Techbanq

AI · speech recognition

Arabic speech recognition is a dialect problem before it is a model problem

The Arabic your customers speak on the phone is not the Arabic most recognition systems were trained on. Gulf, Egyptian and Levantine dialects, English product names dropped mid-sentence, and your own domain vocabulary decide real transcription quality — so we engineer for those, and measure on your calls before anything goes live.

Enterprise delivery credentials

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

The real problem

The gap between broadcast Arabic and the Arabic on your phone lines

We do not have an office in Saudi Arabia

We serve organisations there from our Riyadh office and our engineering hubs in Lahore and Karachi, travelling for the parts of an engagement that need to happen in person.

Most Arabic recognition systems perform best on Modern Standard Arabic — the language of news broadcasts and formal documents — because that is what dominates training data. But nobody calls a contact centre in MSA. A Riyadh caller speaks Najdi or Hijazi, an Egyptian expatriate speaks Egyptian, a Levantine caller something else again, and the spelling-free, vowel-light nature of dialectal speech means an engine tuned for the standard register degrades in ways its brochure never mentions. The distance between the demo and your phone lines is the project.

On top of dialect sits code-switching. A real Gulf sentence moves between Arabic and English mid-stream — an English product name, a policy number, "invoice", "upgrade" — and an engine locked to one language at a time garbles exactly the words your workflow needs most: the entities. Recognising mixed-language speech, and biasing the engine toward your own product names, branch names and domain terms, is where engineering effort actually changes the outcome.

What breaks

The four places Arabic transcription quietly fails

Each of these is invisible in a vendor demo and decisive in production. The evaluation is designed to expose all four before you commit.

  1. 01

    Dialect coverage

    An engine strong on MSA and Egyptian may be weak on Gulf speech, and vice versa. Coverage must be tested per dialect your callers actually use, because an average hides the customer segment it fails.

  2. 02

    Code-switching mid-sentence

    Arabic sentences carrying English product names, numbers and loanwords defeat engines that decide the language once per utterance. The words lost are usually the entities — names, amounts, references — that the downstream workflow runs on.

  3. 03

    Domain vocabulary

    Your product names, tariff names, place names and abbreviations are absent from any general model. Without custom vocabulary and phrase biasing, the engine substitutes its best acoustic guess — plausible, wrong and hard to catch.

  4. 04

    Audio reality

    Compressed telephony audio, cross-talk, hold music and speakerphones cost recognition quality before the model even starts. The evaluation must run on your real channel audio, not clean studio samples, or it measures a system you do not operate.

The measurement

How we measure Arabic recognition quality without quoting a brochure number

We will not tell you an accuracy figure before the evaluation, because any figure quoted before it would be someone else's audio. This is the methodology that produces your figure.

STAGE 01

Sample your real audio

A test set is drawn from your actual recordings across the dialects, call types and audio conditions you serve — consented and handled under your data governance — so the evaluation covers the traffic, not a favourable slice of it. A representative, governed test set.

STAGE 02

Build the human reference

Native speakers transcribe the sample under written conventions — how dialect words, numbers and code-switched English are rendered — because a word-error-rate is only as meaningful as the reference it is measured against. A reference transcript with agreed conventions.

STAGE 03

Score per segment, not on average

Candidate engines and adaptations are scored on word error rate broken out by dialect, call type and audio condition, with entity errors — names, amounts, references — tracked separately because they cost more than ordinary words. A per-segment quality picture tied to your traffic.

STAGE 04

Decide against the workflow, then adapt

Whether the measured quality is sufficient depends on the use: searchable archives tolerate more error than automated actions. Where a segment falls short, we adapt — vocabulary biasing, audio handling, engine choice — and re-run the same evaluation until the workflow's bar is met or the honest answer is that it is not. A go or no-go per workflow, backed by your own numbers.

Our role

From recognised speech to a working contact-centre workflow

  1. 01

    The engineering, end to end

    We select and adapt the engine, build the custom vocabulary, handle the telephony audio path and run the evaluation — the work between "speech recognition exists" and "it works on your calls".

  2. 02

    The workflow integration

    Transcripts flow into the systems that use them — call summaries into the CRM, searchable archives for quality teams, triggers for compliance review — because a transcript nobody consumes is a cost, not a capability.

  3. 03

    The confidence discipline

    Low-confidence passages are flagged rather than silently trusted, and workflows that act on a transcript — not merely store it — get a human checkpoint where the evaluation says the error rate warrants one.

What error rate each workflow can tolerate is an operational judgement that belongs to you; we produce the per-dialect, per-call-type numbers from your own audio that make the judgement an informed one.

Common questions

What buyers ask

What accuracy can we expect from Arabic speech recognition?+

We deliberately do not quote a figure, and you should distrust anyone who does before hearing your audio. Recognition quality varies enormously with dialect mix, audio conditions, code-switching and domain vocabulary — the same engine can be excellent on clean MSA and poor on compressed Gulf telephony audio. The professional answer is a methodology: a test set sampled from your real recordings, a human reference transcript, and word error rate measured per dialect and call type. That produces your number, on your audio, before you commit to building on it.

Can one system handle Gulf, Egyptian and Levantine callers together?+

It can be engineered to, but no engine should be assumed to cover all three equally out of the box. Dialect coverage differs by engine and is rarely documented honestly, which is why the evaluation scores each dialect separately on your own calls. Where one segment lags, the options are adaptation, routing — for instance different handling per detected dialect — or accepting a higher review rate for that segment. The mistake to avoid is deploying on an overall average that quietly serves one caller population well and another badly.

How do you deal with Arabic-English code-switching in transcription?+

Three ways, applied together. First, engine selection favours systems that recognise mixed-language speech within one utterance rather than deciding the language once per call. Second, your English-origin entities — product names, plan names, technical terms — go into custom vocabulary and phrase biasing, because these are the words a purely acoustic guess mangles most. Third, the reference transcripts used in evaluation write down conventions for rendering switched words, so the measurement actually counts code-switching errors instead of excusing them. The entities are usually what the downstream workflow needs, so this is where we spend the effort.

Next step

Evaluate it on your data

An evaluation on a sample of your own recordings that shows, per dialect and call type, whether recognition is good enough to build your workflow on.