The race video commentary is my own voice, cloned zero-shot in Chatterbox by ResembleAI from three reference recordings at three energy levels. Claude writes each line from race facts that Python works out first.

The first seven seconds of each recording.
The first seven seconds of each recording.
calm
build
hyped
This is the energy the clone picks up.

A flat sample gives a flat clone at every setting

Zero-shot cloning copies prosody along with timbre, so exaggeration cannot add energy the reference recording does not contain.

  • The first setup used one recording plus exaggeration, a parameter that raises delivery energy.
  • Now there are three recordings, same person, same microphone: calm for the intro, build mid-race, hyped for overtakes and the finish.
  • Every script line is tagged with one, so the energy shifts the way a real broadcast does.
  • The settings still apply on top of each recording:
style exaggeration cfg_weight
calm 0.50 0.50
build 0.90 0.40
hyped 1.10 0.35
That Time I Got Reincarnated as a Slime
Attack on Titan
Hunter x Hunter
Written by Claude from the facts, spoken by the copy of the hyped recording.

Python works out what happened, Claude only writes it down

If something is not in the fact list Python computes, it cannot end up in the commentary.

  • Reading a race is math: who led at nine seconds, who climbed the most, how big the final gap was.
  • Given raw numbers, a language model writes an excited call about an overtake that never happened.
  • Claude turns the fact list into eight short lines that sound natural said aloud.
  • The model never writes the first and last lines, "Did the chat pick your favorite?" and "Did yours win?". Those belong to the channel.

Each line is capped at 9 words and 48 letters

Letters predict speaking time better than words, so the letter limit does most of the work.

  • Lines sit at eleven fixed times, about three seconds apart.
  • The first rule was a 10-word limit. The Tokyo Revengers script stayed under it and still ran eight seconds long.
  • A line like "Ken Ryuuguji closing in on Manjirou Sano" fits the word limit and is still a mouthful.
  • A cloned voice takes 40 to 65 percent longer to say the same words.

Every line is recorded before it is scheduled

The schedule is built from real clip lengths, and a line that runs into the next one pushes it later.

  • Chatterbox pads a random amount of silence onto each clip.
  • It sometimes makes a bad take: the same six-word line once came out at 0.39 seconds.
  • Each line is generated several times and picked by length.
  • Under the same limits, lines running long range from 0 of 11 to 6 of 11.

One narrator for every series

My voice is the narrator on every series.