Swar स्वर

NepTTS-Bench · 205 sentences · one take per system

The lowest pronunciation error on NepTTS-Bench.

NepTTS-Bench is the public Nepali text-to-speech benchmark by Ampixa: minimal pairs, questions, tongue twisters and short phrases built to trip a TTS up. A speech recogniser writes down what each system said. Pronunciation error counts only the mistakes a listener would hear. Lower is better.

Scored with the benchmark's own recogniser (Meta MMS), identically for every system. Swar rows use the studio voice (AI4Bharat Rasa, CC BY 4.0). Our Nepali training cut IndicF5's pronunciation error by more than two thirds.

Nepali Voice Cloning Benchmark · FLEURS

Number one on the Nepali voice cloning benchmark.

Each system clones four people recorded in Nepal from one short clip each, then says 82 new sentences. Clean track: Swar has the fewest errors (4.55 against 4.89 for OmniVoice and 8.44 for IndicF5) and the closest voice match (0.691 against 0.676 and 0.663). This is our own open benchmark; the scorer and all the audio are public.

Nepali mixed with English · 30 held-out sentences

Half the errors on everyday mixed speech.

Real Nepali has English in it: "मेरो laptop को battery...". On 30 sentences like that, written before any output was heard, Swar makes 5.6% character errors against 11.4% for OmniVoice and 14.6% for IndicF5.

NepTTS-Bench official toolkit

Top scores on the benchmark's own scorer.

With NepTTS-Bench's neptts-eval toolkit, Swar's Maya voice has the highest SCOREQ naturalness score in the benchmark's published results (4.57, ElevenLabs v3 4.45), and the studio voice the lowest Whisper round-trip error (0.329, Gemini 0.371). Both are automatic scores, not listeners.

How it works

Most of the gain needs no extra training.

A text frontend writes Nepali as it is spoken in Nepal and English the way Nepali speakers say it. Enrolment turns a casual voice note into a clean prompt with a transcript from Kriti, our Nepali speech recogniser. Then Swar makes several takes and keeps the one whose pronunciation checks out, before it looks at voice match.

Model

F5-TTS, taught Nepali from Nepal.

F5-TTS Base initialised from AI4Bharat's IndicF5 (MIT), trained on open Nepali speech (Rasa, IndicVoices-R, FLEURS, OpenSLR 43 and 54), English (LibriTTS-R), and podcast speech used with the speakers' permission.

Use it responsibly

Only clone a voice with permission.

Clone a person's voice only with their consent. Do not use Swar to impersonate anyone or to deceive. The ready-made voices are built only from model-made audio; no podcast recording is released.

At a glance