do tts models actually read punctuation?

when you write a question mark, does the TTS model raise its pitch at the end? when you use a comma, does it pause longer than an em-dash?

you'd hope so. but hope isn't a benchmark.

what i did

i designed 28 test utterances across 5 categories:

• sentence-end: period vs question mark vs exclamation point
• pause hierarchy: comma vs em-dash vs ellipsis vs semicolon
• quotation: reported speech vs quoted speech
• trailing: ellipsis vs period
• capitalization: ALL-CAPS vs normal vs lowercase

i generated audio with 3 models (Chatterbox, XTTS-v2, Kokoro) using the same VCTK p229 reference voice. then i extracted pause metrics, terminal F0 slope, amplitude decay rate, and overall F0 statistics.

results

Model	Question F0 Rise	Pause Hierarchy	Quotation Shift	Trailing
Chatterbox	+68.5 Hz (best)	0.67	YES	wrong direction
XTTS-v2	-5.0 Hz (none)	1.00 (perfect)	NO	correct direction
Kokoro	+25.7 Hz	0.67	NO	wrong direction

the split is clean. XTTS-v2 dominates temporal cues - it gets pause hierarchy perfect (comma shorter than ellipsis shorter than semicolon) and handles trailing ellipsis correctly - but it completely ignores terminal F0. a question mark produces no pitch rise.

Chatterbox is the opposite. it nails F0-based cues: question rise (+68.5 Hz is a real lift), quotation shifts prosody. but its pause ordering is weak.

no model does both well. F0 sensitivity and pause ordering look like independent abilities - neither implies the other.

little surprises

• F5-TTS was excluded from this benchmark entirely. its architecture concatenates reference text and generation text internally, causing reference-text bleed into output. not fixable within this paradigm.
• Kokoro tracks Chatterbox on pause metrics despite being a different architecture and having no voice cloning. suggests pause patterns may be more driven by training data than model architecture.
• the quotation shift in Chatterbox is subtle but measurable - F0 range shifts by >10 Hz between reported and quoted speech. XTTS just reads them the same.

limitations

• energy-based VAD is crude - 30ms threshold, no forced alignment
• F0 extraction (pyin) noisy on very short utterances
• only 2 items per subcategory, too small for publication
• Kokoro has no voice cloning, included as no-adaptation baseline
• quotation detection threshold (10 Hz F0 range shift) is arbitrary

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