38 agents ready to dial — reception, collections, KYC, support, screening

Phone calls answered in under a second,
in the language the caller switches to.

Sphoro Voice runs your inbound and outbound calls in fifteen South Asian languages — Hindi, Tamil, Telugu, Marathi, Bengali and the rest — each with its own prompt, voice and recognizer, and a caller who opens in English and finishes in Hindi is followed mid-call rather than restarted. A human notices delay above roughly 800 milliseconds, so every stage streams into the next instead of waiting for it. Your carrier, your model keys, your account if you want it there.

Start free — 60 minutes Book a 30-min demo

60 minutes free · No credit card · 38 agent templates · Self-hosting available

voice.sphoro.com/portal/calls/call_8f21c4b0 Live calls · 41 of 200

+91 98••• 41207 → Reception

Twilio · en-IN · agent ag_2rk9 · recording on
Speaking
Transcript barge-in at +1.2s
0msDo you have anything on Saturday morning?
+310msAbsolutely — I have a 9:15 and an 11:30 open on Saturday, and I can also…
+1.2sActually, make it Sunday.
+1.2sgeneration cancelled · TTS stream closed · buffered audio flushed
+1.5sSunday works — 10:00 or 2:30?
One turn · caller stops speaking → agent starts speaking
VAD / endpointing150ms
final STT120ms
LLM first token300ms
TTS first audio120ms
network / jitter80ms
< 800ms target, achieved by overlapping every stage
Provider chains probed every 5 min
STT deepgram → assemblyai → speechmatics
LLM anthropic → openai → groq
TTS elevenlabs · withdrawn → cartesia · owns this voice
< 800msconversational latency budget
15languages, switchable mid-call
38agents with the prompt written
69speech and language vendors, chained
Agent library

Thirty-eight agents with the prompt already written.

A blank agent is a text box and a shrug. Nearly every phone agent anyone builds is one of a few dozen well-understood things, and the difference between a good one and a bad one is almost entirely the system prompt — which is exactly the part a blank form asks a newcomer to invent. Each of these arrives with a real prompt, a greeting, the filler line that covers a slow lookup, and the tools it expects to find.

Loan Servicing LineInbound

Inbound line for existing borrowers — payments, fees, arrears, settling early.

Payment Recovery AgentOutbound

Respectful outbound calls about overdue invoices.

Loan Payment ReminderOutbound

Outbound reminder with a keypad option to pay or speak to someone.

Failed Payment Follow-UpOutbound

Outbound call after a direct debit is returned or a card is declined.

Identity Check Follow-UpOutbound

Chases outstanding identity or address documents, without ever sounding like a scam.

Card & Account SupportInbound

Lost cards, unrecognised transactions, statements and payments.

Loan Enquiry & EligibilityInbound

Answers borrowing questions, explains the process, books time with an adviser.

Insurance Claim IntakeInbound

First notice of loss — takes the facts without judging the cover.

Inbound Lead QualifierInbound

Qualifies inbound enquiries and routes the good ones to a rep.

Outbound Sales RepresentativeOutbound

Outbound calls to book meetings, with a clean exit when it is a no.

Appointment SetterOutbound

Books demos with warm leads and confirms the details.

Real Estate Lead Follow-UpOutbound

Follows up property enquiries, qualifies budget, books viewings.

Customer SupportInbound

Fields support enquiries, answers from your documentation, escalates.

Technical Support AgentInbound

Walks through troubleshooting one step at a time, escalates with detail.

Order Status & TrackingInbound

Order lookups, delivery estimates, and the start of a return.

Front Desk ReceptionistInbound

Answers the main line, routes to the right department, takes messages.

Healthcare ReceptionistInbound

Medical front desk — appointments, intake details, triage to a human.

Hospitality ConciergeInbound

Hotel front desk — guest requests, local recommendations, issues.

After-Hours Answering ServiceInbound

Takes messages outside business hours and escalates real emergencies.

Keypad MenuNo AI

Answers, offers numbered options, and routes on the keypress.

