Voice AI Product Tutorial · v1.03

Master Omind Voice AI in 30 minutes.

A complete training programme — 20 chapters, 5 modules, 2 knowledge checks. Build, configure, operate, and measure AI voice agents at any scale.

~30 minutes · audio + captions · resume any time

A product from Omind

In this guide

Five modules · twenty chapters

1

Foundations

Studio tour, call architecture, workspaces & roles.

2

Build

Agent setup, prompting fundamentals & deep dive, dynamic variables.

3

Configure

Voice, STT, LLM, knowledge base, tools, transfers, widget.

4

Operate

Phone numbers, Quick Dial, Predictive Dial, live monitoring, logs.

5

Measure

Post-call analysis, webhooks, reports, API.

★

2 knowledge checks

Mid-guide and final assessment to lock in what you learn.

🌙

Total Calls

5,284

+842 this week

Success Rate

87%

Based on latest data

Avg. Duration

2.3 min

Per call average

Transfers

412

7.8% transfer rate

Welcome back, Maya

This is the dashboard you'll see every time you sign in — total calls, success rate, average duration, and transfer rate at a glance. The four sidebar sections — Main, Outreach, Analytics, Settings — group everything in the platform so you always know where to look. The sidebar also collapses to icon-only mode to give you more screen space. Cost details aren't on the dashboard; they live in Reports.

Three layout patterns to recognise

Once you spot these, every page in Omind Voice AI makes sense.

Split list / detail

Agents, Phone Numbers — pick from a left list, edit on the right.

Full-width dashboard

Reports, Quick Dial — KPIs on top, charts and tables below.

Focused detail

A single call log, a campaign run, an API key — one focal record.

Spot the pattern · find anything in seconds.

The Omind Voice AI real-time pipeline

Seven stages · end-to-end under 500ms.

📞

Telephony

SIP trunk · Vonage / Twilio / custom

→

🎙️

VAD

Speech vs silence

→

🔇

Noise cancel

Filter ambient

→

📝

STT

Deepgram Nova-3

→

🔄

Turn detect

Whose turn?

→

🧠

LLM

GPT-5 Mini

→

🔊

TTS

Deepgram Luna

End-to-end · <500ms

Or — collapse it into one model

Speech-to-speech realtime: fewer stages, lower latency, less per-stage control.

Staged pipeline (default)

STT → LLM → TTS as separate stages. You pick the provider for each. Best when you need fine-grained control — voice cloning, specialist transcribers, model swaps.

Provider choiceTunable per stage~500ms

Realtime model

GPT-4o Realtime or xAI Realtime — one voice-native model handles speech in, speech out. Lower latency, more natural prosody, less plumbing — but no per-stage swaps.

Lowest latencyNatural prosodyLess control

Workspaces · clean separation per team

Switch in the sidebar. Members and data are scoped to each workspace.

My Workspaces

N

Northshore Dental

12 members · Pro

A

Acme Corp

3 members · Starter

P

PropertyGroup

8 members · Growth

A

Admin

Full control of the workspace.

✓ Billing & plan

✓ Members & roles

✓ Phone numbers

✓ Everything in Developer

D

Developer

Build and configure agents.

✓ Create & edit agents

✓ Run campaigns

✓ Manage SIP & webhooks

✓ Everything in Member

M

Member

Read-only access.

✓ View call logs

✓ View transcripts

✓ View reports

✗ No configuration changes

Create Agent

Four ways to start. Pick whichever matches your use case.

✨

AI Generator

Describe your use case — Omind Voice AI drafts a complete prompt and greeting message you can edit.

Fastest

📋

Sample Prompts

Browse our template library — fully configured agents for common use cases.

Library

📝

Custom

Start from a blank slate and write everything yourself.

Full control

📥

Import

Drag-drop a .json agent config — from a backup or another workspace. Validates secrets before import.

Restore / Clone

Agent Name

Maya · Northshore Appointments

Agent detail · eight tabs

Everything you can configure for this agent.

