Human-like Bangla Text-to-Speech Synthesis using the Orpheus TTS Model Fine-tuning

Bangla text-to-speech (TTS) synthesis faces persistent challenges due to limited speech resources and complex linguistic features. This paper presents a fine-tuned Orpheus neural TTS model aimed at generating human-like Bangla speech with higher naturalness and clarity. A multilingual Orpheus model was fine-tuned using an open-source Bangla speech dataset. The fine-tuned Orpheus model achieves a substantial improvement in speech quality, especially natural voice tone. The Orpheus fine-tuned Bangla TTS model shows exceptional performance with speaker similarity of 99.96%, Character Error Rate (CER) of 5.8%, Word Error Rate (WER) of 11.1%, along with high similarity in Mel-Cepstral Spectrum. The results show that the fine-tuned Orpheus TTS excels compared to other TTS systems. The framework put forth outlines a productive and scalable process for developing high-quality Bangla TTS systems and extends the effectiveness of Orpheus to more low-resource languages.