Resonance Cascade - Product overview
Resonance Cascade is Shunya Labs' sovereign voice agent. It holds natural spoken conversations with your customers, uses your knowledge and systems while it talks, and gives you a complete, searchable record of every call.
| Measure | Typical |
|---|---|
| Typical reply time, phone and browser | 0.8 s |
| To stop talking when the caller cuts in | 0.2 s |
| Languages, with four English accents | 80+ |
| Channels from one agent: phone, browser, app | 3 |
Built for, among others
- Customer care and helpdesk
- Order and delivery updates
- Appointment and service booking
- Payment reminders and collections
- Lead qualification and sales calls
- Renewals and upgrades
- Onboarding and verification follow-ups
- Healthcare consultations
- Banking and insurance enquiries
- Travel and hospitality bookings
- Sales training role-play
- Candidate interviews
- Surveys and feedback
- Event and visit reminders
- and any conversation you can describe
Jump to a section
- 01 Overview
- 02 How a call works
- 03 Building agents
- 04 Languages & voices
- 05 Conversation
- 06 Interviews
- 07 Tools & integrations
- 08 Phone, browser & app
- 09 Calling at scale
- 10 After the call
- 11 Privacy & security
- 12 Workspaces & billing
- 13 Resonance vs Cascade
- 14 Reliability
How a call works
Every turn of the conversation follows the same path, and every call ends with a full record for your business.
Every turn
About 0.8 s from the caller's last word to the reply.
After every call
Every step runs on Shunya's own technology and servers.
Predictable by design
An agent's instructions, knowledge and tools are fixed for each version, so the same question gets the same kind of answer on every call. Every reply is checked before it is spoken: a price, date, name or promise that did not come from your information is held back, and the agent says it will find out instead.
Ready on day one
Start from a template, add your documents and systems, and try the agent with a test call that dials no one. The same agent then answers on the phone, in the browser and in your app.
Shunya's own technology
A sovereign voice agent: every part is Shunya's own, and your conversations are not sent to any outside AI service. It runs in Shunya's cloud, in your own cloud, or in a fully isolated network.
Building agents
An agent is set up once, in the Playground or through the API, and then answers every call the same way. It combines instructions, knowledge, tools and a voice.
Versioned agents, with their history
Every change is saved as a new version, and each agent’s history lists every version. A call always runs the version it started with, and any earlier version can be brought back. Prompts keep their own history too.
Deleting agents
An agent no longer needed can be deleted in one step: it leaves the agent list and takes no new calls, and the records of its past calls stay as they were.
Ready-made templates
25 to start from: a customer care desk and 24 agents for banking, ecommerce, healthcare, logistics and telecom, with sample details so a test call works at once.
Call inputs
Placeholders such as the customer’s name or order number are filled in for each call, and checked against their type before the call starts.
Knowledge base
Upload documents and keep them versioned. On every turn the agent looks up the passages that answer the caller’s question.
From SOP to agent
A standard operating procedure document becomes guided questions, a draft set of instructions and a call flow, ready to review.
Call flows
Fixed branches for steps that must always go the same way, such as identity checks. They are edited visually in the Playground.
Greeting and closing line
The first and last words are set exactly, in each language. On recorded calls, the greeting includes the recording notice.
Tone and speaking speed
Choose the agent’s tone and how fast it speaks.
Silence handling
If the caller goes quiet, the agent checks in twice, then says its closing line and ends the call. The wording and timing can be changed.
Call length
Calls last up to 30 minutes by default; an agent can allow up to two hours, for example for interviews or long consultations. At the limit the agent closes politely and the record states the reason.
Per-call settings
A single call can set its own voice, language, greeting, closing line and silence prompts, and can carry background information, such as the customer’s account details, for the agent to use.
Capturing details
The business defines the details it needs, such as a booking date or a reference number. The agent saves each one the moment the caller gives it, checked against its type.
Role-play agents
The agent plays a person, for example a customer, so that staff can practise real conversations.
