releases
Every public Ampixa release: demos, sites, models, datasets and code. Newest first where the date is verified (the project's earliest independently verified public date, not necessarily the latest tag); the rest are undated rather than guessed. Links go straight to where each one lives.
Dated
Yak tokenizer
An open, lossless, Nepali-first tokenizer. Its phrase build is #1 lossless on Nepali of the 42 we measured (1.267 tokens/word), and v2.0 is the lowest-cost lossless tokenizer on 18 of 19 languages of Nepal.
barga
A decision model: given a state and a set of candidate options, it returns a probability for each option in one forward pass.
NepTel
A 75-segment, 2,375-word Nepali call-centre ASR benchmark with every system output published.
sanoTTS e12nano
A 294,642-parameter browser/microcontroller TTS voice; RTF 0.383 on an ESP32-S3, 30 browser voices across 16 languages.
Nepali MFA
Forced-alignment toolkit and tiered manifests for Nepali ASR training data; 509.5 hours represented, 934 of 1,000 TextGrids aligned.
NepaliG2P
A comparison of neural and rule-based Nepali grapheme-to-phoneme conversion.
Undated
- —
Audio8 Nepali + Astadhwani
Nepali TTS packaged as the Astadhwani desktop app, exported to ONNX.
- —
Barnamala
A 1.11M-parameter distilled Devanagari character classifier; not detectably different from its 17.32M-parameter teacher ensemble on DHCD (p = 0.345) — not proven equal, but not shown worse.
- —
Bemine
The blog pipeline: review, style checks, deploy and the git mirror behind ampixa.com/blog.
- —
Dhimal G2P
Rule-based and neural grapheme-to-phoneme models for Dhimal.
- —
Gurung frontend
A Gurung text-to-speech frontend.
- —
Jyoti codec
A speech codec kept and published as a negative result: it damages Nepali ASR accuracy.
- —
Kala TTS
A Nepali VITS-based text-to-speech model with a live demo.
- —
Language toolkits
A shared toolkit registry for the languages of Nepal.
- —
Limbu TTS
A Limbu text-to-speech Hugging Face Space.
- —
LimbuOCR
OCR for the Limbu (Sirijonga) script; 0.474% codepoint CER, 0.979% grapheme CER, 85.9% exact lines of 902.
- —
Listening test
A live Nepali TTS listening-test / MOS collection tool at ampixa.com/listen.
- —
Nepal TTS training
Training code and configs for Nepal TTS models.
- —
nepal-ttf2utf
Hash-pinned conversion maps from legacy fonts to Unicode across 10 scripts of Nepal and its neighbours.
- —
Rani
Two from-scratch 121M-parameter Nepali ASR encoders (offline + streaming); 36.3% WER on NepTel (corrected from an earlier, unreproduced 33.8%).
- —
NepTTS-Bench + Arena
A live Nepali TTS benchmark and rating arena; 205 synthesis prompts plus 160 references, 6,962 ratings from 180 raters, 5 approved submission reports.
- —
newsTTS
A live newsroom and reading-room Nepali speech demo.
- —
NSL fingerspelling poses
Whole-body pose sequences for the Nepali Sign Language manual alphabet.
- —
Pocket TTS bilingual teacher
A ~340M-parameter bilingual Nepali/English TTS teacher; English WER 3.5%, Nepali CER 12.3%, CPU RTF ≈ 0.45.
- —
proper-voices
A consent-gated Nepali speech recording app; native-speaker review due 2026-11-11.
- —
Qwen bold neuron
Code and findings from probing a single neuron inside Qwen3.5-2B.
- —
Ratna streaming ASR
A login-gated live streaming ASR demo.
- —
Real Nepali TTS
A Nepali text-to-speech Hugging Face Space.
- —
Shruti
A live Ampixa page at ampixa.com/shruti.