Skip to content

Triton 🐚

Triton is an audio signal processing toolkit for speech-in-noise research in the CONCH Lab.

Triton is moving to a project-first workflow: you create/open a project, define canonical audio settings, add files into that project, and run processing inside that boundary.

Core

Dependency-light math for SNR mixing, babble generation, vocoding, and filtering.

CLI

Batch processing for large datasets with reproducible settings. Label and organize files for cleaner workflows.

GUI

Project dashboard for create/open, file management with labeling, and interactive mixing.

Key Features

Pipelines & Reproducibility: Define processing pipelines in your project config and run them on any file with consistent settings. Use Pipeline Matrix to sweep across multiple files and parameter combinations in a single batch.

Project-first workflow: Centralized settings, file management, and reproducible outputs.

File Labeling: Organize your project files with custom labels (e.g., bab-f1, bab-m1, background), apply one label to a whole upload batch during import, and filter by label in CLI and GUI for easier asset management.

Babble Speech Generation: Mix labeled talker groups with per-file RMS normalization, concatenation per talker, intended-length planning, and balanced male/female selection when counts are not specified. If a talker is short, files repeat randomly to reach target length. GUI and CLI share the same core babble generator, and generated babble can be added back into the project as a labeled derivative (bab-t#) with provenance metadata.

Quickstart

  • Launch GUI and create/open a project.
  • Add files, label them for organization, and manage project assets.
  • Generate babble from labeled talker groups, or mix at target SNR in GUI or CLI.
  • For batch processing, define a pipeline and run Pipeline Matrix for parameter sweeps.

See the Quickstart for commands.