Keypad Menu (no AI)No AI

A pure phone menu. No speech recognition, no language model, no per-call cost.

Forwarding Line (no AI)No AI

Answers, says who is picking up, and puts the caller straight through to a person.

Appointment SchedulerInbound

General-purpose booking — checks availability, books, confirms.

Appointment ReminderOutbound

Outbound confirmations and rescheduling for upcoming appointments.

Salon & Spa ReceptionistInbound

Service pricing, bookings, and stylist requests.

Restaurant HostInbound

Reservations, party sizes, dietary notes, and opening hours.

Home Services QualifierInbound

HVAC, plumbing and electrical — qualifies the job and books a technician.

Property ManagementInbound

Tenant calls — maintenance requests, rent questions, emergencies.

Restaurant Phone OrdersInbound

Takes pickup and delivery orders with customisations and totals.

IT Help DeskInbound

Internal IT — password resets, VPN, access requests, software faults.

Employee Benefits & HR HotlineInbound

Answers staff questions on leave, benefits and payroll.

New Hire Onboarding AssistantInbound

Walks a new starter through day one and what they need to do.

Internal Knowledge AssistantInbound

Answers process and policy questions from internal documentation.

Customer Feedback CollectorOutbound

Outbound follow-up for ratings and open-ended feedback.

User Research InterviewerInbound

Qualitative discovery interviews — open questions, never leading.

Language Practice TutorInbound

Conversation practice that adapts to level and corrects gently.

Training Roleplay CoachInbound

Plays a realistic counterpart for practice, then scores the attempt.

Assessment & Quiz AgentInbound

Conducts a spoken assessment, scores it, and explains the result.

Every one is a starting point rather than a cage — the prompt, the voice, the tools and the languages are yours to change the moment it exists. Tools are named, not assumed: a template that wants book_appointment is filtered against what your account has actually registered, so it never creates an agent with a broken hand-off in it.

Languages

A caller who changes language mid-sentence is not an edge case here.

Every language an agent speaks is its own entry, with its own prompt, its own voice, its own synthesizer and recognizer and its own greeting. There is no separate field for "the" language anywhere on the agent — the first entry is simply the one calls open in. Switch detection on, and the agent follows a caller from one language to another partway through a call, voice and prompt and all.

हिन्दीHindi
বাংলাBengali
தமிழ்Tamil
తెలుగుTelugu
मराठीMarathi
ગુજરાતીGujarati
ಕನ್ನಡKannada
മലയാളംMalayalam
ਪੰਜਾਬੀPunjabi
ଓଡ଼ିଆOdia
অসমীয়াAssamese
اردوUrdu
नेपालीNepali
සිංහලSinhala
EnglishEnglish
Hearing it, not just speaking it Multilingual is one feature with two halves, and the second is the one usually missing. Answering in Tamil is easy; noticing that the caller moved to Tamil, and having a voice that pronounces it and a prompt written in it, is the half that decides whether the call survives the switch.
Hinglish is a thing callers do, not a typo The table is India-first because the market is. A generic BCP-47 library treats hi-IN and is-IS as equally likely and leaves out the facts that actually matter on the line: that Hindi and Urdu are mutually intelligible in speech but written in different scripts, and that a Marathi speaker will accept a Hindi voice where a Tamil speaker will not.
One entry per language, compared properly An operator types "Hindi", a recognizer reports hi, a synthesizer advertises hi-IN, a carrier passes hi_IN in a SIP header. Compared with string equality that is four languages and an agent that switches to Hindi while already speaking it. Every tag is normalized on the way in, once.
Anything else, carried by its tag Languages outside the table are first-class — they simply travel as their tag, costing nothing but the extra knowledge nobody has for them. Following is off by default, because an agent that starts changing language because a field appeared in a struct is a change nobody opted into.
PATCH /v1/agents/agt_123
// Calls open in the first entry, and its prompt is the agent's.
{
  "languages": [
    {
      "code": "en-IN",
      "system_prompt": "You are {{company}}'s collections agent.",
      "greeting": "Hello, this is {{company}} about your account."
    },
    {
      "code": "hi-IN",
      "voice_id": "hindi-voice",
      "greeting": "नमस्ते, मैं {{company}} से बोल रहा हूँ।"
    },
    { "code": "ta-IN", "voice_id": "tamil-voice" }
  ],

  // "off" is the default. "follow" moves language mid-call.
  "language_detection": "follow"
}

A voice is required on every entry after the first, and a prompt only on the first. The two missing halves fail differently: a missing prompt can be derived from the first, a missing voice cannot be derived from anything. Voices & languages

How it works

Answering a real number, this afternoon.

