Arabic Dialects · Speech to Text

Tunisian · تونسي

Tunisian Arabic
Speech to Text

An STT model built for Tunisian. It handles the dialect's unique speed, its vowel deletion, and its natural mixing with French and Italian, the features that make Tunisian one of the hardest dialects for models that were not trained on it.

dialect: "tunisian"12M+ native speakers · real-time · five-language mix

12M+

Native Tunisian Arabic speakers

<29.9%

WER on Tunisian colloquial speech

<300ms

First-word streaming latency

Five-layer

Arabic + French + Italian + Amazigh + Turkish

Coming Soon

Demo launching soon

The interactive demo for this dialect is coming soon. Sign up to be notified when it launches.

Join the waitlist →

Dialect overview

What is Tunisian Arabic?

Tunisian Arabic is a Maghrebi variety with unique phonological features: high speech rate with intensive vowel reduction and syllable compression, making it among the most acoustically dense Arabic varieties. This is precisely why Tunisian is one of the hardest dialects for untrained generic models.

Key features that distinguish Tunisian from Modern Standard Arabic:

  • Systematic short-vowel deletion produces dense sound clusters.
  • French appears naturally in technical and professional vocabulary.
  • Italian leaves its mark on popular everyday words.
  • Amazigh and Turkish traces sit in successive historical layers.

Native speakers

12M+

Tunisia

Distinctive feature

Speed + vowel reduction

Fastest rhythm among Maghrebi dialects

Linguistic influences

Five-layer

Arabic + French + Italian + Amazigh + Turkish

Hamsa WER

<29.9%

on colloquial Tunisian audio · 2024

The problem

Why generic Arabic STT fails
on Tunisian Arabic

Tunisian packs more sound into less time than almost any other Arabic variety. Here is why generic models break.

Speed and vowel reduction mislead generic models

Tunisian systematically drops short vowels and compresses syllables, producing dense sound sequences that look nothing like the MSA or other-dialect data generic models were trained on.

A five-language blend

Arabic, French, Italian, Amazigh, and Turkish in the same sentence is not a rare event. It is normal everyday speech. Models not built for this mix fail to transcribe French and Italian words accurately.

Sound clusters created by vowel deletion

Vowel deletion creates unusual sound clusters that confuse models trained on a different syllable structure, whether Modern Standard Arabic or the Eastern Arabic dialects.

Regional variation across Tunis, Sfax, and the south

Speech in the capital differs from Sfax and southern Tunisia in several phonetic and lexical patterns. A generic model reads these differences as errors rather than natural variation.

The solution

How Hamsa handles
Tunisian Arabic natively

✓

Acoustic models tuned for Tunisian speed

Trained on hours of real Tunisian speech at its natural pace, with its vowel deletion patterns intact. Not slowed-down speech or formal read-aloud recordings.

✓

A language model that understands the Tunisian mix

The language model is trained on real Tunisian vocabulary, including the French, Italian, and Turkish words woven into everyday Tunisian speech.

✓

Smart code-switching layer

The model detects language switches automatically, separates Arabic, French, and Italian segments, and processes each one in its own language.

✓

Tunisian sub-dialect coverage

Models cover the speech of Tunis, Sfax, and southern Tunisia to keep accuracy high across the whole country.

Capabilities

Everything you need for
Tunisian Arabic STT

✓

Native Tunisian Transcription

Purpose-trained on colloquial Tunisian speech at its full speed and acoustic density. Not MSA with a dialect layer on top: native Tunisian acoustic and language models, built from the ground up.

✓

Vowel Reduction Handling

Tunisian systematically drops short vowels, producing dense sound clusters. Hamsa is trained on these patterns and decodes compressed Tunisian speech accurately.

✓

Code-Switch Detection

Tunisians move between Arabic, French, and Italian mid-sentence. Hamsa handles these transitions without losing context.

✓

Smart Punctuation

Automatic sentence boundaries, commas, and question marks inserted correctly for Tunisian syntax.

✓

Real-time Streaming

Under 300ms first-word latency. Transcript fragments arrive as the speaker talks, at full Tunisian speed.

✓

Custom Vocabulary

Add Tunisian brand names, tourism terminology, or hybrid French-Tunisian vocabulary to improve recognition accuracy.

Transcript examples

Real Tunisian Arabic.
Accurate transcription.

Telecom support

عسلامة، نحب نبدّل الفورفي متاعي، الكونيكسيون ما تمشيش مليح.

"Hello, I'd like to change my plan. The connection isn't working well."

Tunisian: "عسلامة" (hello), "متاعي" (my), French loans "forfait" and "connexion"

Hotel front desk

يعطيك الصحة، الشمبرة فيها برشا حس، تنجم تبدّلهالي؟

"Thank you. The room is very noisy, could you change it for me?"

