A supervised multilingual human operations center for Voice AI

Real human conversations for better Voice AI.

Synergy Voice AI Lab produces consented multilingual conversational audio, tests voice agents with real speakers, and delivers structured human quality operations across French, Arabic, Italian, Spanish and regional dialects.

Request a paid pilot Download sample QA report Discuss a data partnership

Documented consent on every recording. No real customer PII.

Two operators recording a human conversation on two separate audio channels
Dual-channel Consent documented
0+
Years of Contact Center Experience
0
Active operational agents
0+
Inbound & Outbound Operations
FR & AR
French & Arabic core expertise
Tunis
Supervised operations hub
Europe
aligned
Time zone (UTC+1) — same working day

60 refers to the active operational agents on duty today — not to the number of available positions or seats. Capacity scales with contracted production, without inflating the number of permanent operators.

Built by contact center operators, not just AI engineers. Our quality standards come from ten years of managing live customer conversations — not from a lab.

Services

Three production-grade offers

Every offer is supervised, consented, measured and delivered with documented evidence. Start small with a calibrated pilot, then scale.

01

Consented Conversational Audio Data

Newly produced human-to-human conversations — natural, semi-guided or role-based — recorded on two separate channels with documented consent and no real customer PII.

  • Natural human-to-human conversations
  • Semi-guided conversations
  • Role-based customer-agent conversations
  • Scripted and unscripted speech
  • Dual-channel recording
  • Separate audio track per speaker
  • Participant recruitment
  • Speaker consent management
  • Language and accent qualification
  • Demographic metadata where legally permitted
  • Project supervision
  • Audio quality review
  • Secure delivery
Newly produced data Documented consent No real customer PII Controlled production Scalable speaker recruitment
02

Human Voice Agent Testing

Real people, real calls, real behavior. We test your voicebot the way your users behave — not the way a script expects.

  • Real calls to voicebots
  • Scripted and natural test scenarios
  • Accent robustness testing
  • Interruption and barge-in testing
  • Hesitation and silence testing
  • Objection handling
  • Background noise testing
  • Intent recognition checks
  • Response quality evaluation
  • Transfer-to-human testing
  • Task completion scoring
  • Linguistic and cultural QA
  • Reproducible defect reporting
  • Regression campaigns

We complement automated simulations with real human perception and unpredictable caller behavior. Automation finds what machines can script; we find what humans actually do.

03

Audio Annotation & Human Evaluation

Structured human labeling and evaluation of audio, transcripts and LLM conversations — with measurable inter-annotator quality.

  • Audio transcription
  • ASR post-editing
  • Timestamps
  • Speaker identification
  • Speaker diarization
  • Agent/customer role labeling
  • Intent classification
  • Entity annotation
  • Sentiment analysis
  • Emotion labeling
  • Interruption and overlap labeling
  • Silence and acoustic-event detection
  • Call outcome classification
  • Conversation quality evaluation
  • LLM response comparison
  • Instruction-following evaluation
  • Linguistic quality assessment
Call Center Audio

Call center audio, sourced the right way

Two distinct offers. One rule: no audio enters an AI pipeline without documented rights.

New, consented dual-channel production

Synergy recruits and records participants specifically for the AI project. Every speaker is informed of the intended use, and every conversation is produced under controlled technical and quality requirements.

  • Dual channel
  • Native / local speakers
  • Natural or semi-guided conversations
  • Documented consent
  • Controlled topics
  • No actual customer information
  • Secure transfer
  • Quality report

Pre-existing audio eligibility review

Synergy can assess whether historical audio is legally and technically eligible for AI use, without promising that existing customer recordings can be licensed.

  • Contract-rights review
  • Provenance assessment
  • Customer authorization requirements
  • PII risk assessment
  • Channel-format audit
  • Language and duration inventory
  • Licensing eligibility report

Existing customer calls are never reused or licensed without verified contractual rights, lawful basis and documented authorization.

Testing offers

Voice Agent QA Sprint

A focused, repeatable testing campaign for one voice agent — structured to fit a release cycle, not to replace your QA team.

A tester calling a voice agent with a headset
Flagship campaign

Voice Agent QA Sprint

One language, one voice agent, full behavioral coverage — with a reproducible defect log and an executive report your team can act on.

1
Language
1
Voice agent
50–100
Scenarios
200–500
Human calls
  • Accent and edge-case coverage
  • Reproducible defect log
  • Executive QA report
  • Final review meeting
Request a QA Sprint
Quick start

20-Call Discovery Benchmark

A fast, low-commitment snapshot of how your voice agent performs with real speakers.