The portal and the API are the same three steps against the same objects — the portal is a client of the public API, not a privileged path into it. Whatever you build by clicking can be read back, diffed and recreated by a script, which is the only version of "no-code and code" that survives contact with a second environment.

Start from a template

Pick one of the thirty-eight and you have a working agent: prompt, greeting, filler line, temperature and the tools it wants. Add the languages it should speak and a voice for each.

Talk to it in the browser before it ever touches a phone line — the Build page dials it over the same runtime a real call uses.

Point a number at it

Bring your own carrier over SIP, or a number from Twilio, Vobiz or Plivo. Numbers are yours and stay yours; the platform is the thing behind them, not the thing that owns them.

Outbound is the same agent with a list in front of it: upload it, see what is wrong with the list before a single number is dialled, then let the dialer pace it.

Watch what it did

Stereo recording with caller and agent on separate channels, full transcript, per-stage latency, CSAT, and cost per call broken down by vendor — so a bad call is a thing you can open rather than a thing you are told about.

Every call fires a signed webhook when it ends, so none of this has to be somewhere you go and look.

Or skip the portal entirely — the same three steps are four API calls, in curl, TypeScript, Python, Go, PHP or Ruby. See the API

The one number

Voice AI lives or dies on the gap between turns.

You cannot buy your way under 800 milliseconds by waiting for each stage to finish and then starting the next — the arithmetic does not close. The stages have to overlap, and that is an architecture decision, not a tuning one.

Sequential
  • Wait for the caller to stop. A fixed silence timeout, usually 700ms, because nothing is watching for the end of a sentence.
  • Wait for the transcript. The recognizer finalises, then hands over.
  • Wait for the whole reply. The model finishes writing before a synthesizer is asked for anything.
  • Wait for the audio. Synthesis runs to the end of the sentence, then playback starts.
≈ 1,900ms — the pause everybody has heard
Overlapped
  • Turn detection fuses three signals. Silence, punctuation and an end-of-utterance score, instead of a fixed timeout — the cheapest 200 milliseconds on the board.
  • Generation starts on the partial. The model is already writing while the recognizer confirms the last few words.
  • Synthesis starts before the model finishes. A sentence aggregator releases whole clauses to text-to-speech as they form.
  • The state machine is lock-free. A small compare-and-swap machine — IDLE → LISTENING → THINKING → SPEAKING — so the goroutine carrying live media never blocks on a mutex to learn what the call is doing.
< 800ms — the budget in the panel above

Falling silent is three things at once.

Barge-in is the feature most platforms get wrong, and they get it wrong in the same place. Cancelling the generation is the obvious part. Closing the synthesis stream is the part people remember second. The one that is missed is discarding audio that is already buffered but not yet played — skip it and the agent keeps talking over the caller for a second and a half after it was interrupted, which is the exact moment a caller decides they are talking to a machine.

How a turn is actually run
Interruption is an event, not a poll.The media goroutine raises it the moment voice activity is detected during SPEAKING, so the cancel does not wait on the next tick of anything.
Keypad input never goes through the recognizer.Digits arrive on their own channel, because a keypress is an exact input with an exact meaning — and callers reach for the keypad precisely when speech recognition is struggling.
A slow tool never becomes dead air.Function calls run out of band behind a filler phrase that is served from cache, so the lookup happens while the caller is being told it is happening.
Every stage is measured, per call.Per-stage latency percentiles, token spend and cost breakdown land in analytics through a pipeline that never stalls the media path — overflow is dropped and counted.
Agent modes

Not every phone line needs a language model.

