do tts voices drift over 5 minutes of speech?

if you make a TTS model talk for 5 minutes straight, does the voice drift away from the original? does it start sounding like someone else by minute 5?

i built a benchmark to find out.

what i did

i generated 5-minute monologues from 3 models (Chatterbox, XTTS-v2, Kokoro) using the same 868-word narrative text. same reference voice (VCTK p229) as the earlier papers.

i sliced each into 15-second windows and measured acoustic features per window: pitch (pyin), breathiness (Praat CPP), spectral features, MFCCs, RMS loudness, spectral centroid. then i measured drift as the IQR-scaled Euclidean distance from the reference clip across time.

results

Model	Drift Mean	Drift Increase	Drift Slope
Chatterbox	0.491	+0.080	+0.00422
XTTS-v2	0.583	+0.076	+0.00431
Kokoro	1.290	+0.070	+0.00473

all three models drift. the rate is suspiciously similar across all of them (~0.004 per 15-second window). this could mean drift is a fundamental property of long-form generation - or it could mean my test text was too structurally uniform to stress models differently. that's a real methodological limitation.

Chatterbox starts closest to the reference (mean 0.491) but drifts the fastest (+0.080 increase). XTTS-v2 starts a bit further (0.583) but is more erratic. Kokoro isn't trying to clone the voice - it's the no-adaptation baseline at 1.290.

breathiness is the least stable channel. for both voice-cloning models, CPP (breathiness) drift exceeded pitch drift. the texture of the voice degrades faster than the pitch. this connects directly to the breathiness benchmark finding - breathiness was the hardest feature to preserve in short-form cloning, and now it's also the fastest to drift over time.

little surprises

• XTTS-v2 had 3 windows with audio clipping (signal hits max amplitude), which may explain its high RMS drift (1.135 vs Chatterbox's 0.286).
• Chatterbox had to be generated in 10 chunks of ~100 words due to CUDA context limits - chunk boundaries could introduce artifacts not present in continuous generation.
• the near-identical drift slopes across all three models is suspicious enough to call out as a likely confound of text homogeneity.

limitations

• single continuous narrative text - real speech has dialogue, questions, lists, emotional shifts
• 5 minutes may be too short for perceptible drift
• Euclidean distance in feature space is not validated against human perception
• single-listener check only, no formal ABX test
• only one reference speaker tested

full code and results in the tts-research repo. download the full report (PDF)