Braat shows you, in real time, the pitch and resonance of your voice as you speak. It's a practice aid aimed at voice training — including trans voice training, where pitch (F0) and vowel resonance (F1-3) are commonly used reference points. It may also be useful for singing practice, accent work, or just exploring how your voice works.

Open Braat →

Braat's signal-processing algorithms are adapted from Praat, a widely used tool in phonetics research. Praat is primarily oriented toward offline analysis; Braat takes its algorithms and runs them on a live microphone in the browser, with no install and no upload.

• Live spectrogram — your voice's frequency content as you speak
• Pitch (F0) tracking — Praat's filtered autocorrelation method
• Formant (F1-F3) tracking — Burg's method LPC, plotted on a vowel chart
• Voice activity detection — Silero VAD picks out voiced segments
• File import — drop in a recording and analyze it offline
• Private — audio stays in your browser; nothing is uploaded
• Offline — once loaded, works without a network connection

Braat uses AudioWorklet, Web Workers, and SharedArrayBuffer, which need a recent browser (Chrome, Firefox, or Safari).

Braat is alpha and in active development. Core features work, but expect changes to the UI and algorithms. A usage guide is planned but not yet written.

All audio processing happens in the browser. An AudioWorklet captures microphone PCM into a SharedArrayBuffer ring, and three Web Workers read from it in parallel:

• a spectrogram worker (FFT-based, with a Bark-scaled colormap),
• a formant worker running pitch (filtered autocorrelation) and formant (Burg LPC) analysis,
• and a VAD worker running Silero v6 via ONNX Runtime Web.

DSP code ported from Praat via LLM is attributed in each source file's copyright header.

# Development
npm run dev            # Run dev server on port 3000
npm run build          # Build for production

# Code Quality
# Prefer `npm run check` over running the underlying tools (oxlint, oxfmt)
# directly: it bundles format + lint-fix into one step.
# Note: `npm run check` or oxlint is how you typecheck, tsgo isn't installed directly.
npm run check

# Testing
npm run test           # Run tests with Vitest
npm run e2e            # Run slow end-to-end tests
npm run test -- --silent=false --disable-console-intercept  # Show logs

# Reference media
npm run media:fetch    # Mirror reference clips into media/references/ (see below)

• UI Framework: TanStack Router (React 19 SPA with file-based routing under src/routes/)
• Styling: Tailwind CSS v4 (via @tailwindcss/vite plugin)
• Components: shadcn/ui with Base UI (not Radix), configured via components.json ("style": "base-nova"). Components are vendored into src/components/ui/ with minor local changes on top.
• Icons: lucide-react

1. Real-time Priority: Spectrogram, waveform, and formant data must remain responsive. Slower computations should not block visualization. If necessary, defer or make features optional rather than blocking the UI.

2. Audio Worklet for Low-Latency: Realtime DSP runs in an AudioWorklet; UI and DSP communicate via message passing (not direct function calls).

3. Vendored DSP Code: Algorithms are ported to TypeScript from reputable sources (primarily Praat) with clear attribution. Avoid WebAssembly when reasonable TypeScript alternatives exist.

4. Stream & Batch Processing: Each algorithm should provide:

   • A stream wrapper (avoids array allocations during processing when possible)
   • A batch wrapper (for offline/file import processing)

5. Browser-Only: All processing runs in the browser. The server only builds and serves static assets.

To add new shadcn/ui components (pulled from the Base UI registry configured in components.json):

npx shadcn@latest add <component-name>

Components are vendored into src/components/ui/ and can be imported directly. We carry minor local changes on top of the generated code, so review diffs before re-adding or updating a component.

The practice route plays ~140 MB of synthesized reference clips (per-sentence MP3s). These are not in the repo. media.braat.app holds them, along with manifest.json.

• Generating clips: npm run synth:references, very slow
• Fetching clips: npm run media:fetch
• Local development: flip USE_LOCAL_MEDIA to use local references
• URL resolution: see src/lib/mediaConfig.ts
• Host headers: media/_headers (tracked) sets cross-origin isolation headers Grebedoc adds Content-Type, caching, and ACAO automatically.

Anonymous, cookieless usage stats via GoatCounter, via src/lib/analytics.ts (trackPageview / track). Events have no properties, so dimensions are encoded into the event name (family/value) from a small, fixed set. See src/routes/privacy.tsx for the user-facing disclosure. Do Not Track / GPC are intentionally ignored. It's not really aiming to stop first-party, anonymous counts. See https://www.arp242.net/dnt.html.

CI runs on Forgejo (.forgejo/workflows/ci.yaml). The workflow lints, tests, builds, and deploys to Grebedoc on every push to main.

Forgejo supports GitHub Actions syntax, but compatibility with third-party marketplace actions is not guaranteed. Prefer runner-agnostic shell steps where possible.

You can, but do not need to, use just and nix for development. Common tasks are wrapped as just recipes (run just to list them). You can use VS Code / Codespaces to open a devcontainer, use direnv + Nix, or simply nix develop.

The canonical source and issue tracker for Hacklily is my Forgejo instance, https://slop.nettek.ca/jocelyn-stericker/braat.

I love receiving emails from people using Braat. My preferred way of receiving suggestions, issues, or patches is by email, jocelyn@nettek.ca.

You can also:

• send a PR or issue to the GitHub mirror: https://github.com/jocelyn-stericker/braat
• Reach out to me on Mastodon: @emilyskidsister@hachyderm.io
• send me an email if you wish to get a login for my Forgejo instance
• comment on this Reddit post

Contributions must be licensed AGPLv3-or-later. Feel free to use AI, but please be mindful that I will need to read the patch.

Copyright (C) 2026 Jocelyn Stericker jocelyn@nettek.ca

This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Affero General Public License for more details.

You should have received a copy of the GNU Affero General Public License along with this program. If not, see https://www.gnu.org/licenses/.

This project contains code derived from Praat. That code is:

Copyright (C) 1992-2008,2011,2012,2015-2020,2022-2024 Paul Boersma

Copyright (C) 1993-2020 David Weenink

This project also contains a TypeScript port of the Bournemouth Forced Aligner (BFA) for phoneme-level forced alignment, in src/lib/alignment/. The upstream BFA is:

Copyright (C) Tabahi tabahi@duck.com

The upstream BFA is licensed under GPLv3; this port is distributed under AGPL-3.0-or-later. It runs the CUPE acoustic model via onnxruntime-web; the model weights are not bundled. See src/lib/alignment/README.md and ATTRIBUTION.md for details.