Can you direct a model on how to say it the way you direct an actor, and have it actually follow?
A director tells an actor to whisper the line, to slow the next one down, to lean on that word and not the one after it. I wanted to know whether you can do the same thing to a model, in Hindi, and have it actually follow.
You can make it whisper, and the whisper is a real one. You can make it exhale slowly, and the extra time is real. You can stretch a word out. You can move the emphasis onto a chosen word.
The instructions work. What takes the practice is knowing which instruction produces which result, because words that mean the same thing to you do not mean the same thing to the model.
One voice, held constant. One thing changed at a time. Everything generated in the model's most expressive mode.
The lines are deliberately flat, from single sentences up to long passages. They state facts and carry no feeling in the words, so whatever feeling comes out has to have come from the direction rather than from the writing.
This is the first thing to get right and it changes everything after it.
The model has three stability settings. I ran the same whisper direction through all three.
Audio shown as original model output. Playback levels have not been normalized.
The whisper direction worked in all three settings, but Creative produced the strongest and most complete whisper response.
Everything below was generated that way.
Ask it to whisper and it whispers, properly. Not a quiet read of the same performance, an actual whisper: the voice stops vibrating for most of the line and the sound moves up into breath and air. By the end of it the voice is only a fraction as engaged as it is in the untagged version.
It holds over length too. The third clip is the same direction carried across a long dialogue rather than a single line.
It is worth knowing what this sounds like before you judge it. A real whisper is a person lowering their voice, not a stage whisper. It is smaller than you expect and it is correct.
Ask for a sigh, then ask for the same thing in different words. "Sigh" and "exhale slowly" describe one idea to a person and produce two different performances from the model.
"Sigh" was less predictable across different generations. One take stayed close to the normal voice, while another became more pause-heavy. "Exhale slowly" produced a clearer pacing change, making the line about 1.3 seconds longer. The slower timing happened throughout the sentence instead of appearing as one long pause.
Similar instructions can make the model perform in very different ways.
This is the single most useful thing I learned. When you write direction into a script:
You are not describing a mood, you are choosing a mechanism.
"Sigh" and "exhale slowly" are not interchangeable. Pick the phrase whose result you have actually heard.
Ask it to draw a word out and it does. When the name was generated by itself, the spoken part increased from about 0.58 seconds to about 1.00 second, around 70% longer.
Inside a full sentence, the stretching effect did not work reliably. Most of the effect disappeared.
Production takeawayIf one word needs to be stretched very clearly, generate that word separately or use more detailed pronunciation control.
Emphasis is the one that matters most in drama. The same sentence means different things depending on which word carries the weight. Devanagari does not have uppercase and lowercase letters, so the usual capital-letter emphasis trick used in English is not available.
I approached it two ways. First with a written direction, the same kind of tag as sigh or exhale slowly, asking for the stress on a chosen word. Then phonetically, writing that one word out in the International Phonetic Alphabet with the stress mark on it and leaving the rest of the line in Devanagari.
Both the written stress tag and the IPA version clearly changed the performance and made the target word feel more important in this test.
Production takeawayThe written stress tag is faster and easier to read. IPA takes more work, but gives more control over which syllable is stressed.
Separately from the direction work, I built five characters: a Punjabi truck owner, a Bihari railway contractor, a Bengali bookkeeper, a Tamil clearing agent and a Hyderabadi landlord.
The voice alone was not enough. In these tests, changing the spelling and pronunciation inside the dialogue gave better control over the regional speech style. The selected voice still helped shape the tone, pace and pitch.
When someone says "oye preetay, tu aaj-uh ischool nai gya", every accent marker in that sentence is a spelling or a word choice. The vocative. The name's vowel raised. A vowel added to the middle of one word and to the front of another. The negation shortened. Those markers sit in the text you hand over, which is why the script is where they can be controlled most directly.
So each character got their dialogue rewritten. Same Hindi words, same Hindi grammar, only the spelling changed to match how that person pronounces it.
The Punjabi character loses the breath on his heavy consonants, so भाई is written पाई. Clusters take a vowel in front, so स्कूल becomes इस्कूल. नहीं shortens to नई.
The Tamil character breaks his clusters from the inside instead, so the same word becomes इसकूल and platform becomes पिलाटफारम. That single difference, whether the vowel goes before the cluster or inside it, is the fastest way a listener places where someone is from.
The Bengali character rounds his vowels, so सब becomes शोब and बात becomes बाओत.
| Standard Hindi | Punjabi | Tamil | Bengali |
|---|---|---|---|
| स्कूल | इस्कूल | इसकूल | इस्कूल |
| भाई | पाई | बाई | भाई |
| नहीं | नई | नई | नहीं |
| सब | सब | सब | शोब |
| बात | बात | बात | बाओत |
The vowel goes before the cluster or inside it. That single difference is the fastest way a listener places where someone is from.
Do not apply every accent rule to every word. If the rules are used too much, the character can start sounding exaggerated or fake. Using them only on selected words sounds more natural.
There is also a practical production problem. The regional-accent versions were longer than the neutral version. When matching dialogue to a fixed video, both the accent and the speaking duration need to be controlled.