The Lamp & the Machine

Under a banyan tree, a storyteller lights the brass oil lamps. Temple bells wake the night. The village settles onto the earth. And on a stretched cotton screen, backlit by fire, a legend begins to move — coloured leather puppets, perforated like lace, telling of a chieftain who dreamed a city into being.

That screen is Togalu Gombeyaata, the traditional shadow-puppet art of Karnataka. The legend is Kempegowda, the 16th-century Nadaprabhu who, in 1537, yoked a pair of white bullocks to a plough and furrowed the four cardinal streets of the town that became Bengaluru. And the whole thing — every puppet, every lamp, every voice — was generated and edited with AI, over a handful of evenings.

This is the build log: what I made, how I made it, the bugs I hit, and the two reusable skills that fell out of it.

The Images

The visuals came from Sora 2 (on Azure AI Foundry). The trick with a multi-scene film isn't generating a clip — it's stopping every clip from drifting. Independent generations wander: the hero's turban changes, the palette shifts, the backlit screen becomes a photoreal sunset. So I built a locked Style Bible and a five-beat grammar, injected verbatim into every scene.

Invocation → Character → Journey → Conflict → Resolution → Moral.the scene grammar, repeated for every clip

A 25-scene, ~5-minute film rendered at twelve seconds a clip, then stitched with a shared warm-firelight colour grade and crossfades. Kempegowda stayed Kempegowda. The bullock-pair stayed a pair. The lamps never went out.

The Voice

The first cut narrated itself beautifully… in Kannada. Sora speaks whatever language your style-bible names — my audio field had said "a wise elder Kannada storyteller." I confirmed it by transcribing the film with gpt-4o-transcribe, which read back fluent Kannada prose. Lovely — but I wanted an English cut.

Simply asking Sora for an English narrator worked. But I wanted more control — and, crucially, I wanted to re-voice the film without losing its soul: the veena, the mridangam, the temple bells and crickets Sora had generated.

Keep the music. Change the voice.the whole editing problem, in four words

The catch: Sora bakes narration and music into a single track. So I used Demucs (a source-separation model) to split the film's audio into vocals (the old narration) and everything else (music + night ambience). I dropped the vocals, kept the music bed, and laid a fresh Azure Neural voice on top — gently side-chain ducking the music under the narration and remuxing onto the untouched video. The visuals never re-rendered; the music survived; the story changed.

The Cast

One narrator became a company. A warm male elder (en-IN-Prabhat) carried the through-line; a female voice (en-IN-Neerja, with empathetic and cheerful express-as styles) took the lyrical and joyful beats; and for the invocation, the rally, and the closing moral, both spoke — alternating lines, then a unison chant to end:

Remember Kempegowda.the finale — two voices, one line

The first multi-voice cut had two problems, and both were honest lessons. The opener sounded robotic — I'd pushed the prosody too far (-8% rate, -2st pitch), and pitch-shifting artifacts crept in. And the narration stopped after ninety seconds — a real bug (more below). I softened every line to a gentle storyteller pace, fixed the offsets, and then reached for Azure's most human voices: the DragonHD family (en-IN-Arjun, en-IN-Neerja). They read so naturally that no segment needed any time-stretching at all. That became the flagship cut.

The Four Voices

Same film, four narrations — identical visuals, identical Sora music/ambience, four different storytellers. Here they are, side by side.

These are compressed 480p previews, self-hosted so the page works out of the box. The complete branded episode below is that same DragonHD cut — now with its opening jingle in front and end-credits at the close.

The Finished Episode — jingle in front, credits at the end

The four voices were the middle of the tale. To make it a channel, I built a third skill — togalu-brand-bumpers — that tops and tails any film with a generated opening jingle and a scrolling end-credits. The ident holds on the Togalu Gombe Aata medallion, then the backlit screen parts open like a curtain onto a Sora journey — a leather map of India gliding to a Karnataka sunset, the theatre rising, and a hand-perforated leather Ganesha entering in invocation. The channel's name is spoken in Kannada — ತೊಗಲು ಗೊಂಬೆ ಆಟ — by a native voice. Then the film plays, and the credits roll to a Directed by card.

Under the Lamp — the Pipeline

Everything runs on Azure with Microsoft Entra (AAD) auth — no API keys — orchestrated from the terminal.

The Craft Notes — every change, step by step

Two reusable skills came out of this. Here is the exact sequence of changes I made to each.

Skill 1 — togalu-gombe-video (the film generator)

Skill 2 — voice-dub (re-voice any video, keep the music)

Skill 3 — togalu-brand-bumpers (opening jingle & end credits)

Moral

A city is not built by stone alone, but by far-seeing vision and selfless service.the elder's closing line

Kempegowda ploughed four furrows and imagined a city that would outgrow his sight by four centuries. Working on this, it struck me that AI is best when it does the same for our heritage — not replacing the puppeteer's lamp, but carrying its light to a new screen. The craft is 500 years old. It just found a longer run.

What folk story from your culture would you bring back to life under the lamp?