10
Scenarios
20
Human calls
1
Language
30′
Review call
  • Short findings report
  • Top priorities identified in 2 weeks
Request benchmark

Paid pilots only. Scope and pricing depend on language, volume, speaker diversity and technical requirements.

Languages & locales

Native speakers, regional accents, real code-switching

Every speaker is sourced and qualified for language, accent, dialect and voice quality before any recording starts.

Several speakers representing different languages and accents

French — France

Metropolitan French with regional accents, for ASR/TTS and conversation benchmarks.

Highlight

North African French

Maghrebi French accents — the real challenge for voice models deployed in France and North Africa.

Modern Standard Arabic

Formal Arabic for news, education and professional domains.

Signature

Tunisian Arabic

Darija Tunisienne — authentic, unscripted, produced by native Tunisian speakers in Tunis.

Signature

Maghrebi Arabic

Regional Arabic varieties spoken across the Maghreb, sourced per-country and dialect.

Rare specialty

French–Arabic code-switching

Natural mixing of French, Arabic and dialect inside a single conversation — hard to synthesize, essential for realistic training.

Italian

Standard Italian plus regional accents for IT and EU markets.

Spanish

Castilian and Latin-American Spanish, subject to native-speaker sourcing.

English is used for project coordination — not as an initial collection specialty.

Additional languages available subject to native-speaker sourcing and qualification.

Conversation types

Scenario libraries built around your use case

We design dialogue flows, scripts and edge cases for your domain — from formal support to everyday free-flowing speech.

Customer support

Product, service and account questions handled by a support agent.

Appointment booking

Calendar-based scheduling dialogues with confirmation steps.

Technical assistance

Troubleshooting steps, diagnostics and resolution paths.

Sales qualification

Needs discovery, budget and decision-process questions.

Complaint handling

Escalating emotions, explanations and resolution commitments.

Billing enquiries

Invoices, charges, payment plans and refund processes.

Subscription management

Plans, renewals, upgrades, cancellations and retention.

Travel and hospitality

Bookings, changes, itineraries and on-trip assistance.

E-commerce and delivery

Orders, tracking, returns and delivery follow-up.

Telecom support

Plans, devices, connectivity and service incidents.

Insurance administration

Policies, claims intake and documentation follow-up.

Healthcare scheduling

Appointments, reminders and practice follow-up.

Automotive services

Maintenance booking, roadside assistance and service follow-up.

Everyday natural conversations

Free-flowing speech on daily topics for unstructured corpora.

Multilingual code-switching conversations

Natural switching between French, Arabic and dialect in one call.

Methodology

Nine controlled steps, from scoping to secure delivery

The same discipline we apply to live contact-center operations — applied to data production and testing.

Production pipeline: audio, transcription, annotation, quality assurance and secure delivery
1

Project scoping

Objectives, use cases, volumes and constraints defined with your team.

2

Legal and usage-right definition

Usage rights, consent wording and data boundaries documented before production starts.

3

Scenario design

Dialogue flows, scripts and edge cases written for your AI use case.

4

Speaker sourcing and qualification

Native speakers screened for language, accent, dialect and voice quality.

5

Consent and onboarding

Informed consent, briefing and rehearsal before any recording is made.

6

Pilot and calibration

A small calibrated batch aligns quality criteria with your team before scale-up.

7

Controlled recording

Supervised sessions, dual-channel capture and acoustic checks on site.

8

Multi-level QA

Automated checks plus human listening on every deliverable.

9

Secure delivery and reporting

Encrypted transfer, file manifests and acceptance reports.

Quality gates applied at every stage

Automated audio checks Human listening review Speaker and language validation Duplicate detection Channel verification Silence and noise thresholds Metadata validation Re-recording workflow Final acceptance report
Sample report

What a QA campaign report looks like

An excerpt of the executive dashboard delivered after a Voice Agent QA Sprint. All figures below are illustrative.

Sample data — for demonstration only
A QA analyst reviewing waveforms, transcripts and defects

Voice Agent QA Sprint — Executive Summary

Voice agent « Atlas » · French (North African accents) · Week 22
Delivery status: report delivered
0
Calls tested
0
Scenarios executed
0%
Task completion
Above pilot target (80%)
0
Confirmed defects
3 critical  9 major  15 minor

Success rate by language

Task completion per language

Defects by severity

27 confirmed defects
  • Critical — 3 (11%)
  • Major — 9 (33%)
  • Minor — 15 (56%)

Failure categories

Share of failed scenarios by cause

Performance by accent

Task completion by speaker accent

Conversation duration

Distribution of human-agent call length

Fallback & transfer

Where the agent gave up — and what happened then
  • Fallback rate — 12%
  • Human transfer success — 87%
  • Direct resolution — 88%

Key finding: task completion drops by 18% with a marked North African accent — consistent across 3 critical defects (intent misrouting on code-switched utterances).

