Synthetic Data Pipeline - Sound Instruct

1. Introduction

The Synthetic Data Pipeline is designed to process large raw-text dataset into trainable sound. This pipeline leverages:

• First, WhisperSpeech to convert text into speech.
• Second, Encodec to convert into trainable sound tokens.

However, as WhisperSpeech and Encodec do not support batch processing, this pipeline is designed to process data in parallel across multiple GPUs in multiple processes.

2. Features

• Multi-GPU and multi-process for each GPU
• Text-to-speech conversion using WhisperSpeech
• Audio tokenization using EncodecModel
• CSV output for easy integration and recovery
• Configurable batch processing
• Progress tracking and logging

3. Requirements

3.1 Hardware

• Multiple CUDA-capable GPUs (tested with 4 GPUs)
• Sufficient RAM to handle large datasets
• Fast storage for input/output operations

3.2 Dependencies

• torch
• pyarrow
• datasets
• WhisperSpeech
• encodec
• Dependencies listed in requirements.txt

4. Installation

1. Move to the directory:

   cd synthetic-data
   

2. Install the required packages:

   pip install -r requirements.txt
   

5. Project Structure

synthetic-data-pipeline/
├── synthetic_data_pipeline.py
├── audio_tokenizer.py
├── tts_processor.py
├── requirements.txt
├── README.md
└── <save_dir>/
    ├── audio_tokens_*.csv
    └── failed_indices_*.json

6. Configuration

Edit the synthetic_generation_cfg.yaml file to configure the following parameters:

# Dataset configuration
dataset:
  name: # Dataset name from Hugging Face
  split: # Dataset split to use
  remaining_indices_file: # File to store remaining indices (List[int])

# Processing configuration
processing:
  devices: # List of GPUs to use
  num_procs_per_device: # Number of processes per GPU
  save_dir: # Directory to save processed data
  save_batch_size: # Batch size for saving processed data for each process
  max_retries: # Maximum number of retries for processing a sample
  sample_rate: # Sample rate for audio data
  speaker: # Speaker class to use for speech generation
  format: # Format for saving processed data

test_mode: # Whether to run the test
test:
  num_samples: # Number of samples to process
  devices: # List of GPUs to use
  num_procs_per_device: # Number of processes per GPU
  save_dir: # Directory to save processed data
  save_batch_size: # Batch size for saving processed data for each process
  max_retries: # Maximum number of retries for processing a sample
  sample_rate: # Sample rate for audio data
  speaker: # Speaker class to use for speech generation

# Logging configuration
logging:
  log_file: # Log file name
  console_level: # Console log level
  file_level: # File log level

upload_to_s3: # Whether to upload processed data to S3
s3:
  save_dir: # Directory to save processed data
  bucket_name: # S3 bucket name
  s3_folder: # S3 folder name

6.1 Speaker

There are currently 3 supported speakers for text-to-speech conversion:

• default_speaker
• speaker_5304

6.2 Format

There are currently 2 supported speakers for text-to-speech conversion:

• parquet
• csv

7. Usage

7.1 Running the Pipeline

To run the full pipeline:

python synthetic_data_pipeline.py --config_file synthetic_generation_cfg.yaml

7.2 Test Mode

To run the pipeline in test mode with a smaller dataset. modify the test_mode parameter in the configuration file:

...
test_mode: true
...

8. Pipeline Components

8.1 Text-to-Speech Processor

The TTSProcessor class uses WhisperSpeech to convert text prompts into audio waveforms. It's initialized with a specific device (GPU) and can generate audio from text inputs.

8.2 Audio Tokenizer

The AudioTokenizer class uses EncodecModel to process audio waveforms into tokenized representations. It can also decode audio tokens back into waveforms if needed.

9. Data Flow

graph TD
    A[Load Dataset] --> B[Split into Chunks]
    B --> C[Process Chunk]
    
    subgraph PROC ["Process x N"]
        C --> D[Text to Speech Conversion]
        D --> E[Audio Tokenization]
        E --> F[Save to CSV]
        C --> G[Error Handling]
        G -->|Retry| D
        G -->|Max Retries Exceeded| H[Log Failed Indices]
    end
    
    F --> I[Combine Results]
    H --> J[Aggregate Failed Indices]

    classDef PROC padding-left:23em;
    class PROC PROC;

Loading

10. Acknowledgements

This pipeline is built on top of the WhisperSpeech and Encodec libraries. We would like to thank the developers of these libraries for their contributions to the field of audio processing.