---
license: cc-by-sa-4.0
language:
- ne
- en
pipeline_tag: text-to-speech
tags:
- text-to-speech
- voice-cloning
- nepali
- f5-tts
base_model: ai4bharat/IndicF5
---

# Swar: open Nepali voice cloning

Swar speaks Nepali the way it is spoken in Nepal and clones a voice from 20 to 30 seconds of someone talking. It reads Nepali, English, and Nepali mixed with English. By [Naamche Labs](https://github.com/Naamche-Labs).

**Try it in your browser, no setup:** [Swar Playground](https://naamchelabs.com/playground): nine ready voices, or clone your own from 30 seconds of speech, free to try. **Run it yourself:** [Colab notebook](https://colab.research.google.com/github/Naamche-Labs/swar/blob/main/notebooks/Swar_Clone_Colab.ipynb) (free T4 GPU). **Code and paper:** [github.com/Naamche-Labs/swar](https://github.com/Naamche-Labs/swar).

## Files

- `podvoices_ema.pt`: **the default model** (F5-TTS Base architecture, EMA weights). Every result below is this file.
- `swar_v6_ema.pt`: a second checkpoint for English-heavy text, trained further with 96 h of English (LibriTTS-R). Mixed Nepali-English errors 5.24% (default 5.6%), but pure Nepali is weaker (NepTTS-Bench pronunciation error 4.04%, FLEURS cloning errors 5.41%). Use it with `ENGLISH_LATIN=1`.
- `vocab.txt`: character vocabulary (IndicF5).
- `kriti_onnx/`: Kriti, our Nepali speech recogniser (MIT), used to transcribe uploads and check takes.
- `presets/`: ready voices Maya (female) and Aarav (male), built only from model-made audio.
- `app/`: the cloner app (Gradio) used by the notebook.

## Results (NepTTS-Bench, 205 sentences, one take)

| System | Character error | Pronunciation error | UTMOS |
|---|---|---|---|
| Swar, studio voice | **6.19** | **2.92** | 3.82 |
| Swar, Maya voice | 7.45 | 4.52 | **4.09** |
| OmniVoice (same prompt) | 6.68 | 4.32 | 3.93 |
| Indic-Speak Srijana | 10.25 | 8.12 | 3.97 |
| ElevenLabs v3 | 15.70 | 12.91 | |

[Nepali Voice Cloning Benchmark](https://github.com/Naamche-Labs/nepali-voice-cloning-bench) (our open benchmark on FLEURS, v1.1, clean track, 64 distinct prompt + sentence pairs from 4 unseen voices): Swar errors 4.72% and voice match 0.697, OmniVoice 5.08% and 0.683, IndicF5 8.30% and 0.669.

Nepali mixed with English (30 held-out sentences): Swar 5.6% character errors, OmniVoice 11.4%. UTMOS is an automatic predictor, not a listener test.

## Training data

Initialised from ai4bharat/IndicF5 (MIT). Fine-tuned on AI4Bharat Rasa and IndicVoices-R (CC BY 4.0), Google FLEURS (CC BY 4.0), OpenSLR 43 and 54 (CC BY-SA 4.0), LibriTTS-R (CC BY 4.0), and podcast speech used with the speakers' permission (not released). The weights are released under CC BY-SA 4.0 to respect the share-alike data.

## Responsible use

Only clone a voice with that person's permission. Do not use Swar to impersonate anyone, to deceive, or to commit fraud.