quick-convert documentation

quick-convert is a modular framework for building speech privacy experiments from reusable datasets, systems, components, and executable pipelines.

Experiments are configured through Hydra, allowing architectures, datasets, feature extractors, trainers, and evaluation systems to be composed without rewriting the surrounding infrastructure.

Start here

New users should begin with the Quickstart.

The Quickstart starts from a downloaded copy of LibriSpeech and walks through a complete workflow:

1. training a SentencePiece tokenizer;
2. precomputing token IDs;
3. building a CSV manifest;
4. training a VQ-ASR model from that manifest.

It also introduces the main abstractions used throughout the project, including datasets, resources, pipelines, trainers, and Hydra configuration composition.

Begin the Quickstart →

Core concepts

quick-convert is organized around three conceptual layers:

Pipelines
    │
    ▼
Systems
    │
    ▼
Components

⚠️ WARNING: Most of the following pages don’t exist yet. ⚠️

Pipelines

Pipelines define complete executable workflows, such as:

• training;
• evaluation;
• anonymization;
• feature precomputation;
• manifest generation.

Pipelines coordinate configuration, data loading, execution, and output handling.

Systems

Systems implement task-level capabilities, such as:

• automatic speech recognition;
• automatic speaker verification;
• speech anonymization and voice conversion.

A system is generally independent of the exact architecture used to implement it.

Components

Components are reusable model and signal-processing building blocks. They are analogous to PyTorch nn.Module objects and may be composed recursively.

Examples include:

• encoders and decoders;
• self-supervised speech models;
• feature extractors;
• speaker embedding models;
• neural network layers;
• losses.

Documentation

Using the framework

• Installation
• Quickstart
• Configuration and Hydra
• Running pipelines

Data

• Datasets
• Resources
• Manifest datasets
• Dataloading

Pipelines

• Training
• Evaluation
• Feature precomputation
• Building manifests
• Anonymization

Systems

• Automatic speech recognition
• Automatic speaker verification
• Anonymization and voice conversion

Components

• Encoders and decoders
• Self-supervised models
• Feature extractors
• Quantizers
• Losses

Development: TODO

• Repository structure
• Adding a dataset
• Adding a component
• Adding a pipeline
• Contributing

How experiments are assembled

Most experiments begin with a run configuration under configs/run/.

A run configuration selects and combines the relevant configuration groups:

defaults:
  - /global: default
  - /pipeline: training
  - /trainer: vq_asr
  - /dataset@train_dataset: manifest
  - _self_

Hydra composes these files into a complete runtime configuration. The selected pipeline is then instantiated and executed.

Run configuration
       │
       ▼
Hydra composition
       │
       ▼
    Pipeline
       │
       ▼
     System
       │
       ▼
   Components
       │
       ▼
     Outputs

Project status

quick-convert is an active research codebase.

Some parts of the package structure reflect earlier iterations of the design. In particular, several task-level systems currently live beneath quick_convert/pipelines/; these are expected to move gradually into quick_convert/systems/.

Optional dependency groups are also being revised and tested. Consult pyproject.toml when installing dependencies for a specific workflow.