> For clean Markdown of any page, append `.md` to the page URL.
> For a complete documentation index, see https://docs.sarvam.ai/llms.txt.
> For full documentation content in one file, see https://docs.sarvam.ai/llms-full.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.sarvam.ai/_mcp/server.

# Dubbing Best Practices

Practical guidance for getting a clean dub on the first pass, organized by where in the workflow it applies.

## Before You Upload

> **Note**
>
> A clean source recording matters more than any setting downstream - noise or overlapping speech in the original carries through into the dub.

* Count your speakers accurately before upload - this drives diarization quality more than any other single setting.
* When unsure of the exact speaker count, round upward. Merging an unused extra slot is a smaller error than blending two real speakers into one.
* Match **Source language** to what's actually spoken - this affects transcription accuracy, not just translation.

## Choosing a Speaker Strategy

* Choose **Clone speaker voice** when audience recognition of the specific speaker matters - hosts, founders, recurring instructors.
* Choose **Use Sarvam voices** when you want a fast, consistent persona across many pieces of content and the original speaker's identity isn't the point.
* Leave the style dropdown on **Auto** unless you have a specific reason not to - it picks a style automatically.
* Override it to `Formal` for official or institutional content, or to `Urban colloquial` for trend-aware, casual content - a serious subject translated in `Urban colloquial` style will undersell it, and vice versa.

## Reviewing Output

> **Note**
>
> Review multi-speaker dubs speaker-by-speaker, not line-by-line - it's the fastest way to catch a diarization error where two voices were merged into one.

* Spot-check pronunciation on names, brand terms, and technical vocabulary - these are the most common source of small errors in an otherwise good dub.
* Use per-block regeneration for isolated fixes. Re-running the whole file for a one-line correction wastes review time on lines you've already approved.
* If a dub's tone feels flat or overly formal, revisit the style dropdown before assuming the voice model is at fault - style has an outsized effect on perceived tone.
* For a recurring term or name that's consistently mistranslated across blocks, use **Find & Replace** in the editor instead of fixing each occurrence one by one.

## Balancing the Final Mix

> **Note**
>
> The editor's three volume sliders - **Dubbed**, **Original**, and **Background** - let you blend rather than fully replace the source audio.

* Keep some **Background** volume when the source has ambient sound or music you want to preserve - muting it entirely can make a dub feel sterile compared to the original.
* Lower, rather than mute, **Original** if you want a subtle sense of the source voice underneath the dub; mute it fully for a clean, dub-only listen.

## Managing Multi-Language Projects

* If you dubbed into several target languages at once, use the language pill in the editor header to review and fix each one independently - edits in one language don't affect the others.
* Export each language separately once you're satisfied with it, rather than waiting for every target language to be fully reviewed before exporting any of them.