Narrator bake-off

Round 5: a Gemini-only scholarly voice audition after all three new Qwen women missed in different ways. The British woman don was too breathy and sexual, the British historian was nasal, and the American professor was too high-pitched.

Every track below reads the exact same 56-word passage. Gacrux tests Google’s mature female preset in both British and American accents, Erinome tests its clear female preset in British English, and Sadaltager and Rasalgethi test knowledgeable and informative male presets in British English. Iapetus now has a true same-text American control.

Round 4B: a correction after the female mezzo and male scholar both sounded too sultry. The word “warm” and the emphasis on low pitch were removed. The four new Qwen briefs instead ask for medium pitch, crisp consonants, dry analytical authority, and public-lecture delivery. They explicitly prohibit breathiness, intimacy, sensuality, vocal fry, languor, and seductive pacing.

Tara stays here as the promising preset reference. Orpheus cannot add scholarly sharpness to that preset, so the new tests use Qwen VoiceDesign: two British female academics, one dry American woman professor, and one dry British male academic.

Round 4: five guarded local tests prompted by the Round 3 verdict. The local candidates all read the same 56-word Maineville passage. Qwen VoiceDesign auditions two deeper scholarly female designs and one male control, with explicit Charlus and Jupien pronunciation guidance. Orpheus auditions its Tara and Zoe female presets, which do not accept a style or pronunciation prompt.

The Gemini track is the existing full Maineville note, included as a voice reference rather than a same-text sample. MOSS was stopped when its real 17.6 GB model-plus-decoder footprint and throttled transfer breached the time guardrail. KugelAudio was gated out because the current MLX port requires 18.7 GB and ignores its voice selector.

Round 3: four real “How to read this” notes, with the Gemini 2.5 Flash/Iapetus test voice first and free local models after it. Every engine reads the exact same title, summary, and explainer text. Kokoro uses a neutral American preset. Chatterbox Turbo uses its built-in stock voice, with no voice clone, reference recording, or style prompt. The older optional Qwen trial failed before producing audio and is not listed. Julia ruled Chatterbox “super off” and Kokoro “nothing like our Gemini voice.”

Round 2 changes: style prompt now demands a consistent American accent with French names in the “wine register” (Merlot: silent final consonant, English stress). Charon was too deep, so three lighter voices audition on P1. P3 appears both prompt-steered and lexicon-respelled (the model literally reads “Vair-duh-ran”, “Sant-oo-vairt”, “Frahnk-toh”) to compare the two levers.