100% Offline Speech Translation Engine • Zero Data Required

The easiest way to understand each other.

Real-time voice-to-voice conversation across the Horn of Africa. Powered by on-device AI running directly on low-cost smartphones—without needing internet, cellular data, or cloud servers.

Download Android APK Listen to Live Demo
አማርኛ (Amharic) Afaan Oromoo ትግርኛ (Tigrinya) Soomaali English

Built upon premier African NLP & Open-Source Speech Research

HornMT Corpus Meta NLLB-200 ONNX Runtime INT8 Mozilla Common Voice Flutter Native
Why Offline Matters

Cloud translation fails when you need it most.

In rural clinics, border posts, and bustling outdoor markets across the Horn of Africa, mobile data is expensive, spotty, or completely unavailable. LISAN is engineered from the ground up to run fully offline without needing cloud servers.

Everyday Cloud Translators

Designed for high-bandwidth cities with permanent fiber and 5G.

  • Total failure during network outages: Cannot translate even basic greetings without connectivity.
  • High latency (1,800ms+): Awkward pauses break natural, spontaneous conversation.
  • Ignores Ethiopian cultural nuance: Translates blessings and honorifics literally, losing respect and context.
  • Muffled by street noise: Fails on real-world smartphone microphones in open markets and busy transit.

LISAN On-Device Intelligence

Optimized specifically for real everyday life in Ethiopia and the Horn.

  • 100% Offline & Zero Data: Runs entirely on your phone’s CPU via INT8 quantized neural models (~640MB).
  • Sub-second instant response (<350ms): Instant Indexed Translation Memory + CTranslate2 speed.
  • Culturally authentic: Preserves healing blessings (“እግዜር ይማርህ” $\leftrightarrow$ “fayyuu kee haa ta'u”) and elders’ honorifics.
  • Built-in Speech Repair & Noise Gate: Dynamically eliminates mic distortion, repetitive stutters, and background chatter.
Interactive Experience

Hear the difference.

Click below to play actual real-world recorded speech, observe LISAN’s automated speech repair healer, and listen to the synthesized translated speech.

Amharic (አማርኛ)
Original Voice
Spoken Audio Clip 0:03
Spoken Words

"የት አካባቢ ያመዎታል? የደም ምርመራ ያስፈልግዎታል።"

Speech Healer: Filtered mic noise & normalized morphological verb suffix.
Wav2Vec2 / Hohe ASR Latency: 142ms
Afaan Oromoo
Instant Speech Output
Synthesized Native Voice 0:03
Translated Output

"Eessa kee si dhukkee? Qorannoo dhiigaa si barbaachisa."

Cultural Context: Medical clinical domain pack matched via indexed Translation Memory.
CTranslate2 INT8 + MMS-TTS Total Latency: 290ms
Native Flutter Client

Designed for true face-to-face dialogue.

Place the phone flat on a consultation table or marketplace stall. LISAN’s interface allows two speakers to communicate naturally, displaying real-time transcribed audio waveforms and synthesized native speech at the push of a button.

Dual-Direction Mode
Instant toggle between speaker mic channels
Live Waveform Feedback
Visual feedback for low-gain speaking
Instant Audio Replay
One-tap neural speech synthesis repeat
Saved Phrases & Memory
Fast access to frequent daily questions
LISAN Flutter App Interface
Tested in Real Life

Engineered for the places
where words matter most.

From hospital emergency triage to local municipal Kebele offices, LISAN comes pre-equipped with verified bilingual domain glossaries.

Clinics & Healthcare

Accurate medical triage between physicians and patients across language borders.

"እግዜር ይማርህ"
➔ "Fayyuu kee haa ta'u"
Preserves traditional healing prayers

Kebele & Public Admin

Civil registration, birth certificates, and ID issuance without translators on staff.

"የመኖሪያ መታወቂያ ለማደስ"
➔ "Waraqaa eenyummaa haaromsuuf"
Accurate municipal terminology

Merkato & Commerce

Haggling, agricultural wholesale trade, and goods transactions in bustling noise.

