English + Arabic TTS benchmark
7 models, 100 English + 100 Arabic call lines, RTX 4090 (baseer-dev), one model at a time. "Normalized" = numbers and codes written as words (what our code will send); "raw" = digits as typed. "Heard" = what Qwen3-ASR heard; a red row means "listen to this one", not a confirmed failure. Clone models copy one reference clip (a synthetic Arabic voice, not a real person); the others use a built-in voice.
Summary
First audio = median time until the first audio is ready (streaming where the model has it, else sentence by sentence like our worker). Speed = time to make audio / length of audio (lower is faster; 0.1 = 10× faster than real time). Error = ASR word error (English) or character error (Arabic, plain lines only, without numbers) against the spoken text. Lines with errors = clips above 15% error.