A voice that starts with the language.

Speech begins before the waveform.

I developed a Nepali text-to-speech system, building a custom grapheme-to-phoneme (G2P) frontend and training a VITS2 model with a RingFormer vocoder. G2P converts written characters into the sound sequence a speech model uses. For Nepali, getting that sequence right means dealing with pronunciation that the spelling alone does not fully resolve.

The schwa deletion problem.

A Devanagari consonant can carry an inherent vowel. That vowel is sometimes spoken and sometimes dropped—a process called schwa deletion. Dropping every final inherent vowel would mispronounce words that retain it; retaining every vowel would introduce unwanted syllables. My G2P work addressed deletion and exceptions at the pronunciation stage, before synthesis.

Why a blanket rule breaks down.

Nepali pronunciation depends on the surrounding letters, consonant clusters, and grammatical form. For example, देश drops the final inherent vowel, while समय retains it. Even a cluster-based rule needs exceptions: मञ्च is an example where the final vowel is omitted. These linguistic examples explain the problem the frontend needs to resolve; they are not outputs from a live model.

• Distinguish an explicitly written vowel from an inherent vowel that may be deleted.
• Account for the word’s pronunciation context when deciding whether to retain the vowel.
• Treat exceptions as part of pronunciation handling before passing phonemes to the speech model.

Where VITS2 and RingFormer fit.

The synthesis architecture pairs VITS2 with RingFormer in the waveform-decoding stage. At a high level, VITS2 connects the input sequence to timing and a learned speech representation; the vocoder turns that representation into audio. The custom Nepali G2P frontend sits ahead of that model, giving pronunciation handling a distinct place in the system.

Local detail, wider context.

RingFormer, introduced by Seongho Hong and Yong-Hoon Choi, combines ring attention with convolution-augmented transformers. The architecture brings local signal detail and broader sequence context into waveform generation. I used that published architecture in my Nepali TTS work; the original research and implementation are linked below.

Two different engineering problems.

Pronunciation and audio generation have different responsibilities. A waveform decoder can produce convincing sound while speaking the wrong phoneme sequence. This project brought language-specific processing and neural speech synthesis together, with the G2P frontend responsible for the pronunciation representation and the model responsible for generating speech from it.