"ዋጋውን ቀንስልኝ እባክህ"
➔ "Gatiisaa naaf hir'isi mee"
Respects conversational bargaining

Banking & Telebirr

Remittance claims, microloans, and mobile money transactions made clear.

"ገንዘብ ማስተላለፍ እፈልጋለሁ"
➔ "Mallaqa erguun barbaada"
Financial terminology verified
Under The Hood

An entire AI lab,
quantized inside your pocket.

LISAN doesn’t compromise on accuracy. We engineered deep optimization pipelines to squeeze multi-gigabyte models into a compact on-device runtime.

Morphological Speech Normalizer

Context-Aware Speech Repair & Disfluency Healer

Smartphone microphones in the field capture stuttering, dropped syllable endings, and conversational contractions. Our pipeline reconstructs clean grammatical text before passing it to translation.

● Raw Mic: "ማን... ማንት ነው እንድ... ወንድሜ" (Disfluent / Glitchy)
✔ Healed: "ማን ነው ወንድሜ" (Clean Grammatical Amharic)
Normalizes colloquial contractions & speech artifacts `speech_repair.py`
Tier 1 & Tier 2

Hybrid Smart Translation

Tier 1: 57.3 MB indexed SQLite FTS5 database for sub-millisecond (≤1ms) lookup of verified idioms, blessings, and civic terms.

Tier 2: CTranslate2 INT8 NLLB-200 neural machine translation for open-ended, complex sentences.

Lookup: ≤1ms NMT: ~210–350ms
640MB Total Footprint

ONNX Mobile Runtime

Exported via INT8 dynamic quantization. Runs smoothly on quad-core Android chipsets with 2GB RAM without heating the battery.

Zero server dependency Android / iOS
Meta MMS-TTS & Piper

Natural Local Voice Synthesis (TTS)

Clear, natural speech synthesis is essential for effortless spoken dialogue. LISAN generates expressive vocal inflections calibrated to authentic Amharic, Afaan Oromo, Tigrinya, and Somali prosody using 16kHz VITS neural vocoders.

Sample: `test_oromo_tts.wav`
16kHz Neural Vocoder Sub-100ms Synthesis
Standardized Benchmarks

Validated against HornMT reference corpora.

Evaluated rigorously on standardized parallel test sets across BLEU, chrF++, and conversational accuracy metrics.

41.2 chrF++
Amharic ➔ English (HornMT)

Character n-gram F-score evaluated on gold multi-parallel test sets.

113,726 Pairs
20 Parallel Language Pairs

Stratified bitext across Amharic, Oromo, Tigrinya, Somali, and English.

57.3 MB
Indexed SQLite Memory

≤1ms instant lookup for everyday idioms, clinic terms & greetings.

Why chrF++ over word BLEU? Agglutinative languages (Afaan Oromoo) and root-pattern Semitic languages (Amharic & Tigrinya) attach rich grammatical affixes directly to word stems. Standard word-level BLEU artificially penalizes valid inflections; character n-gram F-score (chrF++) provides a far more accurate measure of true linguistic comprehension.
Engine / Architecture Offline Capable Memory Footprint Speech Normalizer Inference Latency
LISAN (Our System) 100% Yes ~640 MB (INT8 ONNX) Rule-Based Normalizer ≤1ms (TM) / ~315ms (NMT)
Commercial Cloud API (Google/Azure) No (Requires 4G/WiFi) 0 MB (Cloud-only) Standard Cloud ASR ~1,800+ ms (Network RTT)
Vanilla NLLB-200 (PyTorch FP32) Heavy (Needs 4GB+ RAM / GPU) 2.4 GB None ~1,200 ms (CPU)
Get Early Access

Ready to speak without
boundaries?

Download the standalone Android APK or run the open-source backend locally on your workstation. No credit card, no cloud accounts, no tracking.

Download Android APK (v1.0-beta) View GitHub Repository
Android 8.0+ Compatible Apache 2.0 Open Source Zero Analytics Telemetry