How to Make a Video With a Marathi, Bengali or Gujarati Voiceover
You can narrate a video in 11 Indian languages, not just Hindi: write the script in the target language, pick a voice recommended for it, set the pace to about 0.9, and attach the audio to your video. Here is the voice-by-language table and the workflow.

Short answer: you can narrate a video in 11 Indian languages with one text-to-speech model, Bulbul v3: Hindi, English (India), Bengali, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati, Punjabi and Odia. The workflow is the same for all of them. Write the script in the target language, choose a voice recommended for that language, set the pace to about 0.9, generate the audio, and lay it over your video or slides. The main thing that changes between languages is which voice you pick.
Most guides stop at Hindi and English. If your audience is in Pune, Kolkata or Ahmedabad, a Hindi voiceover is often not the right answer, and recording a human narrator for each language is slow and expensive. This guide covers what the model supports, which voices fit which language, and the exact steps.
Which Indian languages does AI text-to-speech cover?
The Bulbul v3 speech model lists these 11 languages, each with its own language code:
| Language | Code | Voices recommended for it |
|---|---|---|
| Hindi | hi-IN | Priya, Suhani, Shubh, Ashutosh |
| English (India) | en-IN | Ishita, Ratan |
| Bengali | bn-IN | Roopa, Suhani, Rehan |
| Tamil | ta-IN | Ritu, Ishita, Rohan, Ratan |
| Telugu | te-IN | Priya, Neha, Shubh, Ratan |
| Kannada | kn-IN | Neha, Ishita, Shubh, Ratan |
| Malayalam | ml-IN | Pooja, Shubh |
| Marathi | mr-IN | Ritu, Priya, Ratan |
| Gujarati | gu-IN | Ritu, Priya, Ratan |
| Punjabi | pa-IN | Roopa, Suhani, Mani |
| Odia | od-IN | Ritu, Pooja, Shubh |
The model offers 36 voices in total, 14 female and 22 male. The voices not named above have no language recommendation attached, so treat them as general-purpose and audition them before committing. Arcframe also offers other speech models for other needs, but this is the one built for the Indian languages above.
How do you make the video, step by step?
- Write the script in the target language. Do not paste English and hope the voice translates it. The model reads what you give it. For a short explainer, 100 to 150 words per minute of finished audio is a sensible starting budget.
- Write it the way it should be spoken. Mixed-language scripts read naturally, for example Marathi with English product names left in English. Spell out numbers and abbreviations the way a narrator would say them.
- Pick the language and a recommended voice. For a Marathi explainer, start with Ritu, Priya or Ratan. For Bengali, start with Roopa, Suhani or Rehan. For Gujarati, Ritu, Priya or Ratan.
- Set the pace. The pace control runs from 0.5 to 2.0, and 1.0 is natural. 0.9 is a good default for narration, and 1.1 to 1.2 suits news-style reads. Leave expressiveness at the default of 0.6 unless the take sounds flat or unstable: lower is steadier, higher is livelier but can slip.
- Generate and listen. Audition one paragraph first. If a word sounds wrong, respell it phonetically in the script rather than regenerating the whole file.
- Attach the audio to your video. Add it to an existing clip as a voiceover, or start from a PowerPoint, PDF or a single image and let the narrated video be built around the audio.
What does it cost?
Speech with this model costs 2 credits per 1,000 characters of script, whatever the voice or settings. That makes a 1,000-character paragraph 2 credits, and a script's price is easy to estimate before you generate anything. The free tier gives you 20 one-time credits with no card, which is enough to test a few languages. For current plan prices and top-ups, see the pricing page.
Which voice for which job?
- Training and e-learning: choose a steady voice and keep expressiveness low so a 20-minute module stays consistent from the first slide to the last.
- Product and e-commerce listings: a mid-paced voice at 0.9 to 1.0, with English product names left in English.
- News-style updates: raise the pace to 1.1 to 1.2.
- Stories and drama: one voice, Varun, is described as deep and dramatic and is kept for suspense-style reads.
What does this not do?
Being straight about limits will save you a re-render:
- It does not translate for you. You supply the script in each language. Have a native speaker read it before you publish, especially for formal, legal or medical content.
- This model does not speak in your cloned voice. Cloned voices are a separate feature and work with a different speech model. If you need the same voice across languages, see the guide on dubbing a video into another language with the same voice.
- Names, acronyms and rare words can be mispronounced. Respell them phonetically in the script and re-listen.
- Quality varies by language and voice. The recommendations above come from the voice list itself, not from a benchmark we ran, so audition two or three voices on your own script.
Can you do this from Claude, ChatGPT or Cursor?
Yes. Arcframe has an MCP server, so an AI assistant can list the voices for a model, generate the narration and attach it, all on your own account and credits. That is useful when you want to batch the same script across Marathi, Bengali and Gujarati and compare the takes side by side.
Quick checklist
- Script written in the target language, with spoken-style numbers.
- Language code set correctly (mr-IN, bn-IN, gu-IN and so on): the language is required.
- A voice recommended for that language, auditioned on one paragraph.
- Pace near 0.9 for narration.
- A native speaker checked the final audio before it goes public.
Start with a single 30-second paragraph in your target language. It costs well under 1 credit, takes seconds to generate, and tells you whether the voice fits your audience before you commit to a full video.