Aarjan/nepali-voice-cloning-bench

Nepali Voice Cloning Benchmark

An open test of zero-shot voice cloning in Nepali. A system hears one short clip of a person it has never heard, then says a new Nepali sentence in that voice. We score how correctly it speaks, how much it sounds like that person, and how natural it sounds.

• Audio: this dataset
• Code, scorer and leaderboard: github.com/Naamche-Labs/nepali-voice-cloning-bench
• Built by: Naamche Labs, alongside Swar

Leaderboard

Clean track (prompts denoised, the main track)

#	System	CER ↓	Voice match ↑	UTMOS ↑
1	Swar (Naamche Labs)	4.55	0.691	2.67
2	OmniVoice (k2-fsa)	4.89	0.676	2.53
3	IndicF5 (AI4Bharat)	8.44	0.663	2.63
	Real recordings (ceiling)	9.77	0.740	2.24

Raw track (prompts as recorded)

#	System	CER ↓	Voice match ↑	UTMOS ↑
1	Swar (Naamche Labs)	4.58	0.694	2.39
2	OmniVoice (k2-fsa)	5.28	0.681	2.49
3	IndicF5 (AI4Bharat)	8.59	0.651	2.35
	Real recordings (ceiling)	9.77	0.740	2.24

Ranked by CER. Bold is best in each column. Scored 2026-10-25. The real recordings' CER is high because the recogniser also struggles with phone-quality speech; their voice match (a person against their own other clip) is the best possible.

Higher voice match and UTMOS are better; lower CER is better. Every system gets exactly the same prompt clips, transcripts and target sentences, and the same scorer (score.py). A failed or empty output is scored as fully wrong, never skipped.

The data

• Source: the Nepali (ne_np) test split of Google FLEURS (CC BY 4.0), read speech recorded in Nepal.
• 82 items from 4 speakers. FLEURS has no speaker labels, so test recordings were grouped into speakers by ECAPA voice embeddings (average-linkage clustering, cosine distance 0.6, within gender). For each speaker, one 4-12 s clip is the prompt and up to 8 of their other sentences are targets.
• Each item (items.jsonl) has the prompt clip and its transcript, the target sentence, and the real recording of that person saying the target sentence.
• The real recordings are the ceiling. Scored the same way, they show the best voice match possible (a person against their own other clip).

Two tracks

FLEURS clips are recorded on phones and are noisy, and cloning models copy the noise.

• Raw track: the prompt clips exactly as FLEURS recorded them (prompts/).
• Clean track: the same clips denoised with DeepFilterNet3, attenuation capped at 18 dB (prompts_clean/). The cap was chosen on FLEURS validation speakers, which are disjoint from these.

Report the track you used. Do not apply other preprocessing to the prompts without saying so.

Scores

score.py computes three numbers over all 82 items:

Score	How	Better
CER	Character error rate (%) of facebook/mms-1b-all with the Nepali adapter (Latin output masked) against the target text. Digits spelled out on both sides; punctuation and spacing ignored.	Lower
Voice match	Cosine similarity of ECAPA speaker embeddings between the output and the real recording of the same person saying the same sentence.	Higher
UTMOS	Predicted naturalness, UTMOS22 strong via SpeechMOS v1.2.0, 1-5.	Higher

All three are automatic. UTMOS was trained on English, and FLEURS audio is phone quality, so absolute UTMOS values are low for everyone; compare systems, not absolute values.

Submit a system

1. Download the dataset: huggingface-cli download Aarjan/nepali-voice-cloning-bench --repo-type dataset --local-dir data
2. For every line of data/items.jsonl, clone the voice of prompt_audio (raw track) or prompt_audio_clean (clean track) using prompt_text as its transcript if your system takes one, and say target_text. Save it as <id>.wav.
3. pip install -r requirements.txt, then python score.py --data data --audio YOUR_DIR --name YOUR_SYSTEM --track raw.
4. Open a pull request adding results/YOUR_SYSTEM.json and a row to the leaderboard, with a link to the model and the exact settings used. Uploading the audio as well (for example to a Hugging Face dataset) lets others check it.

How the listed systems were run

• Swar (Naamche Labs): final model, NFE 16, CFG 2.0, seed 0, one take per item, pace matched to the prompt.
• OmniVoice (k2-fsa/OmniVoice, 0.6B): default settings, language="npi", prompt transcript given.
• IndicF5 (ai4bharat/IndicF5), unchanged: NFE 16, CFG 2.0, seed 0, the same pacing as Swar.

Outputs of every listed system are in the dataset under outputs/<track>/<system>/.

Licence

Code: MIT. Audio and text derived from FLEURS: CC BY 4.0 (Google). System outputs: released under CC BY 4.0 by Naamche Labs for checking the results.

Citation

@misc{nepalivcbench2026,
  title  = {Nepali Voice Cloning Benchmark},
  author = {Chaudhary, Aarjan},
  year   = {2026},
  note   = {Naamche Labs. https://github.com/Naamche-Labs/nepali-voice-cloning-bench}
}

FLEURS: Conneau et al., "FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech", SLT 2022.