Productivity & Voice

sunAI

Automating regional voiceovers for education content, while making sure code and technical terms are said, not translated.

Timeline
2024 – now
Status
Shipped
Stack
Python, Sarvam TTS, FastAPI, Custom dashboard, Analytics
Links
Try it live ↗
Problem
Localising thousands of slides and scripts into regional audio takes months of recording, and generic TTS mangles code and jargon.
Bet
Automate the voiceover pipeline, and protect technical terms so they are pronounced, not translated.
Built
A TTS engine on Sarvam's regional voices, a phoneme dictionary for code jargon, and a monitoring console built for this one job.
Result
About 60% faster than the manual process, 99% phonetic accuracy, under 1.2 seconds per generation.

Works today

  • Regional voiceovers from scripts and slides
  • Code and jargon said as terms, not translated
  • A console for speed, tokens, overrides and language mix

Limits

  • Voices are limited to Sarvam’s regional set
  • New jargon needs a dictionary entry

Status

Shipped, live demo

A 1-minute read

Speech, stacked, mid-sentence

Scale is the natural enemy of quality in regional localisation: turning thousands of slides and scripts into audio guides by hand takes months of recording and audits. sunAI automates that pipeline across regional Indian languages without losing the syntax of technical speech, so a for loop still sounds like a for loop in Bengali instead of a literal translation of the words.

How it works

  1. Scripts and slide text go in.

  2. A regex and NLP detector flags programming keywords and equations.

  3. Flagged terms are swapped for phonetic strings from a custom dictionary, so they are pronounced rather than literally translated.

  4. Sarvam’s TTS renders natural regional voiceovers without robotic pacing.

  5. A console tracks generation speed, token use, phonetic overrides and language density.

What I built

  1. Sarvam TTS model selection

    Chose and integrated Sarvam's TTS API to generate natural-sounding voiceovers across regional Indian languages, avoiding robotic pacing.

  2. Phoneme dictionary layer

    A custom regex and NLP parser flags English programming keywords and equations and replaces them with phonetic strings, so they are spoken correctly in regional guides.

  3. Custom monitoring console

    A dashboard built for this application rather than a generic toolkit: audio generation speed, token consumption, phonetic overrides and regional language density, for far clearer auditing.

Decisions

  1. Voices chosen for pacing

    Sarvam was picked for natural pacing in regional Indian languages, the part generic TTS gets most wrong.

  2. Fix jargon before the model

    Pronunciation is handled by an explicit phonetic layer in front of TTS, so a wrong word is a dictionary entry, not a retrain.

  3. A console for one job

    The questions were specific (which override fired, which language is slow), so the dashboard was built for them instead of borrowed.

Outcome

~60%
Faster than manual
99%
Phonetic accuracy
< 1.2s
Per generation

Now

Version 1.0.4

Automatic detection of English programming jargon, overriding robotic regional pronunciation.

Stack

Python, Sarvam TTS, FastAPI, Custom dashboard, Analytics

Keeper of this file

Hoot

Up all night listening. Repeats everything back, correctly pronounced, code included.

← Poke it

Review

“Hoo. I fed it pure Sanskrit and it read it back so cleanly the neighbourhood cats got confused and left. Five stars. Best platform for my global takeover announcements.”
Hoot, owl★★★★★

Try it