Finding a voice
by listening.

Make the audition blind.

“actually have 26 of them and the labels are letters of the alphabet only.”

Ben → Claude, 11:36 a.m.

Claude built a local HTML listening page with A–Z samples, automatic advance, and copyable ratings. The first rounds mixed Kokoro, Supertonic, Pocket and Soprano voices with cheap processing. The next brief replaced integer stars with a continuous 0–5 slider. A later comment box let Ben say what a number could not.

The first round’s “humanize” chain averaged 4.25. In round two, raw samples averaged 4.03 and humanize 3.66. The samples differed across rounds, so this is no controlled head-to-head effect estimate. It was enough reason to question whether more processing was helping.

Turn an adjective into a technical change.

“the background noise is not part of the spoken performance on the soundstage idk how to say it”

Ben’s round-three comment, 12:34 p.m.

“More human” became a specific mixing problem. Room tone added after the voice chain sounded detached. The next round put room sound through the microphone chain with the voice, narrowed the search to af_heart and bm_lewis, and compared lavalier, condenser and close-mic colouring.

Ben liked the dry lavalier’s clarity and the condenser’s timbre on af_heart; bm_lewis could sound like a voice actor. Inserted pauses sounded “weird,” and a de-esser made little audible difference. The useful output was a few reusable presets, not a maximal effects stack.

Let the live voice warm up.

The work moved to GPT-Live through the Codex live-voice connection. An elaborate “voice actress” prompt produced performances Ben found fake. At 1:17 p.m., he replaced it with a simple request:

“we’d like you to perform this line for us: ‘line’”

Ben’s suggested wording, 1:17 p.m.

Record the whole reply, allow its “Absolutely, Ben…” preamble, then cut to the requested line. Preserve the pauses inside it: “dont cut out pauses btw.” The second short-line batch had twelve takes across nine voices. Ben called the overall quality very high; natural short lines still did not settle how a whole film should sound.

Exact words were not enough.

The first full-film experiment used a local male voice to speak the instructions. The quotation boundary disappeared in speech, and the live voice improvised. Typed turns fixed that: eight separate Vale replies matched the requested paragraphs in the session transcript and could be placed at the original beat starts.

“revoice is very bad. try directing them w/ text again and to the full script and one take.”

Ben, 1:45 p.m.

A text match answered the engineering question. It did not answer the listening question. Ben changed the unit of performance from a paragraph to the whole script.

Less direction. More continuity.

“continuous is always better.”
“we are overprompting”

Ben, 2:00 p.m. · judgments within this session

Claude compared continuous one-take readings with versions split at paragraph pauses and shifted back onto the old beat starts. Ben preferred continuous delivery; Ember was the best take in that round. The next batch used only “Unhurried documentary narration” plus the script.

Sol won: “best so far,” followed by “reeealy good.” Maple’s voice changed over the narration; Ember fell from the prior round’s favourite to 1.4. Recording several takes and listening remained necessary.

Keep the failed idea in the story.

In a parallel experiment, the live voices listened to the best local male model read the script, then received “Done. Now your own take:”. Ember and Maple treated “take” as an invitation to give an opinion. Maple even chatted during playback. The longer listen-then-perform setup failed as run.

That does not establish that audio references cannot help. The wording was ambiguous. The saved notes propose “Now please perform it yourself, in your own style:” for a future attempt; that repair was not tested in this session.