Tuning by measurement
Chatterbox exposes three settings that shape the voice:
| Setting | What it does | Too low | Too high |
|---|---|---|---|
| exaggeration | Expressiveness / emotional intensity | Flat, “reading” voice | Over-acted, unstable |
| cfg_weight | How strongly generation follows the guidance | Loose, follows the reference’s pace and accent | Slow, over-articulated |
| temperature | Randomness when sampling speech tokens | Monotone, repetitive | Erratic |
I can’t judge a voice by ear for every combination, so I scored them.
The scoring
Section titled “The scoring”Each combination generated the same three Hindi chunks with fixed seeds, scored on:
- Speaker similarity to the reference (the model’s own voice encoder).
- Pitch movement: how much the voice rises and falls, compared with the real speaker.
- Pace: onset rate (a syllable proxy) compared with the real speaker.
- Failed chunks, from the new chunk check.
What the sweep showed
Section titled “What the sweep showed”Seeds matter as much as settings
Section titled “Seeds matter as much as settings”Re-running the two best stable options with a second set of random seeds:
| Settings | Similarity | Pitch movement (reference 5.53) | Failed chunks |
|---|---|---|---|
| 0.5 / 0.4 / 0.75 (chosen) | 0.946 – 0.955 | 3.2 – 4.3 | 0 of 6 |
| 0.4 / 0.3 / 0.85 | 0.952 – 0.955 | 3.7 – 4.8 | 1 of 6 |
Changing only the seed moved the scores as much as changing the settings. The chosen values were the only ones with no failures, so they stayed.
The real limit
Section titled “The real limit”Every stable setting is still a little flatter than the real speaker (about 30% less pitch movement), because the model takes speaking style from only 6–10 seconds of reference. The biggest remaining lever is a better reference recording (15–30 s of natural, expressive speech, no clipping, quiet room), not more tuning.
Next: why it isn’t online 24/7.