Live pipeline

What happens between your mouth and their ear

Four stages, roughly 38 milliseconds. Here is every one of them, and the numbers behind them.

STAGE 01

Capture

Audio is captured at 48 kHz in 20 ms frames, with automatic gain control applied before anything else touches the signal.

STAGE 02

Clean

Echo cancellation and the neural suppressor run on device, stripping room noise while leaving speech untouched.

STAGE 03

Encode

Opus encodes the cleaned frames, adapting bitrate between 12 and 64 kbps as the network moves under you.

STAGE 04

Deliver

Frames travel DTLS-SRTP encrypted over WebRTC, with an adaptive jitter buffer smoothing playback on the far side.

Latency budget

Under 40 ms, end to end

Each stage gets a fixed slice of the budget. Frames are 20 ms, the jitter buffer starts at 20 ms and only grows when the network forces it, and everything else is measured in microseconds.

  • Talkmos push-to-talk38 ms
  • Typical VoIP call150 ms
  • Live streaming relay400 ms

<40ms

End-to-end latency

48kHz

Full-band sample rate

SRTP

Encrypted transport

12kbps

Low-data mode floor

Adaptive bitrate

The encoder watches round-trip time and loss, then walks the bitrate between 12 and 64 kbps. On a clean connection you get full-band stereo-width speech; on a train in a tunnel you still get intelligible words instead of a dropped call.

Reference

Technical specifications

The exact configuration the client and relays negotiate.

Codec
Opus 1.5 (full-band)
Sample rate
48 kHz / 20 ms frames
Transport
WebRTC over UDP, TCP fallback
Encryption
DTLS-SRTP (AES-128-GCM)
Bitrate range
12 – 64 kbps adaptive
Jitter buffer
Adaptive, 20 – 120 ms
Packet loss concealment
Opus PLC + selective retransmit
Echo control
AEC3 with delay estimation

A note on encryption

Media keys are negotiated per session over DTLS and the audio itself travels as SRTP. Relays forward encrypted packets; they are not decryption points. Nothing is written to disk during a live session, so there is no recording to leak — unless a room owner explicitly turns recording on, which is announced to everyone in the room.

Hear the difference yourself

Numbers only go so far. Install Talkmos and put it against whatever you use today.

Privacy policy