10 ready agents · same narration text on both models · voice rotates per agent · Speechify via the product TTS path, subtitles from the product .vtt builder. Scores are machine-transcribed (Deepgram nova-3); trust your ears over them.
word = English word the transcriber did not recognise (fr/es only) · CER = character error vs. script (numbers/dates read aloud inflate it) · clip = samples at full scale · pause = longest silence inside the cue · ▶ plays that model's clip; the other stops.