash
cd pruned_transducer_stateless7_ctc_bs/exp/tensorboard
tensorboard --logdir . --port 6008
```

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

### Install CTC Dependencies

Source: https://github.com/k2-fsa/icefall/blob/master/icefall/ctc/README.md

Installs the necessary Python packages for CTC decoding in Icefall.

```bash
pip install kaldifst kaldi-decoder
```

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

### Get Training Help

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/librispeech/tdnn_lstm_ctc.md

Display available training options for the train.py script using the --help flag.

```bash
./tdnn_lstm_ctc/train.py --help
```

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

### Start Training with 6 GPUs

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/librispeech/pruned_transducer_stateless.md

Initiates the training process using 6 GPUs, specifying epochs, experiment directory, and maximum duration.

```bash
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5"
./pruned_transducer_stateless4/train.py \
   --world-size 6 \
   --num-epochs 30 \
   --start-epoch 1 \
   --exp-dir pruned_transducer_stateless4/exp \
   --full-libri 1 \
   --max-duration 300
```

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

### Verify sherpa-onnx Installation

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/TTS/ljspeech/vits.md

Verify that sherpa-onnx has been installed correctly by checking its command-line tool.

```bash
which sherpa-onnx-offline-tts
sherpa-onnx-offline-tts --help
```

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

### Invoking the Training Stage

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/contributing/how-to-create-a-recipe.md

Shows how to run the training script for a custom recipe, including how to access help for available options.

```bash
$ cd egs/foo/ASR
$ ./bar/train.py
$ ./bar/train.py --help
```

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

### Verify PyTorch Installation

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/installation/index.md

Checks if PyTorch has been installed correctly by printing its version number.

```bash
(test-icefall) kuangfangjun:~$ python3 -c "import torch; print(torch.__version__)"

1.13.0+cu116
```

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

### Prepare for MVQ Training

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/librispeech/distillation.md

This command initiates the preparation for MVQ (Multi-Vector Quantization) training by running stage 3. It's a prerequisite for the actual knowledge distillation training.

```bash
./prepare.sh --stage 3 --stop_stage 3
```

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

### Install ONNX

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/model-export/export-onnx.md

Install the ONNX package using pip. This is a prerequisite for exporting models.

```bash
pip install onnx
```

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

### Install Dependencies

Source: https://github.com/k2-fsa/icefall/blob/master/egs/wenetspeech4tts/TTS/README.md

Installs necessary Python packages for kaldialign, lhotse, tensorboard, bigvganinference, and sentencepiece.

```bash
pip install kaldialign lhotse tensorboard bigvganinference sentencepiece
```

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

### Start Stateless Transducer Training

Source: https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/transducer_stateless/README.md

Use this command to initiate the training process for a stateless transducer ASR model. Ensure you are in the correct directory and have set the CUDA visible devices.

```bash
cd egs/librispeech/ASR

export CUDA_VISIBLE_DEVICES="0,1,2,3"

./transducer_stateless/train.py \
  --world-size 4 \
  --num-epochs 30 \
  --start-epoch 0 \
  --exp-dir transducer_stateless/exp \
  --full-libri 1 \
  --max-duration 250 \
  --lr-factor 2.5
```

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

### Install Icefall Requirements

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/for-dummies/environment-setup.md

Installs dependencies required by icefall using the provided requirements file.

```bash
pip install -r ./requirements.txt
```

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

### Install Extra Dependencies

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/TTS/ljspeech/vits.md

Install necessary packages for TTS recipes, including piper_phonemize, numba, and espnet_tts_frontend.

```bash
pip install piper_phonemize -f https://k2-fsa.github.io/icefall/piper_phonemize.html
pip install numba espnet_tts_frontend
```

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

### Display Training Options

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/aishell/stateless_transducer.md

Use this command to view all available training options for the stateless transducer model.

```bash
cd egs/aishell/ASR
./transducer_stateless_modified/train.py --help
```

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

### Display Training Options

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/librispeech/pruned_transducer_stateless.md

Shows the available training options that can be passed from the command line. Use this to understand configurable parameters.

