Skip to content
All work

Speech, Audio & NLP

Custom Bangladeshi-Accent Bangla TTS

HawkEyes · Custom-TTS-Bangla-Accent-VITS-XTTSv2 · current focus · 2026

A VITS text-to-speech model with a grapheme frontend, built to give production voice agents a natural Bangladeshi accent.

Coqui VITSPyTorchtorchaudiolibrosaBangla NLP
90
commits · very active
VITS
+ grapheme frontend
current
primary focus(CV)

Problem

Natural Bangla speech synthesis is a low-resource challenge, and off-the-shelf voices lack a genuine Bangladeshi accent — a blocker for deploying trustworthy AI call agents for hospital appointment automation, sales, and support.

Approach

  • 01Fine-tune VITS with a purpose-built Bangla text frontend: Unicode NFC normalization, zero-width handling, danda/digit normalization via a grapheme-based approach.
  • 02Build the supporting speech-dataset pipeline (genMaxTxt): transcribe with Google Cloud Speech, normalize transcripts with Gemini, and align audio to spreadsheet entries.
  • 03Target deployment inside the real-time voice agent for inbound/outbound call automation.

Results & Impact

  • An end-to-end Bangla TTS recipe under very active development (90 commits) — the author's current primary focus.
  • Designed to plug directly into the production voice agent to unlock Bangla telephony automation.— CV-sourced