raw Nepali audio infrastructure — speech in, speech out
new NepaliConformer — Nepali speech recognition that holds up on real phone calls, with the benchmark to prove it. try it →
raw Nepali audio infrastructure — speech in, speech out
new NepaliConformer — Nepali speech recognition that holds up on real phone calls, with the benchmark to prove it. try it →
We are ampixa labs — a Nepal-based audio lab working on the raw signal: speech before it becomes text, and speech after text becomes voice.
We build text-to-speech, ASR data pipelines, forced alignment, pronunciation systems, human evaluation, and low-resource speech tooling for Nepali and regional languages.
Audio is messy. It carries accent, breath, noise, code-switching, hesitation, dialect, bad microphones, and the parts of language that plain text throws away. That is the layer we care about.
Sovereign AI for Nepal is not one giant model with a flag on it. It is Nepali data, speech corpora, lexicons, benchmarks, ASR, TTS, alignment, evaluation, and the systems around them.
Nepal does not live in English. It lives in spoken Nepali, mixed Nepali, regional languages, phone calls, classrooms, radio, clinics, government counters, farms, cities, and homes. If AI cannot understand that audio, it is not useful infrastructure here.
The bridge is speech: ASR that can listen, TTS that can answer, alignment that can clean data, pronunciation tools that know the language, and evaluation that uses native listeners instead of imported assumptions.
The future is audio. Ampixa exists to make Nepal audible to machines, and to make machines speak back without forcing Nepal through English first.
Everything below is public — code, weights, datasets, or a demo you can open right now. Numbers link to the repository that shows how they were measured.
Speech recognition for Nepali as it is actually spoken. Two 121M-parameter Conformer models — offline and cache-aware streaming — trained from scratch on ~1,655 hours of mostly conversational Nepali. 33.8% WER on real call-centre audio, against 40.6 for the next-best open system and 96.3 for zero-shot Whisper-large-v3 on the same audio and the same scorer.
A public benchmark of real, spontaneous Nepali — genuine call-centre conversations, 75 segments and 2,375 words, drafted by Chirp 2 and reviewed word-by-segment by a native speaker. Ungated, no login, with per-system outputs published so every number can be re-derived. We would like to be beaten on it.
The data-quality layer between raw Nepali audio and ASR training: forced alignment, pronunciation dictionaries, review dashboards, and the machinery that turns human review labels into clean, noisy and rejected manifests. About 509 hours of MFA-prepared source data exported so far.
सानो = “small”. Neural voices of 745k to 1.8M
parameters — nine voices across six languages (English, Nepali,
Hindi, Vietnamese, Indonesian, Chinese). Under 4 MB per voice, pure numpy
inference, no cloud and no NPU: they synthesize live in your browser via
WebAssembly and run faster than real time on a $3 ESP32-S3.
pip install sanotts
The first open-source Nepali VITS text-to-speech with its own Devanagari
G2P and no eSpeak dependency: five voices, 50× real time on CPU, and
an open training recipe rather than a dropped checkpoint.
pip install kala-tts
A speech synthesis demo for Limbu / Yakthung — a language with no commercial voice and no prospect of one. Type Limbu, hear it spoken.
The same for Gurung / Tamu Kyi: a working voice for a language that has never had one, built on the shared frontend contract below.
The first comprehensive benchmark for Nepali text-to-speech: 365
phonologically designed sentences, twelve systems compared, and
5,760+ human MOS ratings from 164+ native speakers — plus ASR
round-trip metrics and a NepaliMOS predictor for when you cannot convene
164 listeners. pip install neptts-eval
Decades of Nepali writing are trapped in legacy ASCII fonts. This converts
it to real Unicode across the scripts of Nepal, Sikkim and neighbours
— Devanagari (Preeti, Kantipur, NayaNepal, Gorkhapatra), Limbu,
Kirat Rai, Sunuwar, Lepcha, Ol Chiki, Tirhuta, Tibetan, Newa, Gurung Khema
— over 146 routing keys with hash-pinned maps that fail closed rather
than invent a mapping. pip install nepal-ttf2utf
Devanagari to phone tokens for Nepali TTS: eSpeak-free, audit-traced, lexicon-first. It documents the dialect it encodes — Kathmandu Newari-influenced Nepali, per Khatiwada (2009) — and says plainly which listeners it is not right for, instead of pretending a single Nepali pronunciation exists.
The other direction: from sounds back to Devanagari spellings. Lexicon tooling for repairing dictionaries and building the pronunciation resources the rest of the pipeline runs on.
A deterministic Devanagari-to-IPA frontend for Dhimal that preserves what a generic Nepali frontend erases — visarga-marked vowel length, the conscripted glottal-stop spelling, Dhimal schwa behaviour — with alignment, evaluation, and a CPU ONNX demo. Held-out PER: 9.66% rule-based, 5.80% neural, reported separately rather than merged into one flattering number.
The registry and shared engineering contract behind the language work. The programme target is 44 languages and named varieties; six frontends are implemented and demonstrated today — Nepali, Limbu, Magar, Bantawa, Gurung, Dhimal. One of the 44 slots is deliberately left unresolved rather than invented to make the count look complete.
Page-to-book OCR for Limbu in the Sirijonga script: scanned pages and PDFs become editable Unicode with page structure, reading order, and per-line crops that trace every prediction back to its source. 0.474% codepoint CER on 902 held-out lines. Ships with a 50-language registry that refuses to run routes whose data rights are unreviewed.
वर्णमाला, “the alphabet”. A 1.11M-parameter distilled student that matches the 17.32M-parameter state of the art on handwritten Devanagari — 15.6× fewer parameters at statistical parity. We do not claim a win: the benchmark is saturated at a label-noise floor, and the repository ships the prediction dumps so you can check that yourself without a GPU.
Recognition and translation research for Nepali Sign Language. The project page is public; the corpus, pose data and checkpoints are not, pending agreements with the Deaf community organisations the work depends on. Published figures come from hash-pinned held-out evaluations.
| GitHub org | github.com/Ampixa |
| HuggingFace org | huggingface.co/ampixa |
| NepaliConformer | github.com/Ampixa/nepaliconformer |
| NepTel benchmark | huggingface.co/datasets/ampixa/neptel |
| sanoTTS | github.com/Ampixa/sanoTTS |
| NepTTS-Bench | github.com/Ampixa/neptts-bench |
| LimbuOCR | github.com/Ampixa/limbu-ocr |
| Barnamala | github.com/Ampixa/barnamala |
| Language toolkits | github.com/Ampixa/language-toolkits |
| nepal-ttf2utf | github.com/Ampixa/nepal-ttf2utf |
| Nepali MFA | github.com/Ampixa/nepali-mfa |
| Reverse G2P | github.com/Ampixa/nepali-reverse-g2p |
| Voice recorder | tts.ampixa.com/speak |
| Rating platform | tts.ampixa.com/rating |
| hello@ampixa.com |
© 2026 ampixa labs.
Devanagari wordmark uses
"8-bit devanagari"
by colonelhathii (CC BY-NC 3.0).