INSIDE THIS AI SKILL
The craft behind its decisions.
These are the principles and procedures the skill applies. You can understand the reasoning without having to perform every production task yourself.
Matching a voice to a finished performance is a losing game.
Cast a Hindi lead, produce five episodes, then source Tamil, and you are no longer casting a character - you are trying to impersonate a performance in a language it was never built for.
01 / Voice direction
Voice is identity, not delivery
An audience recognises a character by voice faster than by face, particularly on a phone at low volume with the screen half-covered by a thumb. A voice that shifts between episodes is the audio equivalent of a recast, and it is noticed just as quickly.
So the approved voice reference is canon: attached to the character, inherited by every shot, and changed only through review. What it is not is a per-scene delivery choice, which is a separate and much cheaper decision.
- Who this character sounds like
- Approved against every declared track
- Pronunciation of names and places
- Register: who they speak to formally
- Changed only through a scoped canon change
- Pace and emphasis for this moment
- Emotional state inside the approved range
- Where the line breathes
- Volume and proximity
- Adjusted freely, shot by shot
02 / Voice direction
A casting process that survives three tracks
- 01Declare every language track at greenlight
Including the one you are only fairly sure about. Declaring a track you later drop costs a casting session; adding one you did not declare costs the identity.
- 02Audition the same lines in every declared language
Not a translation of the audition - the actual lines the character says, in each track.
- 03Test the hard sounds first
Character names, place names, and any word the season repeats. These are where synthetic voice most reliably gives itself away.
- 04Choose for consistency, not for the best single track
A voice that is excellent in Hindi and unusable in Tamil loses to one that is good in both, because the show ships in both.
- 05Record pronunciation guidance as canon
Names, places, regional phrasing and register, per track. Checked in QA rather than re-decided per session.
- 06Name a native reviewer per track
A gate of their own. A Hindi approver cannot sign for Tamil, and asking them to is how localization quality quietly degrades.
03 / Voice direction
The permission rule
Questions about this skill
Can I direct a voice without acting or audio experience?
Yes. Describe how the character should feel or sound. The Voice Direction Skill translates that intention into delivery, rhythm, pauses and pronunciation notes for your AI Director. Permission for a voice and review of the agreed language track still belong to the production workflow.
Can one voice really work across three Indian languages?
It is the thing we test at casting rather than assume. Some voices carry across and some do not, and finding out before production is the entire reason tracks are declared at greenlight. Where a voice cannot carry, the honest answer is a different voice, chosen knowingly.
What happens if a name is pronounced wrong in a released episode?
Pronunciation is canon, so the fix is a canon correction plus a targeted repair of the affected audio segments. The picture, the other tracks and the rest of the episode are untouched.
Who directs the performance?
You do, within the approved expression range and the treatment. Delivery is directed per scene and adjusted freely; identity is canon and changes only through review.
Does lip-sync tolerance affect casting?
It affects everything downstream. A treatment demanding tight sync narrows which providers can serve a shot and raises attempts per accepted shot in every language, so it is agreed at the treatment stage rather than discovered per track.
Can I use a human performer instead?
Where you have the rights and the budget, yes, and for a lead in a character-driven genre it is often the better answer. Tosheo is not a replacement for performers, and the rights position for a human performance is recorded like any other grant.