```bash
cd egs/librispeech/ASR
./pruned_transducer_stateless4/train.py --help
```

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

### Verify lhotse Installation

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/installation/index.md

Run this Python command to confirm that lhotse has been installed successfully and to display its version.

```bash
(test-icefall) kuangfangjun:~$ python3 -c "import lhotse; print(lhotse.__version__)"

1.16.0.dev+git.7640d66.clean
```

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

### Display Training Help

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/for-dummies/training.md

Prints the usage information for the training script, allowing you to see available options for customization.

```bash
./tdnn/train.py --help
```

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

### Install optimized_transducer

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/aishell/stateless_transducer.md

Install the optimized_transducer library using pip. This library is used for computing the transducer loss efficiently.

```bash
pip install optimized_transducer
```

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

### Download Qwen Model

Source: https://github.com/k2-fsa/icefall/blob/master/egs/speech_llm/ASR_LLM/RESULTS.md

Use huggingface-cli to download the Qwen2-7B-Instruct model to a local directory.

```bash
huggingface-cli download --local-dir models/qwen Qwen/Qwen2-7B-Instruct
```

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

### Multi-GPU Training with All Available GPUs

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/librispeech/zipformer_ctc_blankskip.md

Utilize all available GPUs for DDP training by setting the --world-size argument to the number of GPUs. This example assumes 4 GPUs are available.

```bash
cd egs/librispeech/ASR
./pruned_transducer_stateless7_ctc_bs/train.py --world-size 4
```

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

### Install Git Pre-Commit Hooks

Source: https://github.com/k2-fsa/icefall/blob/master/contributing.md

Install the Git pre-commit hooks for the project. This enables automatic code checks before commits.

```bash
pre-commit install
```

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

### Start Decoding

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/for-dummies/decoding.md

This command initiates the decoding process. Ensure you are in the correct directory and have set up the Python path. Decoding is performed on CPU by default.

```bash
cd /tmp/icefall
export PYTHONPATH=/tmp/icefall:$PYTHONPATH
cd egs/yesno/ASR

# We use CPU for decoding by setting the following environment variable
export CUDA_VISIBLE_DEVICES=""

./tdnn/decode.py
```

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

### Start DDP Training on Specific GPUs

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/aishell/conformer_ctc.md

Configure and start Distributed Data Parallel (DDP) training on specified GPUs.

```bash
#!/bin/bash
cd egs/aishell/ASR
export CUDA_VISIBLE_DEVICES="0,3"
./conformer_ctc/train.py --world-size 2
```

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

### Start Stateless RNN-Transducer Training

Source: https://github.com/k2-fsa/icefall/blob/master/egs/aishell/ASR/transducer_stateless/README.md

Use this command to begin training the stateless RNN-Transducer model. Ensure CUDA devices are configured and adjust parameters like world size, epochs, and experiment directory as needed.

```bash
cd egs/aishell/ASR

export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"

./transducer_stateless/train.py \
  --world-size 8 \
  --num-epochs 30 \
  --start-epoch 0 \
  --exp-dir transducer_stateless/exp \
  --max-duration 250 \
  --lr-factor 2.5

```

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

### View TensorBoard Logs Locally

Source: https://github.com/k2-fsa/icefall/blob/master/docs/source/recipes/Non-streaming-ASR/librispeech/pruned_transducer_stateless.md

Starts a local TensorBoard server to view training logs. Access it via the provided URL in your browser.

```bash
#!/bin/bash
cd pruned_transducer_stateless4/exp/tensorboard
tensorboard --logdir . --port 6008
```

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

### Prepare Data and Features

Source: https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/WSASR/README.md

Prepares the dataset by adding an OTC token and computing SSL features. Ensure 'feature_type' is set to 'ssl' or 'fbank'.

```bash
# feature_type can be ssl or fbank
feature_type=ssl
feature_dir="data/${feature_type}"
manifest_dir="${feature_dir}"
lang_dir="data/lang"
lm_dir="data/lm"
exp_dir="conformer_ctc2/exp"
otc_token="<star>"

./prepare.sh \
  --feature-type "${feature_type}" \
  --feature-dir "${feature_dir}" \
  --lang-dir "${lang_dir}" \
  --lm-dir "${lm_dir}" \
  --otc-token "${otc_token}"
```