Sygma ASR Arabic: test bench

abedbanna/Sygma-ASR-Arabic (8-best beam + word-count bonus + breakdown filter) vs. Cohere greedy on the same clip. Highlighted words differ between the two outputs.

Arabic audio

Decode

Sygma + greedy Sygma only Greedy only

repair fallback (not used in the benchmark)

Optional reference transcript (for WER) · النص المرجعي

Sygma ASR

Cohere greedy (base)

Clips up to 120 s; over 35 s use chunked decoding. WER uses a light normalization (no diacritics/punctuation, unified alef/ya/ta marbuta), not the leaderboard's exact eval.py.