A business buying a phone system is not only buying the clever part of one. Three modes, chosen per agent — because routing "press 1 to confirm" through a model buys nothing but latency, spend and variation.

conversation

The full agent

Listens, thinks, answers, and can act — knowledge, tools, workflows, transfers. Any keypad entries the agent defines are still answered from the menu without troubling the model.

keypad

A menu and nothing else

No recognizer, no model, no network call of any kind inside a turn. The menu answers in milliseconds, costs nothing per call, and keeps working on the day a vendor is down or out of quota.

Digits with no entry still fall through to the model, so an agent can take "press 1 for X" deterministically and handle anything else it is told.

normal

A plain telephone line

Answers, reads its greeting, and puts the caller through to a number. Nothing thinks. For the overflow line, the closed-for-the-holidays number, the one on the back of the card that has always rung a desk.

With nowhere to hand off to it is an announcement line — which is the only honest thing a forwarding line with no destination can do, and a use of its own.

Platform

Everything a real call needs, and nothing it does not.

A live conversation is stateful and streaming; agent configuration and billing are neither. The two planes are split at the architecture, which is why one pins to a session and the other scales horizontally without either pretending to be the other.

Agent runtime

One session object orchestrates a single live call: the streaming pipeline, barge-in, DTMF, transfers, and tool calls executed out of band so a slow lookup never becomes dead air.

Knowledge & retrieval

Grounding runs on sphoro.kb, a separate product this platform is a customer of — integrated over its HTTP API, with its own tenant and key per account, so a bug here cannot reach another tenant's documents.

It is allowed 250ms inside a turn and fails open: a knowledge base that is slow costs a caller a less-grounded reply, never silence.

Function calling

Look up, authorize, parse, validate against JSON Schema, sandbox, run. Every problem in a malformed call is reported at once, and a failure comes back to the model as text it can act on rather than an exception that ends the turn.

Workflows

Book the appointment, update the CRM, send the email, fire the webhook — declarative steps with conditions, retries and stable idempotency keys. Any workflow can be exposed to the model as a single validated function.

Campaigns & dialer

Upload a list, preview what is wrong with it before a single number is dialled, then let the dialer pace it: calls per second the carrier tolerates, a bounded backlog, retry-versus-abandon, and per-tenant fairness.

The list lives in Postgres, never in memory — half a million rows cost the same as five hundred, and a restart resumes rather than repeats.

Voice lines, rendered once

The greeting, the filler, the menu replies and the no-match re-prompt are byte-identical across ten thousand calls. They are rendered once, cached and replayed — so the greeting has no vendor round trip in front of it.

Or upload your own recording: a real voice actor takes precedence over anything synthesized.

Recording & transcripts

A stereo WAV — caller left, agent right — because the answer to "what actually happened" nearly always hangs on who spoke over whom. Caller audio is padded to wall clock; agent bursts are placed where they were played, not where they arrived.

Analytics & CSAT

Per-stage latency percentiles, token usage, caller satisfaction, and cost per call broken down by component — on a non-blocking pipeline that never stalls the media path.

Compliance, before the dial

A do-not-call suppression list, consent records per number, and a pre-flight check you can call before placing a call — so the answer arrives before the dial rather than in a complaint afterwards.

Vendors & failover

Sixty-nine vendors, three chains, and nothing that assumes any of them.

The runtime depends on interfaces, never on a vendor. Every stage is a chain even with one member, so adding a provider is a configuration change rather than a code change — and a vendor having a bad afternoon degrades to the next one instead of dropping the call.

