Blindfold Chess | Play chess blindfolded with your friends

BLACK

+5 SEC

10:00

WHITE

+5 SEC

10:00

Blindfolded player(s)

Clock

Voice access

Checking voice access…

Audio is sent to OpenAI when you release the button for speech to text.

THE MOVE INTERPRETER

System prompt

Transcript -> exactly one line:
UNKNOWN
O-O
O-O-O
M|piece|destination|source-square|source-file|source-rank|capture|promotion|special
R|piece|destination|source-square|source-file|source-rank|-|promotion|-

piece=P/N/B/R/Q/K; square=a1-h8; source-file=a-h; source-rank=1-8; capture=x; promotion=Q/R/B/N; special=ep; unspoken=-. M=move; R=spoken recapture. UNKNOWN if not one coherent move. Preserve only spoken constraints. Never infer legality/SAN/check/mate. No explanation.

I trained a 4B model to understand spoken chess.

It gets one finalized transcript and turns it into a compact move description.

Where frontier models fail

Input “yes, can we take the piece on g eight next”

Trained Qwen3 4B M|-|g8|-|-|-|x|-|- ✓

Claude Haiku 4.5 UNKNOWN

GPT-4o mini UNKNOWN

Correct meaning: capture on g8.

Try base vs. trained Model on Hugging Face ↗

Eval tests comparison

Trained Qwen3 4B 29 / 30 · 96.7%

Claude Haiku 4.5 · best prompt 24 / 30 · 80.0%

GPT-4o mini · best prompt 17 / 30 · 56.7%