Make a Nepali text-to-speech voice
Published 12 October 2026 · prices on this page are live
Few text-to-speech models speak Nepali well, and the good commercial voices are usually locked inside someone else's paid service. With an open model, public Nepali recordings and a rented GPU, you can train a voice of your own in a few hours of GPU time.
Three ways to get a Nepali voice
- Use a ready-made model. AI4Bharat's Indic Parler-TTS (ai4bharat/indic-parler-tts on Hugging Face) lists Nepali among its languages and is Apache-2.0 licensed; accept its terms on Hugging Face to download it, and try it on your own sentences before you build on it.
- Fine-tune a multilingual model. Chatterbox from Resemble AI is MIT-licensed and expressive. Its multilingual release covers Hindi but not Nepali; because both are written in Devanagari, starting from the Hindi model and fine-tuning on Nepali speech works well. We trained a clear, natural-sounding Nepali voice this way on about nine hours of speech.
- Train a model from scratch, such as VITS. It usually needs many hours of one clean speaker and more GPU time, but nothing is inherited from anyone else's model.
Nepali speech for training
| Dataset | What it is | Size | Licence |
|---|---|---|---|
| OpenSLR 43, High quality TTS data for Nepali | Studio recordings of female speakers, with transcripts | About 800 MB of audio | CC BY-SA 4.0 |
| AI4Bharat IndicVoices-R (Nepali) | Cleaned, studio-quality speech from many speakers | About 104 hours | CC BY 4.0; accept the terms on Hugging Face first |
| AI4Bharat IndicVoices (Nepali) | Natural, unscripted speech | About 106 hours | CC BY 4.0; gated the same way |
For one consistent voice, pick clean clips from a single speaker, or a few similar ones, 3 to 15 seconds long, with exact Devanagari transcripts. If you copy a real person's voice, get their written consent first.
Which GPU to rent
Fine-tuning a model of Chatterbox's size, about half a billion parameters, with LoRA fits comfortably on a 24 GB card: our Nepali voice trained on an RTX 3090 Ti. Today's lowest prices:
| GPU | Memory | USD | INR | NPR |
|---|---|---|---|---|
| RTX 3090 | 24 GB | USD 0.24 | INR 24 | NPR 37 |
| RTX 3090 Ti | 24 GB | USD 0.33 | INR 32 | NPR 50 |
| RTX 4090 | 24 GB | USD 0.63 | INR 61 | NPR 96 |
| L40S | 45 GB | USD 0.97 | INR 94 | NPR 148 |
Step by step (Chatterbox)
- Rent a 24 GB GPU with the PyTorch + JupyterLab software and about 100 GB of disk.
- Prepare a folder of WAV clips and a metadata file that maps each file to its transcript.
- Start from the Hindi weights. Extend the text vocabulary so every Devanagari character Nepali uses has its own token, then train LoRA adapters together with the new text embeddings. Community fine-tuning toolkits for Chatterbox handle data loading and the training loop.
- Listen; do not just watch the loss. Generate the same test sentences from each checkpoint and keep the one native speakers prefer.
- Download the adapter, a few hundred megabytes, or push it to a private Hugging Face repository before you destroy the machine.
Speed: fine for narration, slow for live calls
On an RTX 3090 Ti our fine-tuned voice took about 1.2 to 1.5 seconds to produce each second of speech. That is fine for narration, videos and voice notes, but too slow for a live phone agent, which needs a faster model variant, a faster GPU, or both.
What it costs
A working session of data preparation, training and listening tests takes a few hours. Three hours costs, on today's cheapest hosts:
| GPU | Memory | USD | INR | NPR |
|---|---|---|---|---|
| RTX 3090 | 24 GB | USD 0.72 | INR 72 | NPR 111 |
| RTX 4090 | 24 GB | USD 1.89 | INR 183 | NPR 288 |
| L40S | 45 GB | USD 2.91 | INR 282 | NPR 444 |
Questions
Is there a free Nepali text-to-speech model?
Yes. AI4Bharat’s Indic Parler-TTS lists Nepali among its languages and is Apache-2.0 licensed. Fine-tuning your own gives you control over the voice itself.
How much audio do I need to fine-tune a Nepali voice?
Our voice used about nine hours of clean Nepali speech. Less can work when you start from a model that already reads Devanagari, such as the Hindi Chatterbox model.
Can I clone someone’s voice?
Technically yes, from a short reference clip. Legally and ethically, only with that person’s clear permission.
Can I use the voice commercially?
Chatterbox is MIT-licensed, and CC BY data allows commercial use with attribution. Read each dataset’s licence; this is not legal advice.