16 text-to-speech
CartesiaElevenLabsDeepgram Aura OpenAIRimeLMNT PlayHTHumeSpeechify SarvamGoogleAzure + 4 more
49 language models
AnthropicOpenAIGemini GroqCerebrasMistral DeepSeekxAITogether FireworksOpenRouterOllama vLLM+ 36 more
4 speech-to-text
DeepgramAssemblyAI SpeechmaticsGladia
Every vendor is asked, on a schedule, whether it still works.A key does not fail when it is saved. It fails when the account runs out of credit, when a plan will not cover the voice an agent asked for, when somebody rotates it in the vendor's console and not here — and every one of those looks identical from this side until a caller hears nothing. So a probe runs every five minutes: a few characters of synthesis, a one-token completion. Two consecutive failures withdraw the vendor; a permanent error withdraws it at once.
The provider that owns the agent's voice goes first.Chain order says which vendor a deployment prefers in general; it cannot say which one an individual agent's voice belongs to. Following the order alone produced an agent that answered perfectly well in the wrong voice, and nothing errored. Now a voice id outranks the order.
Failover survives the vendor accepting the job and then refusing it.A synthesizer that opens the session and only then rejects the voice used to cost the rest of the sentence. The chain now resumes mid-turn from where it left off, and never re-asks a provider that already said no.
Keys live in the portal, sealed.Vendor credentials are entered on a screen and sealed with one master key, so a deployment sets exactly one secret by hand and the rest is an operator's job rather than a redeploy. Per-vendor default voices and models come with them.
# Order is priority, and it is explicit. A vendor whose key
# is present is a vendor that is wired in — there is no
# second enable flag to forget.

STT_PROVIDERS=deepgram,assemblyai,speechmatics
LLM_PROVIDERS=anthropic,openai,groq
TTS_PROVIDERS=cartesia,deepgram-aura,elevenlabs

# A configured vendor missing from the order is appended
# rather than dropped: a key that is set and then ignored
# is worse than an unexpected ordering.

# One secret by hand. Every vendor key moves to the portal
# behind it.
SPHORO_SECRET_KEY="…"
Telephony

Your numbers, and one stack behind all of them.

Six carriers speak to the same telephony core; only the webhook signature and the media framing differ. We label each one by what it has actually done, not by what the code supports — a carrier badge is a promise, and promising one nobody has dialled through is a support ticket with a sales pitch on it.

TwilioInbound and outbound, on live production traffic.
Live
VobizInbound and outbound, for India. Live production traffic.
Live
SignalWireInbound. Protocol matched and covered by tests, including tamper and replay.
Inbound
TelnyxInbound. Protocol matched and covered by tests, including tamper and replay.
Inbound
VonageInbound, with its own answer/event shape rather than TwiML.
Inbound
PlivoInbound. The outbound client is built and tested but deliberately not enabled.
Inbound
Binary media frames, not base64 in JSON.Caller audio rides the realtime socket as a raw binary frame — a third fewer bytes on every packet of a live call. That is latency you can hear.
Naming an inbound-only carrier for an outbound call is refused.At validation, rather than quietly dialling Twilio instead and leaving somebody to work out why the caller ID was wrong.
Admission control, so a node refuses cleanly.A cap that is merely documented is not a cap. A held inbound call has no speech to process, so holding is far cheaper than talking — and much better than every call in progress degrading at once.
Hand off to a human, mid-call.When the request is beyond the agent, the live call is redirected to a number through the carrier — from a workflow, a tool, or a plain forwarding line.
# Point the number at the platform. Every carrier gets the
# same three routes; only the signature differs.

voice url   → https://voice.sphoro.com/v1/telephony/twilio/incoming
status url  → https://voice.sphoro.com/v1/telephony/twilio/status
media       → wss://voice.sphoro.com/v1/telephony/twilio/stream

# India, same shape:
answer url  → https://voice.sphoro.com/v1/telephony/vobiz/incoming
hangup url  → https://voice.sphoro.com/v1/telephony/vobiz/status
Developers

An API you can hold the whole shape of in your head.