Maya · Northshore Appointments

ID: agt_3K7n2pYbN8wL

📝 Instructions

🗒️ Debrief Prompt

🧠 AI Model

🎙️ Voice & Speech

🛠️ Tools

⚙️ Advanced

▶ Simulate

📞 Call Me

System Prompt

You are Maya, the AI appointment coordinator at Northshore Dental Group. You help patients book, reschedule, and cancel appointments, and you answer common questions about clinic hours, insurance acceptance, and procedures. Be warm, concise, and professional. Always read back the patient's name, phone number, and appointment time before confirming. Transfer any insurance dispute or billing question to the front desk.

Persona, tone, guardrails, read-back, and escalation — covered next chapter.

Greeting Message

Hi, you've reached Northshore Dental. This is Maya, the AI assistant — how can I help today?

Default stack: GPT-5 Mini · Deepgram Luna · Nova-3 — change any stage in AI Model or Voice & Speech.

▼ Test in browser before going live

Anatomy of a strong system prompt

Seven sections, every time.

1

Persona & role

"You are Maya, the appointment coordinator…" Be specific.

2

Tone

Warm, professional, concise. State it explicitly.

3

Response guidelines

Short and conversational — never lecture. Avoid jargon.

4

Guardrails

What the agent must never do — share PII, make medical claims, promise what it can't.

5

Battle cards

"Are you a real person?" "Pricing pushback." Pre-loaded answers for hard moments.

6

Read-back protocols

Phone, email, address — read back, ask "Did I get that right?" before continuing.

7

Escalation & handoff

When to transfer, when to end, what to say while handing off.

Anatomy → behaviour

Read-back protocols

The single most important habit for production agents.

Phone numbers · group in 2s and 3s

Caller

Oh seven seven one zero five five five nine zero two.

Agent

Let me read that back: oh-seven-seven-one-oh · five-five-five · nine-oh-two. Did I get that right?

Email · spell with NATO phonetic

Caller

It's m dot patel at northshore.

Agent

So that's M as in Mike, dot, P-A-T-E-L, at, N-O-R-T-H-S-H-O-R-E dot com. Correct?

Battle card example · "Are you a real person?"

Agent: "I'm Northshore's AI assistant — but I can help you with most appointment and clinic questions. If anything needs a human, I'll transfer you straight away."

First message and the iteration loop

The first message recipe

Name the company
Identify the assistant by name
Ask how you can help — open-ended
Keep it under 3 seconds when spoken

"Hi, you've reached Northshore Dental. This is Maya, the AI assistant — how can I help today?"

Iterate with Test Agent

Open Test Agent · Voice mode
Try the happy path
Then the awkward edges — confused caller, interruptions, pricing pushback
Listen for moments where it sounds wrong
Edit · save · test again

⚡ Most prompts are at v5 before they're production-ready.

Same anatomy · two industries

Dental clinic vs. real estate agency.

Maya · Dental

Northshore

Job: Book, reschedule, cancel appointments. Answer FAQs.

Guardrails: No clinical advice. No billing disputes without verification. No PII without checks.

Read-back: Patient name · phone · appointment time.

Escalate: Insurance disputes & billing → front desk.

First message: "Hi, you've reached Northshore Dental. This is Maya — how can I help today?"

Alex · Real estate

PropertyGroup

Job: Qualify enquiries. Answer listing questions. Book viewings.

Guardrails: No prices outside listing range. No market predictions. No viewing without calendar check.

Read-back: Property reference · viewing date/time · contact details.

Escalate: Offers, negotiations, contract questions → human (warm transfer).

First message: "Hi, this is Alex from PropertyGroup, happy to help with your enquiry."

Same seven sections · different domain

Three common mistakes

✗

Too short

A two-line prompt produces a vague agent. Aim for a few hundred words, structured.

✗

No escalation path

Without a clear handoff rule the agent will either over-promise or refuse useful help.

✗

Vague tone

"Be helpful" is not a personality. Show, don't tell — give examples of phrases the agent should use.

Dynamic variables · one prompt, every caller personalised

Template (system prompt fragment)

Hi {{patient_name}}, this is Maya from {{clinic_name}} calling to confirm your appointment on {{appointment_date}} at {{appointment_time}} with {{doctor_name}}.

Resolved at call time

Hi Sarah Patel, this is Maya from Northshore Dental calling to confirm your appointment on Wednesday 6 May at 2:30pm with Dr Ahmed.

