Voice Capture Studio

Overview

Voice Capture Studio is a local-first browser application for directed voice capture. It helps a speaker move from an empty session to clean, reviewable takes with transcripts, timing, quality metadata and export structure attached.

The application is live at electronicartefacts.github.io/voice-capture-studio and the source code is public at github.com/electronicartefacts/voice-capture-studio.

Purpose

The project exists because professional voice material is often captured before the workflow is ready for it. A recording may sound usable, but the prompt, language, speaker, room, microphone, review state, timing and export manifest can disappear into loose folders.

Voice Capture Studio treats those details as part of the recording. It prepares structured voice material for downstream archives, speech datasets, voice-over review, dubbing and machine-learning workflows.

What It Does Not Do

Voice Capture Studio does not train AI models. It records speech, reviews takes and exports structured material. Any later training, alignment, normalization, archive processing or model evaluation belongs to a separate workflow with its own governance and consent requirements.

That boundary is important: the application prepares source material; it does not claim the authority of a model pipeline.

Recording Modes

The live application exposes three capture modes:

• Dataset ML: prompted recording for calibrated phrases, phonetic progress and training-ready exports.
• Dubbing: pasted text or segmented files transformed into recordable lines.
• Master audio: local text, reference audio and a separate voice capture lane.

The project page uses “dataset” in the preparation sense: accepted recordings and metadata can become useful inputs, but model training remains downstream.

Key Features

• Guided prompt sessions for French and English starter corpora.
• Local microphone capture through the browser.
• WAV PCM mono export at 48 kHz / 24-bit where browser support allows it.
• Room-tone calibration and first-pass technical quality checks.
• Transcript, timing, intent, prosody and quality metadata for each accepted take.
• Keeper, review and reject states so coverage advances only on accepted material.
• Browser-private workspace storage with downloads and folder export where supported.
• PWA manifest and service-worker support for static GitHub Pages distribution.

Architecture Overview

The repository separates the React application shell from domain modules. The app shell coordinates browser capture, storage adapters and export orchestration. The domains own corpus, sessions, workspace, coverage, recording, speakers, settings and export contracts.

The important architectural decision is dependency direction: domain modules do not import React, browser UI, routes or CSS. That keeps recording behavior, corpus planning and export contracts testable without tying them to one interface.

Technology Stack

Voice Capture Studio is built with React, Vite and TypeScript. It uses Web Audio for local capture, browser storage for local workspace state, the File System Access API where available, explicit downloads as a fallback, a PWA manifest and a GitHub Pages deployment workflow.

The validation gate in the repository runs formatting, linting, TypeScript project references, unit tests and the production build.

Design Philosophy

The interface is not a generic recorder. It behaves like a small studio surface: choose a mode, check the environment, confirm the speaker and language, capture calibration material, record takes, then export the session with enough metadata to be reviewed later.

The design keeps technical state visible because local-first software needs trust cues. Storage mode, microphone readiness, WAV support, folder access, corpus version and calibration status are part of the user experience.

Open Source Philosophy

The project is released under the MIT License. Open source matters here because voice workflows should be inspectable. A team preparing voice material needs to know what is stored, what is exported, what is never uploaded and where data boundaries sit.

The repository includes contribution guidelines, security reporting, issue templates, a pull-request template and documentation for architecture, corpus structure, workspace structure, export structure and GitHub Pages.

Privacy

Voice recordings and generated workspaces are user data. The application is designed so recordings remain local to the browser or the folder selected by the user. The public repository should contain code and corpus definitions, not private voice recordings.

Browser storage is useful, but it is not a permanent archive. The app therefore treats explicit export and folder workflows as first-class parts of the product.

Export Model

The primary export target is a structured capture-session folder. The current repository documents session, speaker, corpus and manifest JSON files, take-level WAV and transcript files, timing and intent metadata, quality reports and dataset-readiness reports.

Every accepted take is more than audio. It carries enough structure to support later alignment, review, coverage analysis or archive processing.

Documentation

Useful entry points:

• Architecture doctrine
• System model
• Corpus structure
• Workspace structure
• Export structure

Roadmap

The public roadmap includes workspace restore/import for downloaded backups, explicit workspace schema migrations, corpus tombstones for long-lived compatibility, a stronger first-run folder flow, additional export targets and screenshot/release automation.

Those items are product infrastructure, not decoration. They make the difference between a promising local app and a voice archive workflow that can survive real use.

Related Electronic Artefacts Work

Voice Capture Studio sits near ORETH because both treat audio as structured material. It sits near Web Audio and the Web Audio API because the browser is the runtime. It sits near Open Source because the code and documentation are inspectable.

It also expands the Electronic Artefacts ecosystem into a practical voice-technology artefact: a tool that does one thing clearly before any archive, model, dubbing or production workflow receives the data.