Audion

Browser-based audio analysis. Record from your mic or drop in a file. See the waveform, spectrogram, MFCCs, pitch, formants. Export to JSON/CSV. Everything runs locally — no uploads, no backend.

Try the demo →

What's it for?

• Inspecting recordings visually
• Extracting MFCCs, pitch, formants for ML pipelines
• Finding speech segments in longer files
• Quick phonetic analysis without installing anything

Not for: music production, audio editing, transcription, or live processing.

Features

Input: Record via mic or import WAV/MP3/OGG/FLAC/M4A/AAC/WebM

Live view: Waveform + spectrogram while recording

Analysis:

• MFCCs (13 coefficients)
• Pitch (F0) via autocorrelation
• Formants (F1-F4) via LPC
• Voice activity detection with segment boundaries

Export: JSON with full structure, CSV for spreadsheets, WAV clips for segments

Getting started

npm install
npm run dev

Open http://localhost:5173. Record something or drop a file. Switch to Analysis tab.

Export formats

JSON:

{
  "metadata": { "sampleRate": 48000, "duration": 3.2 },
  "mfcc": { "coefficients": [[...]] },
  "vad": { "segments": [{"start": 0.4, "end": 2.1}] }
}

CSV:

frame,c0,c1,c2,...,c12
0,12.34,-5.67,2.34,...

How it works

Processing runs in web workers so the UI stays responsive.

• MFCCs: pre-emphasis → framing → Hamming → FFT → mel filterbank → log → DCT
• Pitch: autocorrelation + parabolic interpolation + median filter
• Formants: LPC → spectral envelope → peak picking
• VAD: energy + zero-crossing rate with auto threshold

Stack

React + Vite, Web Audio API with AudioWorklet, ComputeKit for workers

Limits

• No noise reduction
• Large files can be slow (it's JS)
• Stereo gets mixed to mono
• VAD threshold is automatic, not adjustable

Deploy

npm run build

Static output in dist/. Works on Vercel, Netlify, or any static host.

License

MIT