Say a sentence and pass it down a line of AI relays. Each one speaks it, sends it over a crackly line, hears it, and retells it in its own words. Watch the meaning drift, round by round, all on this laptop.
Not loaded. Four models (~980 MB in total: a voice, ears, a retelling brain and a meaning matcher) download once, then cache.
The message
or type it. Aim for 6+ words.
How each relay retells it
Meaning drift
–final similarity to the original
–relays before it lost the plot (below 0.50)
Leaderboard
#
Who
Message
Score
How it works
Speak: Kokoro turns the text into speech, with a different voice for each relay.
The line: the audio goes through a phone filter (300–3,400 Hz) plus hiss, crackle and, on tin can, dropouts. The noise is seeded, so the same message on the same settings gives the same chain.
Hear: Whisper transcribes what came down the line. Misheard words are highlighted.
Retell: a small language model (Qwen2.5 0.5B) passes it on in its own words, with greedy decoding.
Score: MiniLM embeds the original and each retelling; cosine similarity is the drift chart.
The lesson: every step is “pretty good”, but errors compound. It's the same reason long agent pipelines go off the rails.
Load the line, set a message and start the chain. Each relay shows up here as it happens.