Adapting Kokoro-82M to Bengali: A Teacher-Forced Fine-Tuning Recipe for a Text-to-Speech Model Released Without Training Code

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Kokoro-82M is a compact, high-quality text-to-speech (TTS) model whose weights are public but whose training code is not. Its released voices cover a handful of languages; Bengali is not one of them. This report describes a complete recipe for adding a new language to Kokoro-82M using only the published inference package and an external forced aligner. We reconstruct the StyleTTS2-style teacher-forced training objective on top of the unmodified model modules, obtain phoneme durations from Meta's MMS forced aligner via romanised word alignment, map espeak-ng phonemes onto Kokoro's fixed 178-symbol vocabulary, and train on 18.3 hours of single-speaker studio Bengali from the IIT Madras IndicTTS corpus. A first attempt that drove the vocoder with the model's own predicted pitch and energy produced muffled, harmonic-poor speech that did not improve between 1,000 and 2,000 steps. Switching the vocoder input to ground-truth pitch and energy during training, as StyleTTS2 does, restored clear harmonic structure at 1,000 steps and raised output loudness by 5.5 dB toward the reference. Training is ongoing; we report the pipeline, the ablation, and objective measurements, and outline the perceptual evaluation planned once training completes.