Nepali TTS: listen and compare

VoxCPM2 + our Nepali LoRA (ampixa/voxcpm2-nepali) next to Swar and other systems, on the same inputs. The number under each player is that clip's character error rate (lower is better) from the scorer named in each section. Our system does not beat Swar overall; this page lets you hear where it is better and where it is worse.

How to read the best-of-8 rows. "Best-of-8" makes 8 takes and keeps the one our internal ASR (LONG1C) transcribes best. On NepTTS-Bench and the Swar bench the scorer is MMS, a different model, so the pick is not graded by itself. On Fresh v1 the scorer is LONG1C, the same model that picked the take, so the best-of-8 Fresh numbers are optimistic by construction; compare the single take there. Swar rows are one take.

Sources and licences

NepTTS-Bench sentences: Ampixa/neptts-bench (MIT), 205 scored sentences. Metric: Swar's own scorer (Naamche-Labs/swar @4d2a7406, tools/rescore_deva.py + tools/pron_cer.py, MMS-1b-all with Latin tokens masked). Run here on Swar's published audio it gives 6.21 / 7.45 (published 6.19 / 7.45).
Swar audio: Naamche Labs, Aarjan/swar (weights CC BY-SA 4.0); NepTTS-Bench Swar clips are their published files (Aarjan/swar-neptts-bench, CC BY-SA 4.0). Maya prompt: Swar preset (model-made audio).
Studio prompt: AI4Bharat Rasa (CC BY 4.0). Swar bench cloning prompts: Google FLEURS (CC BY 4.0), via Aarjan/nepali-voice-cloning-bench.
NepTTS-Fresh v1: our held-out set (frozen 2026-10-31): 12 speakers from public Nepali YouTube videos (short prompt and real clips), plus our own written sentences. Swar's training data includes FLEURS and Rasa, so those prompts likely favour Swar; Fresh v1 audio is new to every system.
A450 is another of our models (Audio8 after DPO, step 450), shown for reference; it is not released.
Audio here is MP3 (mono, 24 kHz, 64 kbit/s) for listening; every score was computed on the original WAV files.