Project Approach Research targets Get in touch
SJM Labs · Speech research

Speech, made smaller and more open.

Jarir is a research project exploring a compact text-to-speech model for English, Modern Standard Arabic and Najdi Arabic.

EnglishModern Standard ArabicNajdi ArabicCompact speech models
The approach

A new student model, trained from its first step.

Jarir is being explored as a compact model with randomly initialized student weights. Existing speech models may help produce teaching examples; their weights are not the starting point for the student.

01

Compare teachers

Assess candidate speech models for naturalness, pronunciation and coverage of the target varieties.

02

Build speech data

Generate and filter text-and-audio examples with clear language, speaker and style labels.

03

Train from scratch

Train a compact student architecture from random initialization and measure what it learns.

04

Evaluate each step

Check intelligibility, dialect fit, voice consistency and expressive control before expanding.

Research targets

A focused starting point for multilingual speech.

These are project goals under investigation, not features of a released model. The pilot will determine which targets fit the available data and compute.

01 · Languages

Three varieties

English, Modern Standard Arabic and Najdi Arabic, evaluated separately as well as together.

02 · Model size

20–40 million parameters

A compact student is the target range; the final size depends on pilot quality and inference needs.

03 · Expression

Voices and emotion

Explore 6–10 male and female voices, with emotion tags as a controlled research goal.

Built at SJM Labs.

Jarir is an early research effort in compact, multilingual speech synthesis.

Contact the team →
Get in touch

Follow the research.

For research, partnership or project enquiries, contact SJM Labs.