โ† All examples

ModelEthio-ASR Amharic (600M)4Ethio-ASR Multi (300M)14Ethio-ASR Multi (600M)7Ethio-ASR Multi (94M)7Meta MMS-1B7OmniASR-CTC (1B)8OmniASR-CTC (300M)9OmniASR-CTC (3B)7OmniASR-LLM (1B)5OmniASR-LLM (300M)4OmniASR-LLM (3B)6OpenAI Whisper-small222
Audio qualityClean Studio1Ambient Room Noise2Harsh Environment2Distant / Low-Volume Mic0Wideband Phone (VoLTE)2Randomized Phone Channel2Narrowband 2G + Packet Loss6

What did the model get wrong?

OmniASR-LLM (3B) Narrowband 2G + Packet Loss   Example 14 of 20  ยท  speaker: Male

๐ŸŽง Listen โ€” Narrowband 2G + Packet Loss
Heavy background noise plus phone-call compression โ€” worst-case stress test.
โœ… What was actually said
แŠฅแˆฝ แ‰ชแ‹ตแ‹ฎ แ‹ˆแ‹ญแˆ แˆ˜แˆ˜แˆชแ‹ซ แŠ แˆˆ แŠฅแŠ•แ‹ฐแ‰ต แŠฅแŠ•แ‹ฐแˆแŒ แ‰€แˆ
๐Ÿค– What this model heard
แŠฅแˆญแˆตแ‹ˆ แ‹˜แ‹ญแˆ แŠแŒˆแˆญ แŠ แˆˆ แŠฅแŠ•แ‹ฐแ‰ต แ‹ฐแŒแˆž แŒ แŒ แˆ
Correct
Wrong (different word)
Missed (skipped)
Extra (added)
2
Correct
4
Wrong
1
Missed
1
Extra
29%
Words right
Out of 7 spoken words, this model got 2 right (29%). Errors: 4 words wrong; 1 missed; 1 extra.

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