Negotiation agents
The agent negotiates a rate within limits the business sets. The system enforces the limits, so the agent cannot agree to anything outside them.
Languages and voices
Each agent speaks in a chosen language and voice. The caller can move between the agent’s languages during the call.
Many languages
More than 80 languages, including English, Hindi, Hinglish, Bengali, Telugu, Tamil, Marathi, Gujarati, Kannada, Bahasa Indonesia, Japanese, Korean, Chinese, Arabic, Spanish, French and German. Hinglish is transcribed in the Latin script, the way people type it.
Language switching
The caller can move between English, Hindi, Hinglish and Bengali mid-call, and the agent carries on in the caller’s language without apologising or saying it can speak only one. An agent can also be held to a single language.
English accents
Indian, British, American and Indonesian-accented English, each with a female and a male voice.
Native voices
Bahasa Indonesia, French and German agents speak with voices native to the language.
Voice by gender
A single call can ask for a female or a male voice. The matching voice for the call’s language and accent is used.
Custom voices
A business can create private voices from short recordings of its own speakers. A custom voice speaks only after its speaker’s consent is on record.
Voice previews
Hear any voice, in any accent, before choosing it.
Speaking styles
One voice can speak in several styles: conversational, customer care, announcement or news. Each style is its own recording of the same speaker. Available for Meera, Oliver, Emily, Ethan, Grace, Budi and Sari, and chosen in the Playground beside the voice.
Even loudness
Every voice is levelled to the same loudness, so switching voices or languages never makes a call suddenly louder or quieter.
Numbers said properly
Amounts, dates and figures are spoken the way each language and country says them. In Hindi, Marathi, Gujarati, Bengali, Telugu and Tamil, dates, ordinals, amounts in lakh and crore with paise, and percentages are said in the language, and codes digit by digit. In Indian English, “₹5 lakh” is “five lakh rupees”. In every English accent, card endings, OTPs, PINs and account numbers are read digit by digit.
Conversation
The agent is built to behave like an attentive person on the line: quick to answer, easy to interrupt, and careful not to say anything it cannot stand behind.
Fast replies
About 0.8 seconds from the end of the caller’s sentence to the start of the reply, on the phone and in the browser.
Interruptions
The caller can cut in at any time. The agent stops within about a fifth of a second, then listens or carries on as the moment needs.
Natural turn-taking
The agent waits for callers who pause mid-thought: in the middle of a sentence, after a word that leaves it unfinished, between two sentences of one answer, and when the caller says “Alright” or “Okay” and keeps going. The reply is prepared while the agent waits, so it follows promptly once the caller has finished.
Short replies to short answers
“Okay” or “yes” gets one short reply, not a speech.
Corrections
When the caller corrects a detail, the new one replaces the old one for the rest of the call.
Long calls
Facts given early in the call are still used correctly many turns later.
“I can’t hear you”
The agent repeats its last answer, more simply.
Noisy lines
Works in street noise, crowded rooms and on telephone-quality lines. Short bangs and the agent’s own echo do not interrupt it.
Grounded answers
The agent does not invent prices, policies, bookings, dates, names, phone numbers or order status. It uses only what its instructions, knowledge or tools provide.
Honest about what it is
Asked whether it is a person, the agent answers honestly, in one light sentence, as a virtual assistant, then carries on with the call. It never claims to be a person. The wording, such as “digital assistant”, can be set for each agent.
Clean endings
One complete goodbye, then the call ends.
Interruptions your way
Callers can cut in, with low, medium or high sensitivity; or the agent finishes each sentence first, for notices that must be heard in full.
Reply style
Set how creative the agent may be and how long its replies run, and describe its tone and accent in your own words.
A gentle nudge
If both sides go quiet, the agent can prompt the caller after a pause you choose.
Today’s date and time
The agent knows the date and time in your time zone, and what today, tomorrow and the day after are, in the caller’s language.
Interviews
An agent can conduct a structured interview: it asks a list of questions one by one, handles the candidate's questions and pauses, and returns every answer.
