GPT Audio and GPT Audio Mini through OpenRouter.
DECISION CONSOLE / OPENROUTER LIVE TEST
Pick Meron’s actual
Pick Meron’s actual
Amharic brain.
Important audio note
Every play button reads that model’s unique Amharic answer through the same native Amharic voice. You are comparing the model’s wording, grammar, and comprehension—not 25 different TTS voices.
IDENTICAL TEST PROMPT
Explain a nuanced weather sentence in two simple Amharic sentences, encourage the learner, translate it, fix a subject–verb agreement error, and explain Ethiopian hospitality.
RANKED RESULTS— models
90+80–89failed
| # | Model | Amharic | Breakdown | Latency | Input / Output | Listen | Compare |
|---|
END-TO-END AUDIO BENCHMARK
Did the model generate the voice?
Yes, only for the cards marked “native audio model.” Those models generated both the Amharic response and the waveform. Dedicated TTS cards received fixed text, and the 25 language models in the other tab are text-only.
Native Google Amharic plus three direct ElevenLabs attempts.
Rate pronunciation, fluency, and naturalness from 1–5.
CAPABILITY LIMIT
Gemini 3.x, GPT‑6, DeepSeek, Claude, and Qwen are text-only here.
OpenRouter currently exposes text output for those models. Their play buttons in the language tab use one shared native Amharic TTS renderer and must not be interpreted as model-generated voices. Google’s configured API key could not access a Gemini Live/native-audio endpoint, so no Gemini voice sample is fabricated.