Download sample report (PDF)
About the Lab

A division of Synergy Growth — built on real operations

Synergy Voice AI Lab is built on ten years of real contact-center operations. Our teams have handled more than twenty inbound and outbound programs across customer care, sales, qualification, technical support and back-office operations. We are converting this operational experience into human data and quality services for the next generation of conversational AI.

The same floor managers, QA auditors, trainers and linguists who run our live programs supervise every data collection and testing campaign.

10 years of experience 60 active agents 20+ programs Live supervision Dual listening Training programs Dedicated QA Secure infrastructure Located in Tunis European proximity Scalable capacity
A supervised team in a modern operations center in Tunis
Security & consent

Privacy and consent, engineered into every workflow

We protect people and data through documented workflows — not through slogans.

NDA support

Mutual NDA signed before any project details are shared.

Participant consent

Informed, documented, revocable consent from every speaker.

Project-specific usage rights

Audio licensed for the declared AI purpose — nothing broader.

Restricted topic design

Sensitive or personal topics excluded from scenario design.

No real customer PII

New collections contain no live customer data — by design.

Role-based access

Each operator accesses only the projects assigned to them.

Secure file transfer

Encrypted transfers and per-project delivery channels.

Controlled workstation access

Locked-down machines, no personal devices on project data.

Retention and deletion policy

Defined retention per project and documented deletion at end.

Incident reporting

Named contact, defined escalation path, written follow-up.

International-transfer review

Cross-border movement of data reviewed per project.

GDPR-aligned workflows

Consent records, data minimization and EU-aligned practices.

Security and privacy controls can be aligned with project requirements. We do not claim certifications we do not hold; we document the controls we actually apply.

Voice-cloning, speaker-identification and biometric uses must be explicitly declared and contractually authorized. We will not produce or deliver audio for undeclared biometric or identity purposes.

FAQ

Questions we get from data and product teams

Do you reuse or license existing customer calls?+
No. New collections are produced with recruited, consented participants and contain no real customer PII. For pre-existing audio, we only assess legal and technical eligibility — and we never reuse or license recordings without verified contractual rights, lawful basis and documented authorization.
Which languages and accents do you specialize in?+
French (France), North African French, Modern Standard Arabic, Tunisian Arabic, Maghrebi Arabic, French–Arabic code-switching, Italian and Spanish — each with native-speaker sourcing and qualification. Additional languages are available subject to sourcing.
How is participant consent documented?+
Every participant is informed of the intended use and records a documented, project-specific consent before any recording. Usage rights are restricted to the declared AI purpose, and consent records are retained for the project lifetime.
How do you test voice agents with real people?+
Real human callers run scripted and natural scenarios covering accent robustness, barge-in, silence, noise, objections and transfer-to-human. We deliver a reproducible defect log, task-completion scoring and an executive QA report. A full Sprint covers 50–100 scenarios and 200–500 calls.
Can you produce dual-channel audio?+
Yes — dual-channel recording with a separate track per speaker is the default for our conversational collections, with channel verification and silence/noise thresholds applied before delivery.
How do we start — and what does it cost?+
We run paid pilots only. Scope and pricing depend on language, volume, speaker diversity and technical requirements. Send the qualification form and we will come back within two business days with a scoped pilot proposal.
Contact

Request a pilot or data partnership

Tell us about your project. A senior member of the Lab — not a sales bot — will reply within two business days.

Please enter your first name.
Please enter your last name.
Please enter a valid work email address.
Please enter your company name.
Please enter your job title.
Please enter your country.
Please enter a valid URL (starting with http).
Please select a project type.
Please describe your project briefly (minimum 20 characters).
Please choose a file smaller than 10 MB.

Your data is used only to respond to your request. No marketing lists, no sharing.

Thank you — request registered.

An advisor from the Lab team will contact you within two business days. A copy of your request has been prepared for your email client.

Let’s build voice AI that understands real people.

Start with a paid pilot, a benchmark, or a data partnership conversation.

Request a paid pilot Discuss a data partnership Download sample report