Questions with every call
The business passes up to 100 questions with each call. The agent asks them one at a time, in order, exactly as written.
Repeats on request
If the candidate did not hear or asks again, the question is repeated word for word.
“I don’t know”
The answer is recorded as not known and the agent moves on to the next question.
Candidates’ own questions
A question about the role or the process is answered briefly from the agent’s documents, then the current question is asked again. If the documents do not cover it, the agent says so plainly.
Time to think
Candidates can pause to think. After 10 seconds the agent says “Take your time”; at 20 seconds it offers to repeat the question; after 90 seconds it moves to the next one. It never asks “Are you there?”
Pauses within an answer
Hesitations such as “umm” or “let me think” get no reply, and the whole answer is recorded.
No response
If three questions in a row get no response, the interview ends politely and those questions are recorded as not answered.
Live progress and follow-ups
The business can read the interview’s progress during the call and add or change questions that have not been asked yet.
Results for every question
Each question’s status and answer arrive on the call record and by webhook when the call ends.
Values
| Measure | Typical |
|---|---|
| Questions per call | up to 100, 500 characters each |
| Gentle prompt | after 10 s of silence |
| Offer to repeat | after 20 s |
| Move to the next question | after 90 s |
| End after no response | 3 questions in a row |
| Interview length | up to 2 hours |
| Settable per call | Every timing above can be set for each call, for example to give candidates longer for a hands-on task. |
Tools and integrations
During a call the agent can look things up and take actions in the business’s own systems, and send messages on the business’s channels.
Connect your systems
The agent calls the business’s web APIs during the call, for example to look up an order or book a slot. Tools can be imported from an OpenAPI file, and each one can be tested before use. Numbers and IDs the caller reads out are passed on in the format the system expects.
Secure connections
Common sign-in methods are supported: API keys, basic sign-in, OAuth 2 and signed requests. Tools reach only secure public web addresses.
Guided tool setup
A step-by-step guide in the Playground walks through setting up a tool with a worked example. Passwords and keys for tools are stored so that no one can read them back.
Safe actions
An action that changes something runs once, never twice, and is described as done only after the business’s system confirms it.
Agreed before acting
Before a change to the account, the agent reads back the details that matter (the amount and card for a payment, the e-mail for paperless) and acts only on the caller’s yes. Every request is tracked until it is done.
Natural waits
If a lookup takes a moment, the agent says so. Results are spoken in the caller’s language, and failures are reported honestly.
Lookups before the call
Information can be fetched before the call starts, from your systems or a secure web address, so the agent begins with the latest details.
Built-in actions
End the call, transfer to a person, raise an alert, schedule a callback, send a WhatsApp message or QR code, check the time and save a detail.
Message templates synced from Meta, delivery status and QR codes, sent from the business’s own WhatsApp Business account.
Alerts during the call
The agent can flag an issue to the team, by notification or webhook, without saying anything to the caller.
Refusals explained
When the business’s system declines a request and says why, such as a declined card, the agent tells the caller the reason and offers what is still possible. A declined action is never reported as done. If the system is not responding, the agent gives the tool’s usual failure message.
Your own details with every call
Each call can carry your own details, such as a customer’s name, account, fees or a role-play persona. Any agent uses them, and they fill placeholders such as {name} in its lines.
Phone, browser and app
The same agent answers on every channel, with the same instructions, voice and records.
Outbound phone calls
Calls are placed within a calling window, with retries, and with protection against dialling the same request twice.
Inbound phone calls
A phone number is answered by a chosen agent.
Your own carrier
A business can place calls through its own Plivo account and numbers.
Browser and app over WebSocket
A web page or mobile app streams the conversation to the agent over a secure connection that works once only.
WebRTC
A second way to connect from browsers and mobile networks. Strict office networks are reached through the standard secure port, and if WebRTC cannot connect the call continues over WebSocket. It adapts to weak mobile connections, so the caller is still heard clearly.
Transfer to a person
On the phone, the agent hands over and connects a colleague. In the browser, it books a callback.
