The Problem
A recently trained Bangla TTS model performed well on clean test data but failed on actual user input: "১২/০৮/২০২৪-এ ৩pm-এ ডা. রহমানের সাথে meeting আছে।" The sentence - a straightforward appointment reminder mixing Bangla numerals, English abbreviations, and code-switched words - produced robotic, inconsistent output.
The issue wasn't model weakness. The text itself was the problem.
Why This Matters
Bangla serves roughly 250 million speakers but remains chronologically under-resourced in speech tech. Enterprise teams deploying voice interfaces for APAC markets face persistent challenges:
Orthographic mismatches: Written Bangla differs significantly from spoken forms. Consonant clusters like "ক্ষ" (written) become "kkh" in speech. Without grapheme-to-phoneme normalization, even common words get mispronounced.
Numeric chaos: "১২৩" could be "একশ তেইশ" (one hundred twenty-three), "একুশে" (21st), or "শূন্য এক দুই তিন" (phone number digits) depending on context. Models can't guess - they need explicit rules.
Code-mixing ubiquity: Real Bangla text, especially in social media and enterprise contexts, routinely mixes English words. "Meeting", "pm", and abbreviations like "ডা." (Dr.) require different handling than pure Bangla.
Token-level accuracy can reach 90.5% with proper vowel and sibilant handling, according to research. But off-the-shelf models like Whisper fail on normalization and Bangladeshi dialect variants.
The Trade-offs
Rule-based text normalization works but requires constant linguistic maintenance. Statistical approaches don't scale without expensive studio data covering diverse accents. Recent low-cost Bangladeshi efforts emphasize local accent tuning and community data collection, particularly for regions like Sylhet.
Early open-source work (Festival-based Bangla TTS from 2011) highlighted these scalability issues. A decade later, the fundamental problems remain: Bangla orthography and Bangla speech aren't the same thing, and no amount of training data fixes that without preprocessing.
What's Next
The developer's solution: structured, rule-based normalization pipelines that convert written Bangla to spoken forms before feeding text to modern architectures like VITS or Piper TTS. It's not elegant, but it works.
For APAC tech leaders deploying multilingual voice systems: budget for linguistic expertise upfront. The alternative is shipping systems that sound confident on demos and confused in production.