REST for configuration, Server-Sent Events for watching a call, a WebSocket for driving one, and signed webhooks for everything that happens while you are not looking. The OpenAPI document is served by the API itself, so it cannot drift from the deployment you are talking to. 

curl https://voice.sphoro.com/v1/agents \
  -H "Authorization: Bearer $VOICEAI_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"name":"Reception",
       "mode":"conversation",
       "system_prompt":"You are a friendly receptionist."}'

# Watch it work. SSE with a ring buffer behind Last-Event-ID,
# so a dropped connection replays the turns it missed.
curl -N https://voice.sphoro.com/v1/calls/call_8f21c4b0/events \
  -H "Authorization: Bearer $VOICEAI_API_KEY"
Errors carry a stable code.RFC 9457 problem+json on every failure, with a machine-readable code and a request id that also appears on the response header. Branch on codes; messages are free to improve.
Validation reports every problem at once.Including fields it does not recognise. A payload with four mistakes takes one round trip to fix — and a typo like temperture is an error rather than an agent quietly running at a temperature nobody chose.
Idempotency keys, generated once.Every POST gets one outside the retry loop — so a network timeout on "create call" cannot place a second call. A retry replays the original response.
Event streams that survive a blip.Call events over SSE with a ring buffer behind Last-Event-ID: a client whose connection drops mid-call replays what it missed instead of losing the turns that happened in the gap.
Four clients, none with a dependency.An SDK's dependencies become your application's dependencies, and the first version conflict is yours to resolve. Adopting one of ours cannot force a Jackson, urllib3 or axios version on you.
Your own request log, in the portal.The last few hundred API calls your integration made, with status and request id — because "why am I getting a 401" should not require somebody to grep a production log.
Documentation OpenAPI spec
Security

Every URL in a voice platform comes from a customer or a model.

The crawler, the webhooks, the HTTP tools an agent can invoke — none of those addresses is yours. That is the threat model the outbound path is written against.

Egress that fails closed

Private, loopback, link-local, CGNAT and cloud-metadata addresses are rejected at connect time, redirects are re-validated, bodies are capped and non-HTTP schemes refused — which is what makes it immune to DNS rebinding rather than merely careful about it.

Keys that cannot be recovered

API keys are 256-bit random secrets stored only as SHA-256 hashes and compared in constant time. A database leak yields hashes, not credentials. Revocation is immediate.

Tenancy that is a partition

A namespace is a hard boundary, not a filter predicate — it fails closed without a tenant and cross-checks the authenticated identity. A dedicated test asserts one tenant's credentials can never resolve to another's.

Webhooks you can trust

HMAC-SHA256 over timestamp.body, so a captured request stops verifying once it ages out of a five-minute window. Multiple signatures during a rotation mean a secret can change without dropping a delivery.

Tools that cannot run wild

Deadlines, panic containment, per-tenant rate limits and concurrency caps, truncated output. The database tool takes a query name and parameters, never SQL — a "run this SQL" tool is an injection vulnerability with extra steps.

Permissions, not role strings

Nothing asks "is this user an admin"; it asks "may this user revoke a key". Every rule that stops somebody promoting themselves lives in one place, because a privilege-escalation bug is a single missing check somewhere. One-time codes on sign-in, and a durable audit log behind it.

Limits you can see coming

Token-bucket rate limits with X-RateLimit-* on every response and Retry-After on a 429 — so a client can slow down before it is throttled rather than after.

Consent, before the dial

The TCPA sets statutory damages per call with no cap; TRAI layers registration and a national registry on top. Neither is a policy code can interpret — what it can do is make suppression, consent and time-of-day checks impossible to skip.

Nothing fetched from a stranger

The server has no third-party dependency, and neither does any SDK. This page loads no font, no script and no image from another host — which is why it renders offline and has nothing to disclose about who watched you read it.

Pricing

Prepaid credit, and one number you can answer.

A customer should be able to ask "how much have I got left" and get one figure back, in one currency, that ties to a ledger they can read line by line. So the model is prepaid tokens against a rate card, not a subscription with estimated overage that cannot answer the question until the month closes. A better plan spends fewer tokens per minute — growing with the platform is never punished.

Trial

Prove it works, with no card.

Free
900 tokens included
≈ 60 minutes · $0.15/min
  • Every provider and every language
  • Knowledge bases, tools, and workflows
  • 1 agent · 2 concurrent calls
Community support
Start free
Starter

A first production line.