Voicemail detection
The agent recognises carrier messages and answering machines, waits for the beep and leaves its message in full.
Ringback tone
Inbound callers hear a familiar ringing tone until the agent speaks.
Values
| Measure | Typical |
|---|---|
| WebRTC to WebSocket | after 15 s without a connection |
| Voicemail by silence | 1 to 30 s, set per agent |
| Wait for the beep | up to 15 s |
| Ringback tone | up to 20 s |
| Audio, browser | 16 kHz in, 24 kHz out |
One agent, every channel. Phone, browser and app calls share the same agent version, guards, recordings, analysis and billing, so a test in the browser behaves like a call on the phone.
Calling at scale
Campaigns, retries and live control for outbound calling, with consent and the time of day respected.
Contact campaigns
Upload contacts who have given consent and set a calling window, then start, pause, resume or cancel. Campaigns can be copied, and pending contacts removed.
Live progress
Figures for each campaign, and the queue of calls waiting over the last hour, six hours or day.
Callbacks and retries
Rules for each result, such as busy, no answer, failed or voicemail, decide when to try again. Callbacks the caller asks for happen inside the calling window.
Fair queueing
When every line is busy, outbound calls wait their turn in order, and inbound callers hear a short, polite message.
Live view and control
See every call in progress, open one to follow its conversation turn by turn as it happens, and hang up, send a line for the agent to say, or transfer.
Test calls
Run a complete call, from greeting to analysis, without dialling a real number and without being billed.
Contact lists as you have them
Upload a list with a phone, number or mobile column and any other columns; each column becomes that contact’s details for the agent. Preview exactly how the agent will open each call.
Calls placed once
A request sent twice with the same reference returns the first call, so no customer is called twice.
Test calls
| Measure | Typical |
|---|---|
| Length | up to 5 minutes |
| At once | 2 per workspace |
| Per day | 100 per workspace |
After the call
Every call leaves a complete record, ready to review, analyse and pass to the business’s own systems.
Call record
The full transcript, timing for every turn, tools used, what the caller agreed to, alerts, transfers, details captured and why the call ended.
Recordings
Caller, agent and combined tracks, with sensitive details bleeped. Shared through signed links that expire.
Re-transcription
After the call, the whole recording is transcribed again for a more accurate final transcript. It can also be requested at any time.
Call analysis
A summary, the caller’s sentiment, the outcome, follow-up actions and any fields the business wants pulled out of the conversation.
Quality evaluation
Each call is scored, with the share of the agent’s statements that were not backed by its information, the share of required details collected and any SOP steps missed. Strengths and failures are shown with the lines around them.
Insight reports
Across many calls: key insights, common questions, reasons for calling, where callers get stuck, objections, recommendations, agent performance and trends.
Analytics
Volumes, outcomes and durations; sentiment by day or week; the busiest hours; campaigns side by side.
Webhooks
Signed notifications for call events, sent to the business’s own systems and retried until accepted.
Alert rules
Rules that fire after a call or while it is running, for example on a low score or repeated interruptions, sent in the app, to Slack or by webhook.
Results with every notification
The end-of-call notification carries the summary and outcome, so your systems get the result without asking for it. A per-call callback address works too.
Find any call
Filter the call history by number, language, outcome, sentiment and length.
Privacy, security and data
Sensitive data is masked, consent is recorded, data is kept for stated periods, and every business’s workspace is kept apart from every other.
Sensitive data masking
Transcripts are masked and recordings bleeped, by the country of the agent’s language.
Verified first
Nothing about the account is looked up, said or changed until the caller has passed the agent’s identity check.
Recording notice
On recorded calls, the agent tells the caller that the call is recorded. Recording can be switched off per workspace.
Data retention
Recordings are kept for 90 days, then deleted. Playback links expire after a day.
Data erasure
Delete a single call, or every call, contact and message record for a phone number. Billing receipts are kept, as the law requires.
Deletion requests
A formal request to delete a workspace’s data is approved by Shunya, then carried out completely.
