Voice Capture Studio

Overview

Voice Capture Studio is a local-first voice instrument that runs in the browser. It now joins five distinct workflows in one surface: immediate recording, ML corpus production, picture-led dubbing, interpreted performance with a separate guide track, and word-by-word segmentation of local audio or video.

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

Studio line

Voice Capture Studio is the first published application in Studio, Electronic Artefacts’ open-source software line. The connection is a shared operating ethic rather than a shared product shell: specialist tools should be independently usable, local-first where possible, free of advertising, published with inspectable source and able to export durable material into other workflows.

The next planned record, Spatial Mapping Studio, approaches a different material — final 3D payloads and their semantic regions — without changing those commitments. Its V0 work is not an integration with Voice Capture Studio or Forge; it establishes the portable mapping boundary that later workflows can build on.

The experience is part of the instrument

The current interface is deliberately quiet: warm white rather than studio black, large carbon typography, rounded capsules, fine borders, ice-blue focus, dusty-rose warnings and translucent panels that keep the page light while preserving depth. Glassmorphism is used as an information layer, not an ornamental effect: the background stays perceptible while readiness, mode, calibration and export state remain legible.

The main curve — the filament — is the product’s perceptual spine. In the recording application it represents fresh audio, never an arbitrary loop. Its halo derives energy from measured amplitude. During replay, the same visual language follows the decoded take. If a live frame becomes stale, the surface returns to an honest idle carrier instead of presenting old data as current.

That rule extends to the entire motion system: nothing that looks like a reading may be simulated. Silence remains data. Capture receives the full rendering budget. Reduced motion removes choreography without removing meaning. The interface is therefore closer to a pane of glass over an acoustic instrument than to a conventional dashboard.

Five modes, five outcomes

• Free Capture starts directly, without a corpus, and records up to ten minutes with manual stop, WAV and provenance.
• Lexical Segmentation takes local audio or video, discards the image, analyzes speech or song and exports a timestamped WAV for each supported word plus JSON and CSV evidence.
• ML Dataset plans phrases from corpus coverage, calibrates the room, records comparable takes and advances coverage only when accepted material is both valid and durably stored.
• Dubbing combines a local script or SRT/VTT cues with local video or an explicitly activated YouTube reference. The performer sees the scene, preserves cue timing and exports an isolated voice lane.
• Performance records continuous or segmented interpretation against an optional headphone guide. The guide supports the performer but is never mixed into the captured voice WAV.

These are product modes rather than file-format presets. Each changes the inputs, the capture gesture, the definition of success and the useful output. Podcasting remains a Free Capture scenario; voice-over remains a Dubbing scenario; audiobook work remains a Performance scenario until those uses require genuinely different engines.

From signal to decision

The recording path begins with progressive capability inspection: microphone permission and inputs, Web Audio, storage durability, folder access, downloads, wake lock, optional speech APIs, Workers and accelerated rendering. A missing optional capability degrades only the workflow that depends on it.

Microphone capture uses getUserMedia and a PCM recorder. It prefers AudioWorkletNode, keeps a ScriptProcessorNode compatibility fallback, encodes mono RIFF-padded WAV at 48 kHz / 24-bit when the runtime permits, and derives first-pass technical metrics from the recorded signal. Imported media is decoded to 16 kHz analysis signals before local segmentation.

After capture, observations remain separate:

• the immutable corpus declares the expected text, language, intent and delivery;
• the PCM signal supplies physical and prosodic measurements, energy VAD, segments, silences and pauses;
• browser speech recognition, when available, remains an optional hypothesis;
• grapheme-to-phoneme and word timing remain preparatory estimates;
• fused decisions cite their evidence, reasons and confidence rather than erasing the observations that produced them.

This is the deterministic research layer: identical bounded observations feed explicit policies, and uncertainty stays visible. Browser ASR can request review but cannot reject or authorize a physical recording. A transcript mismatch does not erase a technically valid take. An acoustic forced-alignment import can replace an estimated timing map without rewriting the raw evidence.

Lexical segmentation pipeline

Lexical Segmentation is the fifth mode and the clearest example of the local pipeline. It accepts up to 200 MB and ten minutes, decodes the media on the device, classifies the scene and chooses a fast, verified or deep analysis budget. A first Whisper pass scouts the source. Difficult speech, singing or music can trigger a base-model pass, a vocal-focus signal, spectral mid/side separation and temporal fuzzy-majority consensus.

Each exported word keeps its start, end, clip context, acoustic support, confidence, consensus votes and evidence class. The ZIP contains 16 kHz / 24-bit mono WAV clips, a versioned manifest and a timeline. Video pixels are not processed by this mode; only the audio track enters the pipeline.

Local-first persistence and open exports

The workspace prefers IndexedDB, can use a user-selected File System Access folder, and always keeps explicit downloads as a fallback. A take cannot credit dataset coverage when its audio was not accepted by durable local storage. Workspace archives are verified before restoration and do not silently overwrite existing WAV files.

The main capture package contains session, speaker and corpus records; take-level WAV, transcript, timing, phoneme, intent, quality, observation and evidence files; dataset reports; a JSONL training manifest; and SHA-256 checksums. Browser estimates remain labeled forcedAlignmentRequired until acoustic validation replaces or confirms them.

Voice Capture Studio prepares material for archives, research, dubbing and machine-learning workflows. It does not train a model, promise research-grade acoustic alignment in every browser or upload private recordings to a remote service.

Architecture and technology

The app is built with React 19, TypeScript and Vite. Web Audio, AudioWorklet, browser media decoding, IndexedDB, the File System Access API, Workers, WebAssembly/WebGPU-capable local analysis, PWA service-worker distribution and GitHub Pages form the browser chassis.

Domain modules own corpus, sessions, workspace, coverage, recording, observations, phonetics, speakers, settings and export contracts. They do not import React, CSS, DOM or browser APIs. The application shell owns orchestration and adapters. This dependency direction keeps planning, keeper rules, evidence fusion and package contracts testable outside a specific interface.

The repository release gate covers formatting, linting, TypeScript project references, unit coverage, bundle budgets and production build. End-to-end suites exercise capture, responsive layouts, Chromium, Firefox, WebKit, offline restart and automated accessibility checks.

Open source, privacy and boundaries

The MIT-licensed repository makes the behavior inspectable. Voice recordings and generated workspaces remain user data; they belong in browser storage, a chosen local folder or an explicit download, never in the public source repository.

Local-first is not a claim that browser storage is permanent. The interface exposes storage durability, capability gaps and export paths because users need to know what will survive a reload, what requires a download and what a particular browser cannot provide.

Documentation and evidence

Useful source entry points:

• Phenomenology — signal, movement and perceptual laws.
• Mode experience grid — users, inputs, gestures, outputs and success criteria for the five modes.
• System model — observed runtime flow and current tensions.
• Capture technology audit — observation pipeline, timing limits and validation boundary.
• Export structure — package layout, quality records and Forge-shaped downstream stages.
• Browser-native engineering manifesto — progressive capability and local-first principles.

Knowledge relay

The six editorial briefs are now complete Knowledge articles rather than release notes: trustworthy audio interfaces, traceable voice capture, deterministic evidence fusion, local lexical segmentation, five mode contracts and browser-native local-first architecture. Each remains useful after interface changes because it addresses a durable problem.

Voice Capture Studio sits near ORETH because both treat audio as structured material. It connects to Web Audio, the Web Audio API, Human–Computer Interaction, Metadata and Open Source as a practical test of those ideas.

The project’s role is precise: make the first voice artifact cleaner, more inspectable and more reusable before an archive, model, dubbing edit or production workflow receives it.