ampixa
ampixaएम्पिक्सा ल्याब्स

we bring human interaction to computers. सुन्छ · बोल्छ · पढ्छhears · speaks · reads

another chatbot the human layer.

speech · writing · reading · listening — on phones, browsers and tiny chips.
Every number on this site links to where it came from.

Human interaction is messy. It carries accent, breath, handwriting, noise, code-switching, hesitation, dialect, old fonts, bad microphones, and the parts of language that clean text throws away. That is the layer we care about.

hear it — the words are in the sound
// demos

Hear it, call it, read it.

scripts & text

How many pieces does it take to say "Kathmandu is the capital of Nepal"?

download onlyFive tokens where Llama-3 needs seventeen.
speech recognition

nepaliconformer

download onlyNepali speech recognition, open weights.
speech synthesis

294k ≠ 294k

liveA tiny voice that runs in your browser.
document ocr

The newspaper already knew how to read itself

liveUpload a page, get the text.
evaluation · sign · music

A score that grades clips but can't rank voices

liveRate Nepali text-to-speech systems.
// everything · 61 projects

The whole wall.

liveshippedresearcha negative we kept→ has a story page

speech recognition7

speech synthesis17

scripts & text8

document ocr5

  • devanagari-ft-v3(live)→
  • PaddleOCR-VL-1.6 lines(research)
  • deva-ocr-75m(shipped)
  • Paddle → PyTorch / MPS(shipped)
  • Kept negatives(kept negative)

kura8

  • Multi-user deployment(live)
  • LLM(live)
  • Receptionist / tool-use eval(research)
  • Echo cancellation + ASR adaptation(research)
  • 33M Kura ASR vs 202M teacher(research)
  • KURA memory ASR(kept negative)
  • Live-session audit(shipped)
  • Voice-agent benchmark(research)

evaluation · sign · music9

systems7

// where it lives · 50 public links, all checked

Every door in.

try it

sites & pages

models

datasets

code