Quantifying the Intelligibility–Expressivity Tradeoff in Emotion-Conditioned Kazakh Text-to-Speech

Abstract

Emotion-conditioned text-to-speech (TTS) is attractive for low-resource narrative, educational, and accessibility applications, but acoustic rendering for expressivity may also reduce content preservation. This paper presents an objective diagnostic study of Kazakh TTS using translated-CARER planning labels and automatic speech recognition (ASR)-based back-transcription. Rather than training a Kazakh emotion classifier or claiming subjectively validated emotional quality, we treat emotion labels as text-side planning metadata and measure how rendering affects Character Error Rate (CER), Word Error Rate (WER), synthesis failures, and label-dependent degradation. A full CARER planning evaluation on 1014 translated Kazakh chunks shows that configured full rendering increases mean CER from 0.417 to 0.865 and mean WER from 0.998 to 1.403 compared with the no-rendering condition. A rendering-intensity sweep over 240 chunks indicates that every tested non-zero intensity increases ASR error in the present configuration. Per-emotion analysis further shows that high-arousal labels are more sensitive to rendering-induced degradation, with the largest CER increase on angry chunks (ΔCER = +0.619), whereas love-mapped neutral chunks remain stable (ΔCER = −0.011). These results establish a configuration-specific ASR content-preservation cost curve rather than a perceptually validated expressivity trade-off. The study motivates calibrated, label-aware rendering and human perceptual validation before drawing conclusions about listener-level intelligibility or emotional benefit.