Natural Tunisian: "يعطيك الصحة" (thanks), "برشا" (a lot), "شمبرة" (room, from French)

Media interview

توا الشباب الكل ولّى يخدم في الـ digital، يعني الحكاية تبدّلت برشا.

"Now all the young people work in digital, so things have really changed."

Code-switching: "digital" mid-sentence, "توا" (now), "ولّى" (became)

Use cases

Where Tunisian Arabic STT
delivers results

Tourism

Better Visitor Experiences in Tunisia

Tunisia is a major tourist destination welcoming millions of visitors every year. Accurate recognition of Tunisian speech in hotels, tourist sites, and airports improves the guest experience.

Tech & Telecom

Digital Customer Service

Tunisia's growing tech sector needs accurate transcription models for customer service calls and user support.

Media & Broadcasting

Tunisian Media Archiving

Turn Tunisian programmes, interviews, and radio content into searchable, archivable text.

EdTech

Educational Content Transcription

Convert lectures and lessons delivered in Tunisian Arabic into text for students, and build a searchable archive of educational content.

Accuracy benchmark

High-accuracy STT for Tunisian Arabic

Measured on a held-out benchmark of colloquial Tunisian speech recordings. Lower word error rate means more words transcribed correctly.

Word Error Rate (WER) · lower is better

Hamsa STT (Tunisian)us29.9%
Generic Arabic STT38.4%
MSA-trained model48.1%
GPT-4o-transcribe52.6%

Evaluated on Tunisian colloquial speech benchmarks · 2024 · Lower is better

Architecture

How it works

01

Audio in

Send audio via WebSocket stream or REST upload. Accepts PCM, MP3, M4A, WAV, and WebM, whatever your microphone, telephony platform, or recording system produces.

02

Dialect routing

Hamsa detects Tunisian Arabic features and routes the audio to the Tunisian-specific acoustic and language models. When code-switching is detected, Arabic and French segments are processed in parallel.

03

Transcript out

Transcripts stream as JSON events with word-level timestamps, speaker labels, confidence scores, and punctuation. Pipe directly into your database, UI, or downstream NLP pipeline.

Developer API

Two lines to Tunisian Arabic speech recognition

Drop in the Hamsa SDK, set dialect: "tunisian", and start streaming transcripts. No preprocessing pipeline, no normalization to MSA, no fallback surprises.

Read the API docs →
tunisian-stt.tsstreaming
import { Hamsa } from '@hamsa-ai/sdk';

const stream = hamsa.stt.stream({
  model:       'hamsa-stt-s4',
  dialect:     'tunisian',
  diarization: true,
});

stream.on('transcript', (t) => console.log(t.text));
// → "عسلامة، لاباس عليك؟" · 00:01.840

Related services

More Hamsa for Tunisian Arabic

Why Hamsa

Built for Arabic from the ground up

✓

Native, not adapted

Hamsa trains on real Arabic dialect data, not English models adapted for Arabic. On fast, compressed speech like Tunisian, that difference decides whether a transcript is usable.

✓

Enterprise-grade SLA

99.9% uptime, SOC 2 Type II compliance, and data residency in the MENA region. Your audio stays in-region.

✓

Fastest Arabic streaming

Under 300ms first-word latency in production. Real-time transcription that keeps pace with human speech rates.

✓

13+ dialects, one API

Switch dialect with a single parameter. Run A/B tests across Tunisian and the other Maghrebi dialects. No separate integrations, no separate contracts.

FAQ

Tunisian Arabic STT:
frequently asked

Why is Tunisian one of the hardest Arabic dialects for generic STT models?

The main reason is short-vowel deletion and syllable compression, which create an unusual acoustic density. A model trained on MSA, or even on Eastern dialects, struggles to break down these Tunisian sound clusters. On top of that come a high speech rate and natural mixing with French and Italian.

Does Hamsa handle Tunisian-French code-switching?

Yes. Switching between Tunisian Arabic and French, and sometimes Italian, is natural in everyday Tunisian speech. Hamsa detects these switches in real time and processes each segment in its own language, so the full transcript stays accurate.

What word error rate does Hamsa achieve on Tunisian Arabic?

Hamsa achieves under 29.9% WER on colloquial Tunisian speech benchmarks, compared to 38–65% for generic Arabic STT models. The gap is widest on fast speech and contact center recordings.

Does the model cover Tunisian sub-dialects outside the capital?

Yes. The model is trained on data from Tunis, Sfax, and southern Tunisia to cover the main regional patterns.

Which industries use Tunisian Arabic STT most?

The leading use cases are tourism (speech recognition in hotels and tourist sites), digital customer service for tech companies, media (archiving TV and radio programmes), and e-learning (lecture transcription).

Explore more

13+ Arabic dialects, supported

Partner ecosystem

Trusted Across the
Technology Ecosystem

Hamsa works with leading technology companies, integrators, and solution providers to bring enterprise Voice AI to organizations across the region.