Duet — private peer-to-peer voice calls

A static web app for HD voice calls straight between two browsers. There is no backend, no accounts, no build step and no dependencies: just HTML, CSS and plain JavaScript modules. It runs on GitHub Pages, Vercel or any static host.

How it works

Part	What it does
Audio	WebRTC with the Opus codec, sent directly peer-to-peer. It is always encrypted with DTLS-SRTP (AES, hardware accelerated, so it adds microseconds of delay, not milliseconds).
Finding each other	The two browsers exchange setup messages through public MQTT brokers (EMQX and HiveMQ, used at the same time for redundancy). Every message is AES-256-GCM encrypted with a key derived (PBKDF2) from the room secret, which lives only in the link's #hash and is never sent to any server. The brokers see only random topic names and ciphertext.
After connecting	Signaling moves to a P2P data channel, so quality changes and mute state never touch a relay.
Extra-private call	Uses no relay at all. You swap two short codes or links yourselves, through any chat app.
Security code	Both screens show an 8-digit code derived from the two DTLS fingerprints. If the codes match, nobody is sitting in the middle of the call.

Interface

• An Apple-style design with Light, Dark and Automatic themes (Automatic follows the device), plus a Larger text option.
• Plain wording everywhere. Connection numbers, routes and data use are hidden unless you turn on Show technical details in Settings.
• An optional name is shown to the other person, with their initials on the call screen.
• Text messages during a call. They travel over the same encrypted P2P channel as call control (or the encrypted relay while connecting), show "Delivered" once received, and are never stored: they disappear when the call ends.

Quality presets

Both sides automatically use the lower of their two choices, so one slow phone doesn't break the call for both people.

Preset	Opus bitrate	Audio band	Packet	≈ data/hour
Best	96 kbps	48 kHz fullband	20 ms	~105 MB
High	40 kbps	48 kHz fullband	20 ms	~34 MB
Standard	24 kbps	24 kHz	20 ms	~25 MB
Data saver	14 kbps	16 kHz wideband	40 ms	~14 MB
Lowest data	8 kbps	8 kHz	60 ms	~8 MB

• Automatic (the default) starts from the Network Information API's hint, then moves up or down based on live packet loss, jitter and round-trip time.
• Every preset uses Opus in-band FEC, which recovers lost packets. Every preset except Best also uses DTX, which sends almost nothing while you're silent.
• Longer packets on the low presets cut header overhead, which costs about 16 kbps at 20 ms.
• Changing quality mid-call applies to both directions within about a second. There is no reconnect.

Limits

• Two people per call.

• Strict networks. Some network pairs block direct P2P traffic, most often a phone on mobile data (carrier NAT) or an office firewall. On those, a TURN relay is needed. TURN only forwards the encrypted audio and can't decrypt it, and only one side of the call needs it. You can add one in two places:

  • For everyone: put credentials in js/config.js. Free tiers include ExpressTURN and Metered.
  • Just for you: use Settings → My own relay server in the app.

  When a call can't connect, the app retries once, then explains why in plain words. Turn on Settings → Show technical details to also see which routes each side found.

• Public brokers and STUN. The app uses Google and Cloudflare STUN servers and public MQTT brokers to find the other person. They never carry audio, and they only see encrypted setup data. If you need zero third parties at all, use an Extra-private call. Note that STUN is still used there to find your public address.

Deploy

It's a static folder:

• GitHub Pages: push the files and enable Pages on the branch root.
• Vercel: run vercel in this folder (framework: "Other", no build command).

Browsers only allow microphone access over HTTPS (or on localhost).

Run locally

python3 -m http.server 8080