✓ Variables substituted before the call dials.

📊

CSV upload

Each column → variable. Each row → one personalised call.

⚡

API request

Pass variables in the outbound call request body.

📚

Data Collections

Pre-loaded datasets mapping phone numbers to attributes.

Configure the model stack

Mix and match LLM, TTS, and STT providers — swap any stage without touching the others.

Language ModelProviders: Open AI · xAI · Groq

🤖

OpenAI

GPT-5 Mini

DEFAULT

🤖

OpenAI

GPT-5 · GPT-4.1

⚡

xAI

Grok 4 · Grok 3

🎯

Realtime

GPT-4o · xAI

Creativity

FocusedCreative

Max Response Length

ShortLong

Pre-generate Responses

On

Service Tier (OpenAI) — Default Auto Flex Priority · Turn Detection (realtime) — Silence-Based Meaning-Based + Eagerness: Auto · Patient · Balanced · Responsive

Voice (Text-to-Speech)Also available: ElevenLabs · Rime · Murf · Cartesia · Sarvam · Inworld

🎵

Deepgram

Luna · warm

DEFAULT

🤖

OpenAI

TTS HD

⚡

xAI

Ara · expressive

Selected voice

Voice preview available for Deepgram · OpenAI · xAI. Select any voice and click Play Voice to hear it before saving.

Transcriber (Speech-to-Text)Also available: Sarvam (Indian langs) · Cartesia (90+ langs) · OpenAI STT

🎯

Deepgram

Nova-3 · 36+ langs

DEFAULT

⚡

xAI STT

stt-1

🤖

OpenAI

gpt-4o-transcribe

Noise Cancellation

Standard

Removes background sounds

Enhanced ✓

+ filters background voices

Telephony Optimized

Phone line audio · voice agents only

Knowledge base · agent-grounded answers

Upload PDF, TXT, or DOCX. Omind Voice AI indexes them automatically.

Uploaded files

Use uploaded files

📄

clinic-hours-2026.pdf

86 KB · PDF

✓

📄

insurance-accepted.docx

214 KB · DOCX

✓

⬆

Uploading…

…

Supports PDF · TXT · DOCX · drag-and-drop or browse

How RAG works in a call

Caller asks: "Do you accept Aetna?"
Agent embeds the question
Vector search hits insurance-accepted.docx
Top passages join the LLM context
Agent answers — grounded, not guessed

⚡ Adds ~80ms · Toggle "Use uploaded files" off for agents that don't need it.

Tools tab · three sections

Call Control · Actions · Integrations

Call Control

📞

End Call

Hangs up cleanly when the conversation is finished.

📵

Detecting Answering Machine

Identifies voicemail; switches behaviour automatically.

Actions

🔀

Transfer Call

Cold or warm transfer to a human or another number.

📅

Look Up Availability

Checks open slots for appointment-style flows.

✅

Confirm Appointment

Books a slot once the caller agrees.

Integrations

🔌

Custom Functions

Register webhooks the agent calls mid-conversation — covered next.

Tools turn talk into action.

Custom function · webhook

Definition

Name

check_appointment_slot

Description

Check open slots for a date/time at Northshore Dental.

Webhook URL

https://api.northshore.com/calendar/check

Parameters

{ "date": {"type": "string"}, "time": {"type": "string"}, "doctor": {"type": "string"} }

Live during the call

Caller

Can I see Dr Ahmed on Wednesday afternoon?

Agent

→ check_appointment_slot(date="2026-05-06", time="afternoon", doctor="Dr Ahmed")

Agent

← 2:30pm and 3:45pm available

Agent

Dr Ahmed has 2:30pm or 3:45pm open on Wednesday — which works better?

SSRF-protected 10s timeout Host allowlist

Transfer Call · configuration options

Cold or warm · with whisper, hold music, and exit delay.

Cold transfer

Faster

Agent says goodbye and bridges directly. No briefing.

1

Agent: "Connecting you now."

2

Caller bridged to human. Agent leaves.

Warm transfer

Recommended

Agent calls destination first, whispers context, then bridges caller in.

