Independent research on TTS quality evaluation and emotion encoding.
Replication of Kacper Wikiel's Zonos emotion channel methodology on Chatterbox. Finding: scalar collapse onto pauses. Vector control routes to timbre. → _1-Chatterbox-emotiononal-channel-analysis/
- Original Zonos: https://kwikiel.github.io/polish-tts-weekend/
Objective benchmark measuring whether voice-cloning TTS preserves breathiness from reference audio. Finding: XTTS-v2 preserves breathiness best (score 0.097), Chatterbox close on retention (0.510), Kokoro baseline zero. → _2-breathiness-preservation-benchmark/
Objective benchmark measuring whether voice-cloning TTS preserves speaker-level F0/prosody from reference audio. Finding: Chatterbox best F0 contour match (DTW 11.27), XTTS-v2 best breathy-modal contrast retention (0.79). Kokoro baseline zero. → _3-prosody-transfer-benchmark/
Measures whether TTS voice characteristics (F0, breathiness via Praat CPP, spectral features) drift over 5-minute synthesized monologues compared to a short reference clip. Finding: all models drift measurably (+0.07–0.08 increase), Chatterbox stays closest to reference (0.491) but drifts fastest. Breathiness (CPP) is the least stable channel. Acoustic drift below perceptual threshold at 5 minutes (single-listener check). → _4-identity-drift/
Measures whether TTS models respect punctuation: periods, questions, exclamations, commas, dashes, ellipses, quotation marks, capitalization across 3 models × 28 utterances. Finding: Chatterbox has best F0-based differentiation (question rise +68.5 Hz/s) but weak pause hierarchy (0.67). XTTS-v2 has perfect pause ordering (1.00) but near-zero question F0 rise (−5.0 Hz/s). Kokoro (no-adaptation baseline) shows intermediate F0 sensitivity (+25.7 Hz/s) with same hierarchy score (0.67). F0 sensitivity and pause ordering are independent dimensions — no model scores high on both. → _5-punctuation-sensitivity/
Tests whether a preceding paragraph changes the prosody of an identical neutral target sentence, using target-only segmentation after full-context generation. Finding: segmentation QC passed and repetitions are real. Kokoro shows strong prior-context carryover, mostly cold vs any preceding paragraph. Chatterbox and XTTS-v2 show no reliable condition effect at n=5. Evidence for semantic content-specific priming is weak. → _6-owl-semantic-priming/
Tests whether a monotone/robotic number prime suppresses pitch variance in a subsequent emotional target sentence — acoustic "hangover" from context window carryover. Finding: Chatterbox shows pilot evidence of pitch compression (-39% f0_cv, p=0.031) with stable speaking rate. XTTS-v2 shows strong trend (-29%, p=0.094). Kokoro inconclusive. Two distinct phenotypes: pitch compression vs tempo acceleration. → _7-subliminal-hangover/
Preregistered, manifest-driven replication of _7. Multi-target, larger-n (N=594). Mixed effects models. Finding: Tempo acceleration confirmed across all three architectures (Chatterbox +0.35, Kokoro +1.36, XTTS +0.70). Pitch compression is Kokoro-specific (-0.045 f0_cv, p=2e-08). Chatterbox f0_cv model unreliable (convergence failures). → _8-subliminal-hangover-v4/
- CodeSOTA TTS ELO: https://codesota.com/text-to-speech/elo