Command Palette

Search for a command to run...

Live demo
You are browsing real pipeline data captured from a live 5-Domia home mesh — latencies, delegation journeys, memories and emotions are genuine. Voices are synthetic. The demo is read-only.

Analytics

Time to first audio, per-stage latency, model performance and the labeled eval corpus.

S2S time to first audio

2.00 s

median, speech in → speech out

Interactions

40

across 2 flow types

On-device

40%

run on the origin device vs delegated

Eval corpus

10

9 good · 1 needs work

TTFA waterfall (avg, voice replies)
STT121 msLLM1.47 sTTS2.83 s

First audio at 2.01 s — the user hears Domia speak 2.42 s before the full 4.42 s pipeline finishes, thanks to per-sentence LLM→TTS pipelining.

TTFA distribution
Volume over time
Latency trend

On-device vs delegated

On-device16

TTFA p50

2.09 s

p95 3.79 s

Total p50

7.39 s

p95 14.53 s

Delegated (gRPC)24

TTFA p50

1.92 s

p95 2.64 s

Total p50

10.73 s

p95 20.83 s

Fastestt2t832 ms

New topic, Marlowe. Tell me one surprising behind-the-scenes fact about how classic movies were made.

Marlowe · total 843 ms

Slowests2s3.79 s

MARLOW RECOMMEND ME SOMETHING FOR MOVIE NIGHT I WANT A HIGHST FILM

Marlowe · total 12.97 s

Latency by flow
FlowRunsTTFA p50TTFA p95Total p50
Speech → Speech362.00 s3.66 s10.67 s
Text → Text41.52 s2.23 s1.54 s

Latency distribution

Speech-to-text

113 ms

p95 187 ms · avg 121 ms

LLM

1.52 s

p95 2.34 s · avg 1.48 s

Text-to-speech

2.03 s

p95 9.27 s · avg 2.83 s

Time to first audio

1.96 s

p95 2.78 s · avg 1.96 s

Total

9.88 s

p95 18.78 s · avg 9.64 s

Headline value is the median (p50) per stage.

Model performance
StageModel / engineRunsAvg
STTstreaming-zipformer-en36121 ms
LLMllama3.2:3b9678 ms
LLMllama3.1:8b311.71 s
TTSKOKORO352.83 s
By Domia
DomiaRunsTTFA p50Total p50
Sous101.69 s11.27 s
Marlowe91.46 s6.50 s
Atlas72.12 s9.01 s
Torque71.92 s5.47 s
Luna72.14 s13.01 s
Eval corpus

Graded

10

Good

9

Needs work

1

Tagged

10