Banking and insurance

Train models on your calls, claims and contracts

For banks, insurers and their contact centres: transcribe, extract and fine-tune per language and product.

Two panels. Left: a waveform over three call transcripts with word-error-rate chips (call_0381, Arabic, WER 4.1%; call_0412, Malay–English switch, WER 6.0%; call_0450, French, WER 5.3%), Agent and Customer speaker tags and masked account digits ending 4417. Right: a synthetic claim form with five extracted fields (Policy no. PX-20931, Date of loss 14.10.2026, Vehicle reg. masked, Amount claimed USD 18,450.00, Signature) and a handwritten note tagged low confidence, to review.

What it's worth

Fit speech models to how customers talk

General speech models miss dialects and product names; hand transcription to fix that is costly.

issues resolved hourly by chat-support agents with an AI assistant; +34% for new agents: +14%
word error rate, TII's own Gulf Arabic test set: general against Emirati-dialect model: 31% → 23%
to transcribe one audio hour by hand: USD 60–180

Sources: NBER working paper 31161, 2023, chat support in English, 5,179 agents at one company; TII's Falcon-ASR release figures, 07.10.2026, as reported by FourWeekMBA, on TII's own Gulf test set (Cohere Transcribe Arabic 31.05%, Falcon-ASR 22.73%); SpeakWrite transcription prices, 2026.

Use cases

Understand every call, claim and contract

Use case	Raw data	Rare cases to find	Adapt per	Release test
Contact-centre speech	Call audio	Noise, mixed languages, rare names	Dialect, product	Word error rate against the last release
Claims	Claim forms, photos, adjuster notes	Handwriting, unusual damage	Insurer, product	Extraction accuracy on a fixed test set
Contracts and KYC	Scans, PDFs, ID documents	Unusual clauses, poor scans	Language, jurisdiction	Extraction accuracy on a fixed test set

Contact-centre speech
Raw data
Call audio
Rare cases to find
Noise, mixed languages, rare names
Adapt per
Dialect, product
Release test
Word error rate against the last release
Claims
Raw data
Claim forms, photos, adjuster notes
Rare cases to find
Handwriting, unusual damage
Adapt per
Insurer, product
Release test
Extraction accuracy on a fixed test set
Contracts and KYC
Raw data
Scans, PDFs, ID documents
Rare cases to find
Unusual clauses, poor scans
Adapt per
Language, jurisdiction
Release test
Extraction accuracy on a fixed test set

How it works

Raw calls and documents in. Tested releases out.

Melt dataPrepare the data · 01–03

01Ingest
Load call audio, claim forms, contracts and ID scans; drop corrupt files, catalogue by language and source, and mask fields each dataset does not need.
02Mine
Find noisy calls, mixed languages, rare names, unusual clauses and handwriting, and drop near-identical routine files.
03Label
Transcribe and extract with large labelling models that run inside the platform; your reviewers correct a sample.

Forge intelligenceTrain and test the model · 04–06

04Train
Retrain your shared base model on the full, versioned dataset when calls in a new dialect land; every release traces to its exact data.
05Adapt
Fine-tune the base model for each dialect, product, language or jurisdiction, and tune it again as usage changes.
06Validate
Optimise each release candidate for your contact-centre servers, then test word error rate and extraction accuracy against the last release.

Languages

One adaptation per language or dialect

Each one, including calls that switch language mid-sentence, gets its own word-error-rate test set.

Examples: eight example adaptations, each with its own adaptation and its own test set. Arabic dialect, base model v13 to adapted v13-ar; French, v13 to v13-fr; Swahili, v13 to v13-sw; Malay, v13 to v13-ms; Bahasa Indonesia, v13 to v13-id; Thai, v13 to v13-th; Tagalog, v13 to v13-tl; Arabic–English switch, v13 to v13-ar-en.

Quality

Regressions stop at the gate

Every release candidate runs the last release's test sets; a drop beyond your margin blocks it until you sign off.

Your team sets the metric and margin for each test set: word error rate per language and dialect, so no customer group gets a worse model, and field accuracy on claims and KYC, each with its 95% confidence interval. Each release keeps a locked record of dataset and test-set versions, metrics and sign-off, which your team exports for model-risk review.

Banking and insurance

release v13 → release candidate v14

Banking and insurance test sets: release v13 → release candidate v14
Test set	Cases	Last release	Release candidate	Margin	Result
Contact-centre calls, Arabic, word error rate	600 h	9.2%	8.7%	+0.5 pts	Pass
Calls that switch language, word error rate	120 h	13.4%	15.2%	+0.5 pts	Fail +1.8 pts
Claim forms, field accuracy	3,000 forms	96.1%	96.8%	−0.5 pts	Pass
KYC scans, field accuracy	2,500 scans	97.4%	97.5%	−0.5 pts	Pass

Contact-centre calls, Arabic, word error rate
Cases
600 h
Last release
9.2%
Release candidate
8.7%
Margin
+0.5 pts
Result
Pass
Calls that switch language, word error rate
Cases
120 h
Last release
13.4%
Release candidate
15.2%
Margin
+0.5 pts
Result
Fail +1.8 pts
Claim forms, field accuracy
Cases
3,000 forms
Last release
96.1%
Release candidate
96.8%
Margin
−0.5 pts
Result
Pass
KYC scans, field accuracy
Cases
2,500 scans
Last release
97.4%
Release candidate
97.5%
Margin
−0.5 pts
Result
Pass

Release candidate blocked: 1 test set beyond its margin

Interface preview · sample data

Your data

Your calls, documents and weights stay yours

Before you upload, we confirm with your compliance team which data may be processed at the site your contract names.

Ownership: Your data, labels and trained weights stay yours. Bytemelt has no access to them: your contract says so, and the platform is built so only accounts your team controls can open them.
Masking: Your team sets versioned rules per dataset: mask fields such as account numbers in calls, keep ID and policy fields for extraction.
Location: Your data is stored and processed only at the partner data centre your contract names, in the Middle East and outside the EU. The contract names the site and country before any data moves, and data leaves the site only when your team moves it. Personal data from the EU moves only under transfer safeguards such as Standard Contractual Clauses, and we give your data-protection officer the site details a transfer impact assessment needs.
Vendor review: Before you sign, we answer your vendor-risk questionnaire, and your contract names the data-centre operator and the data-protection terms.

How Bytemelt handles your data

Questions banks and insurers ask

All questions about Bytemelt

Which languages and dialects can we train for?

The ones in your own calls and documents. Languages, above, shows how each gets its own adaptation and test set.

Why not train in our cloud provider's region?

You can. If your regulator requires in-country processing, an in-country region may be your only option; we say so on the first call. Otherwise, a cloud region bills GPUs by the hour or reservation; Bytemelt includes compute power, storage and the pipeline in one subscription.

Where does the trained model run?

On your servers. Your team exports the trained weights and deploys the model in your contact centre or document systems; the platform trains and tests it first.

See the other three industries

Size compute for calls and documents

In one hour, the founder maps your speech and document models to the six stages and sizes the compute.

Book demo call Email us

contact@bytemelt.ai