Frequently Asked Questions

Yes - uploads are capped at 2 GB per file. Separately, dubs are also metered by video duration: new accounts get 60 minutes of free dubbing time, shown as a live countdown badge in the header, and some videos are marked “fully free” and don’t draw against your allowance or credits at all.

Very large files may still hit browser upload timeouts even under 2 GB. For those, or for anything programmatic, use the Dubbing API instead of the dashboard.

• Clone speaker voice - zero-shot vocal trait tracking from the original speaker. Use this when audience recognition of that specific speaker matters (a host, a founder, a recurring instructor).
• Use Sarvam voices - native artist personas. Faster to review since the persona is already familiar and consistent across projects, but doesn’t carry the original speaker’s identity.

See Creating a Dub for where this is set.

Five options, chosen per dub: Auto (default, picks automatically), Formal (official/institutional content), Urban colloquial (trend-aware, casual), Standard colloquial (everyday spoken language), and Traditional colloquial (preserves local linguistic and cultural nuance).

Style changes register and word choice, not just vocabulary - a serious lecture translated in Urban colloquial style will read as too casual, even if every fact is correct. Leave it on Auto unless you have a specific reason not to.

Diarization isolates overlapping audio waveforms. For clean results, manually specify the exact speaker count during project setup - if you’re unsure, round your estimate upward rather than down. Undercounting merges two real speakers into one voice, which is much harder to fix afterward than simply merging an unused extra slot.

Yes. The Dubbing editor supports per-block correction: edit the text in that specific block, then click Regenerate audio - only that block re-renders, not the full timeline. See Editing a Dub.

Yes. The Dubbing editor has three independent volume sliders - Dubbed, Original, and Background - so you can blend the new voice track against the source instead of fully replacing it. Useful when you want ambient sound or background music from the source to carry through.

A live progress card opens automatically after you click Create dub, showing each pipeline stage (processing, transcribing, translating) with timestamps and durations. You don’t have to wait on this screen - click Go to Projects to navigate away; the dub keeps processing in the background and shows as “Review Ready” once it’s done.

Use the language pill in the editor header to switch between target languages. Each language keeps its own independent transcript and edits, and you can export each one separately once it’s ready - no need to wait for every language to finish review before exporting any of them.

A few things to try, in order:

1. If the same term is wrong in multiple places, use Find & Replace in the editor instead of fixing each occurrence by hand.
2. If it’s a one-off line, edit the translated text directly in that block, then click Regenerate audio - only that block re-renders.
3. If the whole dub reads as too casual or too formal, the issue is likely the translation style, not a specific word. Revisit the style dropdown on a fresh dub - see Which translation style should I use? above.

Check the speaker strategy first: this only happens with Clone speaker voice, not Use Sarvam voices (which intentionally uses a native artist persona instead). If Clone speaker voice was selected and it still doesn’t land, the likely cause is source audio quality for that speaker’s segments - background noise or overlapping speech in the original carries through into the clone.

Credit cost is based on character count, estimated live as you type - for example, the interface shows “327 characters · ~0.98 credits” for a sample paragraph.

Yes. The Cloned voices tab sits right next to Sarvam voices in the persona picker. If you don’t have a clone yet, that tab shows a Clone a voice button that takes you straight into the cloning flow without leaving Text to Speech. See Generating Speech.

It’s a delivery-pace multiplier (shown as e.g. “1.00x”), and it changes more than playback speed - it changes perceived tone. Values below 1.00x read as more deliberate and authoritative (good for narration); values above 1.00x read as more energetic and casual (good for ads and social content).

In order:

1. Language mismatch is the most common cause. Confirm the output language dropdown matches the language your text is actually written in.
2. Check punctuation around the word. Missing commas or periods can merge it awkwardly with neighboring words.
3. Spell out ambiguous numbers or abbreviations if you notice mispronunciation - some readings have more than one valid pronunciation.
4. Try a different voice. Personas can differ slightly in how they handle uncommon names or loanwords.

Sarvam voices are curated, ready-to-use personas. Cloned voices are ones you’ve created yourself via Voice Cloning, and require your explicit consent that you hold rights to the voice being cloned.

My voices is your personal set - favorited and cloned voices. Explore voices is the full browsable catalog. Both live under the same filters (use case, gender) and search box in the Voice Library. See Browsing and Selecting Voices.

No - cloning is a live in-browser recording, not a file upload. You pick a language, read a short provided passage aloud (Shuffle for a different one), and record directly for about 10 seconds. See Creating a Clone.

A separate setting from the personality tags shown in the Voice Library - think of it as a hint for how the clone should be used, not a fixed label. Options are General (default), Conversational, Professional, Warm, and Energetic.

Yes. The final step plays back your recording and lets you hear the clone on three sample lines in different registers - casual, news-anchor, and storytelling - before you confirm consent and save. If you’re not happy, Re-record is available right up to that last step.

Yes, and it’s mandatory, not a formality - you can’t save a clone without checking the consent box confirming you have the rights to record, process, and use that voice for generating new content.

A cleaner recording almost always beats tweaking anything after the fact:

1. Re-record in a quiet room with no music playing - this is the product’s own stated requirement, not just a suggestion.
2. Read at your normal pace. Rushing a short recording tends to distort natural speech patterns.
3. Use Re-record before finalizing, on either Step 1 or the final Preview step - don’t settle for a take you’re unsure about.
4. If a passage feels awkward to read aloud, click Shuffle for an easier one. A clean read of an easy passage beats a shaky read of a hard one.

MP3, WAV, M4A, FLAC, OGG, and AAC, up to 500 MB per file. Video files work directly too - no need to extract the audio track yourself first.

Yes - click or record audio on the upload screen to capture audio directly in the browser. There are also three built-in example clips (“Gujarati finance news,” “Marathi cricket commentary,” “Tamil healthcare conversation”) if you want to see a real transcription before committing your own recording.

Yes, and it transcribes as-is - a real example output includes Gujarati text with embedded English financial terms like “IDBI bank,” “share,” and “disinvestment.” This is normal, expected behavior, not an error.

Provide an accurate speaker count if you know it - for a roundtable or panel recording, this is often the single biggest factor in whether the transcript correctly separates who said what. If unsure, round upward: merging an unused extra slot is a smaller error than splitting an undercounted one. Cross-talk and overlapping speech are a known hard case for diarization generally.

In order:

1. Confirm the language setting matches what’s actually spoken. This affects the model’s phonetic expectations for the whole recording, not just word choice.
2. Check the speaker count, especially for multi-speaker recordings - an inaccurate count is the most common cause of garbled or merged segments.
3. If the errors cluster around overlapping speech, that’s expected - cross-talk is a known hard case for diarization generally, not specific to your recording.
4. If English terms appear inside regional-language text, that’s correct behavior, not an error - code-mixed speech transcribes as heard.

Sarvam’s infrastructure follows standard institutional data-security practices. For specifics on data residency or retention relevant to your use case, check with your Sarvam account contact - commitments here can vary by plan and deployment.