1

Dial destination · play on-hold music

2

Whisper: "I have Sarah — insurance question…"

3

Bridge caller. Three-way message plays. Agent exits.

Transfer settings

Speak Before Execution

Whisper MessagePrompt

On-hold Music

Three-way MessageStatic

Disconnect After Handoff

Agent Exit Delay5 s

How many seconds the agent stays on the line after handoff before dropping — a small delay ensures the transfer completes smoothly.

Ringing Timeout30 s

Transfer Attempts2

Test Agent · two modes

Chat for fast iteration. Voice for the real thing.

Live transcript

Connected

You

Hi, I'd like to book a cleaning.

00:02

Agent

Of course — could I get your name and a phone number?

00:04

You

Sarah Patel, oh seven seven one zero five five five nine zero two.

00:08

Agent

Let me read that back: oh-seven-seven-one-oh, five-five-five, nine-oh-two. Did I get that right?

00:13

Test Agent ✕

💬 Chat

🎙️ Voice

Listening…

Speak naturally — your microphone is captured in real time.

Omind Voice AI Web Widget

Embed once. Visitors talk to your agent from your website — no phone, no app.

One snippet · voice + chat · same agent

Simulate · test without dialling

An AI caller talks to your agent — no real phone, no minutes consumed.

Providers used

Agent LLM

GPT-5 Mini (OpenAI)

Caller LLM

gpt-4o-mini (OpenAI)

always fixed

Analysis LLM

gpt-4o-mini (OpenAI)

always fixed

Agent TTS

Deepgram Luna

Caller TTS

tts-1-hd / nova

always fixed

Knowledge Base

LlamaIndex · vector search

New Scenario

Scenario name

Anxious patient — reschedule

Caller identity

Nervous patient, first visit

Caller goal

Reschedule Wednesday 2:30pm appointment

Max turns

12

Post-Call Analysis

Score: 8 / 10

Goal achieved

✓ Yes

Tool usage

look_up_availability called ×2

Strengths

Read-back protocol followed correctly. Warm tone maintained throughout.

Improvements

Confirmation step was skipped before booking. Add "Did I get that right?" to prompt.

Debrief Mode · Structured prompting

Structured prompting with built-in behaviour criteria — next-gen alternative to the single long prompt.

Agent Debriefs

Structured prompting · behaviour criteria

Instructions tab is disabled while Debrief Prompt is active. Turn off Debrief to use the Instructions tab.

DEBRIEF PROMPT TAB

Greeting Message

Hi, you've reached Northshore Dental. This is Maya — how can I help today?

Role

Objective

Call Flow

Rebuttals

You are Maya, the AI appointment coordinator at Northshore Dental Group. Be warm, concise, and professional.

Debrief Criteria · 6 groups

Opening & Identity 4 criteria ▾

Proper greeting and introduction DEFAULT

Discloses AI identity when asked DEFAULT

Conversation Quality 5 criteria ▸

Emotional Quality 3 criteria ▸

Accuracy & Data 4 criteria ▸

Each criterion has an enable toggle, a DEFAULT badge when on by default, and an EDITED badge when you've customised the prompt. Six groups total: Opening & Identity · Conversation Quality · Emotional Quality · Accuracy & Data · Safety & Compliance · Situational.

Three ways to get a phone number

⚡

Buy new number

Search by country & area code. Vonage default, Twilio available. Live in seconds.

Fastest

🔁

Port your number

Keep the number your customers know. Contact our Telephony team — typically about a week.

Keep your number

🔌

Connect via SIP trunking

Bring your own provider for compliance or pricing reasons.

Advanced

📞 Porting note

Port-in goes through our Telephony team — they coordinate with your source carrier. Typical timeline is around a week, but it depends on your carrier, region, and number type. Some ports take longer. Contact us via the Documentation menu to start.

Connect a custom SIP trunk

Inbound trunk

SIP URI

sip:sayin-in@trunk.northshore.com:5060

Allowed source IPs

198.51.100.0/24, 203.0.113.5

Authentication

northshore_user

••••••••••

Outbound trunk

Destination URI

sip:trunk.northshore.com:5060

Caller ID

+44 7700 900142

Max concurrency

25 simultaneous calls

Quick Dial

Upload a CSV → dispatch a personalised outbound batch.