Consent for custom voices
A custom voice cannot speak until its speaker’s consent is on record, and consent can be withdrawn at any time.
Separate workspaces
Each business’s agents, calls, recordings and keys are kept apart. A key works only in its own workspace.
Shunya’s own technology
The agent runs entirely on Shunya’s own technology and servers. Nothing is sent to an outside AI service.
Logged staff access
When Shunya staff open a workspace to help, every entry and change is logged.
Masked by country
| Country / region | Categories masked |
|---|---|
| India | PAN, Aadhaar, phone, card, account, OTP, CVV |
| Indonesia | NIK, NPWP, phone, card, account, OTP, CVV |
| France, Germany | IBAN, phone, card, account, OTP, CVV |
| Other languages | every category |
Where data goes. Call data leaves Shunya only where the business sends it: its own tools and webhooks, its phone carrier and its WhatsApp account.
Workspaces and billing
Teams, groups of businesses, access and usage are managed from one Playground and one API.
Team members and roles
Invite colleagues with a role. Invitation links work once and can be withdrawn.
Parent and sub-client workspaces
A group or an agency can manage the workspaces of the businesses under it.
Feature controls
Browser calls, phone calls, campaigns and recordings are switched on per workspace.
The Playground
Four areas: Build for agents, voices and tools; Run for calls, tests and campaigns; Analyze for recordings, analytics and evaluations; and Settings.
One API key
The same Shunya API key a business uses for Shunya’s other APIs works for Resonance Cascade, in its own workspace.
Full API
Everything in the Playground is also available through a documented API, with signed webhooks for events.
Working example for developers
A complete example with a small server and a browser page, in plain JavaScript and React, covering microphone access, browser audio rules, reconnecting and busy lines.
Per-minute billing, to the second
Browser calls are measured from connection to end, and phone calls from answer to end. Unanswered calls are not charged.
Webhooks by agent
Up to ten notification endpoints, each with its own secret, events and agents.
Values
| Measure | Typical |
|---|---|
| Invitations | 7 days by default, up to 14 |
| Workspace levels | parent and sub-client |
| Billing unit | per minute, measured to the second |
Resonance and Resonance Cascade
Shunya offers two voice agents. Both are available to every client, and a business can use them side by side, one use case at a time.
| Product | Summary |
|---|---|
| Resonance | Quick to set up for straightforward conversations, with the widest choice of voices: 30, each speaking every language. |
| Resonance Cascade | For conversations that must be exact and connected to your business: answers grounded in your information, your systems used during the call, sensitive data protected, interviews, and a full record and analysis of every call. A sovereign voice agent, built entirely on Shunya's own technology. |
When something goes wrong
Calls meet silence, busy lines, failed systems and poor networks. This is what the agent does in each case.
| Situation | What happens |
|---|---|
| The caller says nothing | After 12 seconds the agent checks in, and again after a second silence. Then it says its closing line and ends the call. The record states the reason. |
| The caller cannot hear | The agent repeats its last answer, more simply. |
| Background noise or echo | Noise and the agent’s own echo are filtered out. Short bangs do not stop the agent. |
| The caller asks for a person | On the phone the agent transfers the call. In the browser it books a callback. |
| The caller asks whether the agent is a person | The agent says, in one light sentence, that it is a virtual assistant, then carries on. It never claims to be a person. |
| A business tool is slow | The agent tells the caller it is checking, and keeps the conversation going. |
| A business tool fails | The agent says it could not get the information. It never invents a result. |
| An action might run twice | Actions that change something run once only and must confirm success. |
| WebRTC cannot connect | After 15 seconds the call continues over WebSocket in the same session. Strict networks are reached through the standard secure port. |
| Every line is busy | Outbound calls wait in order. Inbound callers hear a short message. A browser or app gets a clear busy answer and can try again. |
| The number is busy or does not answer | Retry rules for each result decide when to call again. |
| An answering machine picks up | The agent waits for the beep, leaves its message in full and ends the call. |
| A webhook endpoint is down | Delivery is retried up to 12 times, and every attempt is in the history. |

