โ† All examples

ModelEthio-ASR Amharic (600M)2Ethio-ASR Multi (300M)2Ethio-ASR Multi (600M)1Ethio-ASR Multi (94M)3Meta MMS-1B7OmniASR-CTC (1B)4OmniASR-CTC (300M)4OmniASR-CTC (3B)3OmniASR-LLM (1B)0OmniASR-LLM (300M)1OmniASR-LLM (3B)1OpenAI Whisper-small7
Audio qualityClean Studio2Ambient Room Noise2Harsh Environment3Distant / Low-Volume Mic2Wideband Phone (VoLTE)3Randomized Phone Channel4Narrowband 2G + Packet Loss7

What did the model get wrong?

OmniASR-CTC (300M) Randomized Phone Channel   Example 3 of 20  ยท  speaker: Female

๐ŸŽง Listen โ€” Randomized Phone Channel
Phone-call audio with random extra distortions layered on top.
โœ… What was actually said
แ‰ แˆตแˆแŠญ แ‹ฐแ‹แ‹จ แˆ›แŒแŠ˜แ‰ต แ‹จแˆแ‰ฝแˆแ‰ แ‰ต แ‹จแŒฅแˆช แˆ›แŠฅแŠจแˆ แŠ แˆˆ
๐Ÿค– What this model heard
แ‰  แˆตแˆแŠญ แ‹ฐแ‹แ‹จ แˆ›แŒแŠ˜แ‰ต แ‹จแˆแ‰ฝแˆแ‰ แ‰ต แ‹จแŒฅแˆช แˆ›แŠฅแŠจแˆ‹แˆˆแ‹
Correct
Wrong (different word)
Missed (skipped)
Extra (added)
4
Correct
2
Wrong
1
Missed
1
Extra
57%
Words right
Out of 7 spoken words, this model got 4 right (57%). Errors: 2 words wrong; 1 missed; 1 extra.

Try switching Audio quality โ€” see the same model degrade as the audio gets noisier.