Batches

12

+3 this week

Dialled

5,284

+842 today

Connected

3,847

72.8% answer rate

Running

2

Active now

Batch	Agent	Progress	Status	Schedule
April Reminders · Northshore	Maya	576/847	Running	Now
Insurance Renewals	Maya	102/300	Running	Now
Patient Survey · May	Maya	—	Scheduled	Tue 09:00
New patient onboarding	Maya	450/450	Completed	Yesterday

847 personalised calls · automated.

CSV → calls in three steps

1 · Upload CSV

phone,patient_name,appointment_date,appointment_time,doctor_name +447710555902,Sarah Patel,2026-05-06,14:30,Dr Ahmed +447831200134,Tom Brennan,2026-05-06,15:00,Dr Patel +447900122481,Aditi Rao,2026-05-07,10:15,Dr Ahmed

2 · Pick agent & schedule

3 · Watch it run

14:42

Sarah Patel · ✓ confirmed

14:38

Tom Brennan · ✓ rescheduled to Friday

14:31

Aditi Rao · voicemail · will retry

14:28

→ dialing next…

Personalised per row Cancellable Live progress

Predictive Dial · the campaign engine

Timezone-aware · DNC-compliant · disposition-tracked.

🌍

Timezone-aware dialing

Every prospect called within their local 9am–9pm. No after-hours violations.

🚫

DNC enforcement

Built-in DNC list. Upload your own. State-level exclusions and holiday filtering.

⚙️

Cadence controls

Retry limits, retry intervals, dialling rate caps. Prevent over-dialling.

🏷️

Custom dispositions

Interested · Not Interested · Callback · Wrong Number · No Answer.

📅

Callback scheduling

Agent collects time, contact re-enters the queue automatically.

📧

Automated reports

Scheduled campaign reports delivered by email.

Campaign in flight

Insurance Renewals · 18,420 contacts

Active

Active dials

38

Answer rate

28.4%

Drop rate

1.2%

Pacing

3.5×

Disposition mix · last 24h

Interested32%

Callback Scheduled18%

Not Interested22%

No Answer21%

Wrong Number7%

Compliance & pacing

Local 9–9 enforced✓ ON

DNC scrub✓ Daily

State exclusions3 states

Holiday skip✓ ON

Max 3 retries · 4h apartCadence

Live Dialer Status Report

One row per active campaign · auto-refreshes every 5 seconds.

Campaign

Active Prospects

Callable Prospects

Renewals · Q2 2026

Maya · running

12

2,418

Appointment Reminders · May

Maya · running

8

643

Lead Follow-ups · West

Alex · running

5

1,206

NPS Survey · April cohort

Maya · finishing

2

17

Auto-refresh · 5s Use during campaign launches to confirm the queue is moving and no campaign is stuck.

Call Logs

Every call · three channels · searchable and exportable.

📞 Phone

🌐 Web

💬 Chat

Call ID	Direction	Agent	Customer Number	Disposition	Time & Duration	Details
`cl_a8b2…`	Inbound	Maya	+44 7710 555 902	Booked	14:42 · 3:42	👁 view
`cl_a8b1…`	Inbound	Maya	+44 7831 200 134	Resolved	14:38 · 1:12	👁 view
`cl_a8af…`	Inbound	Maya	+44 7900 122 481	Transferred	14:31 · 5:08	👁 view
`cl_a8ae…`	Outbound	Maya	+1 415 555 0934	Voicemail	14:28 · 0:42	👁 view
`cl_a8ad…`	Inbound	Maya	+44 7700 408 217	Resolved	14:22 · 2:55	👁 view
`cl_a8ac…`	Inbound	Maya	+44 7889 663 005	Transferred	14:18 · 4:21	👁 view

Call Details

Three tabs · Transcriptions · Conversation Analysis · Call Details

`cl_a8b2` · +44 7710 555 902 · 3:42

Booked

📝 Transcriptions

📊 Conversation Analysis

ℹ️ Call Details

▶

1:24 / 3:42

Caller

Hi — I'd like to book a cleaning.