$49/month
6,000 tokens included
≈ 500 minutes · $0.12/min
  • Real numbers on Twilio, or Vobiz for India
  • Call recordings and transcripts
  • Webhooks with signed delivery
  • 3 agents · 10 concurrent calls
Email, next business day
Get started
Growth Popular

Multiple teams, real volume.

$199/month
22,500 tokens included
≈ 2,500 minutes · $0.09/min
  • Everything in Starter
  • Provider failover chains across vendors
  • Analytics history and CSAT
  • Role-based access and a full audit log
  • 15 agents · 50 concurrent calls
Email and chat, same business day
Get started
Scale

High concurrency, lowest rate.

$799/month
72,000 tokens included
≈ 12,000 minutes · $0.06/min
  • Everything in Growth
  • Bring your own provider keys
  • S3-compatible recording storage
  • 99.9% uptime commitment
  • Unlimited agents · 200 concurrent calls
Priority, 4-hour response
Talk to us
Enterprise

Your infrastructure, your terms.

Custom
Contract billing
No token charge on conversation
  • Everything in Scale
  • Self-hosted or private cloud
  • Data residency in the US or India
  • SSO, custom contracts, a named engineer
Dedicated, 1-hour response
Book a call

Top-ups are sold in packs of 1,000 tokens — $10 — and both figures are chosen rather than converted, so neither moves with an exchange rate. Every plan meters conversation minutes; model tokens and synthesis characters are counted but not charged. Moving between plans is done with us or by redeeming a plan code in the portal, so nobody changes tier by clicking the wrong button — mail info@sphoro.com and it takes a minute.

Questions

Before you start.

Anything not here, ask us directly — info@sphoro.com.

Do I bring my own carrier and model keys?+

You can, and on Scale and above that is the intended shape — you keep the commercial relationship with your carrier and your model providers, and Sphoro Voice is the runtime between them. On the lower tiers ours are used and metered as conversation minutes. Nothing about the platform assumes a particular vendor, because the runtime depends on interfaces rather than on any of them.

What happens when a provider goes down mid-call?+

Each of speech-to-text, the model, and text-to-speech is a chain rather than a single provider. A health probe asks every vendor every five minutes whether it still works, and two consecutive failures withdraw it before the next caller reaches it. If one fails during a turn, the chain resumes from where it left off rather than starting over — including the case where a vendor accepts the session and only then refuses the voice.

Can the agent actually do things, or only talk about them?+

Do things. Function calling validates the model's arguments against a JSON Schema and runs them in a sandbox with deadlines and per-tenant limits; the workflow engine composes those into multi-step automations with retries and idempotency keys; and a call can be handed to a human mid-conversation. A failure comes back to the model as text it can act on, so a mistyped booking becomes a follow-up question rather than a dropped turn.

Can I run outbound campaigns?+

Yes. Upload a list, see everything wrong with it before a single number is dialled, then let the dialer pace it against what your carrier tolerates — calls per second, a bounded backlog, retry-versus-abandon, and a refusal to outrun answering capacity, because a call placed with nowhere to answer it spends money to annoy a customer. Suppression and consent are checked before each dial, not after.

Where does it run?+

The data plane is geo-pinned — a live call is stateful and sticky, so it stays close to the caller. The control plane is stateless and scales horizontally anywhere. Current focus is India and the US; Enterprise adds data residency in either. Talk to us about other regions.

How do I evaluate it without a phone number?+

The whole platform runs locally with no network, GPU, database or cloud account — mock speech and language providers ship with it, and a call simulator speaks the carrier's side of the media-stream protocol so nothing on the server is stubbed. You can hear a barge-in work before you sign anything.

Tell us what the call is supposed to accomplish.

Start on the trial and hear a real turn — including the interruption — inside ten minutes. Or book thirty minutes and we will walk one end to end and be straight with you about what is ready and what is not.

Start free — 60 minutes Book a 30-min demo

No credit card · info@sphoro.com

Sphorosphoro.com — the parent company → BugCatchbugcatch.sphoro.com — bug capture and triage → Device Scorescore.sphoro.com — device trust signals →