Skip to content

Building Voice Ninja: How Self-Hosting Our Voice Agents Cut Per-Minute Costs by Almost 60%

We started on Retell, moved to ElevenLabs, then deployed our own models to bring costs down. Voice Ninja's pricing fell by almost 60%, and clients with good hardware now run voice agents on their own systems with almost no AI cost per minute.

~60%lower Voice Ninja pricing after deploying our own models
Close to zeroAI cost per minute when self-hosted on a client's own hardware
3generations of the voice stack: Retell, ElevenLabs, our own
Phone and webone agent on a phone number and as a website widget
Key takeaways
  • Voice Ninja is our own product: a platform for AI phone agents and website voice widgets.
  • We built its voice layer three times: on Retell, then with ElevenLabs, and then on models we deploy ourselves.
  • We switched to cut costs. Deploying our own models reduced Voice Ninja's pricing by almost 60%.
  • For clients with good hardware, the AI cost per minute is close to zero, because the language model, speech recognition and speech synthesis all run on systems they own.
  • A managed platform is still the right place to start. Self-hosting pays once volume makes per-minute fees a real cost.

Voice Ninja is Wope's own platform for AI phone agents and website voice widgets. We built its voice layer three times: first on Retell, then with ElevenLabs, and then on models we deploy ourselves. We made the last move to cut costs. It reduced Voice Ninja's pricing by almost 60%, and for clients who already own good hardware we now run the models on their own systems, which brings the AI cost per minute close to zero. Phone numbers, and premium voices where a client wants them, are still paid for.

What Voice Ninja does

A business builds an agent, visually or through an API, and puts it on a phone number, on its website, or both. The agent listens, understands the caller with a language model, takes actions, and answers in a natural voice.

  • Phone agents and website widgets from the same agent, with the widget's colors, position and lead-capture fields set by the business.
  • A choice of language model, voice and language per agent, to balance quality, speed and cost.
  • Custom tool calls, so the agent can book appointments, look up orders or capture a lead in the business's own systems.
  • Agent-to-agent transfer, so a receptionist agent can hand a caller to a specialist agent.
  • Templates by industry: automotive, beauty and wellness, food and dining, healthcare, hospitality and travel, and real estate.
  • An analytics page for the leads each agent captures, and developer APIs and real-time WebSockets for teams that want to build on it.

Stage one: starting on Retell

We began on Retell. A managed voice platform bundles the hard parts of a phone call, such as speech recognition, turn-taking, the voice and the telephone line, behind one service. That is the fastest way to get a working agent onto a real phone number.

The general trade-off with any managed platform is that it decides what you can change, and you pay its price for every minute. Neither matters much on day one. The per-minute price matters a great deal as volume grows.

Stage two: integrating ElevenLabs

Next we integrated ElevenLabs. Working with a voice provider directly, instead of only through a bundled platform, gave us more say over how an agent sounds and let us offer a choice of voices to match a brand.

Stage three: deploying our own models to cut the cost per minute

With a managed platform or a voice provider, every minute of every call is billed. For a business that takes a lot of calls, that fee becomes the largest part of what a voice agent costs. Some of our clients also had capable hardware of their own sitting underused. So we deployed our own models and built a voice agent service we could host anywhere.

  • On a client's own systems, the AI cost per minute is close to zero. The language model, speech recognition and speech synthesis all run on hardware the client owns, so there is no AI provider to pay for each minute. What remains is the phone numbers, the cost of running the machine, and premium voices if the client wants them.
  • On Voice Ninja, pricing came down by almost 60%. The same self-hosted service runs the product, and we passed the saving on.
  • Speech is processed in-house. With the standard voices, a call is transcribed, understood and answered on the client's own systems instead of being sent to an outside AI provider.

A home care example. The same self-hosted setup can run the phone agents of the New York home care agency we work with: the language model, speech recognition and speech synthesis on hardware the agency owns, paid for once instead of per minute. Phone numbers, and premium voices where used, are still billed. The full story is in our home care agency case study.

The website widget shows how it works: it opens a real-time connection to the voice service, streams the visitor's speech as they talk, and handles the moment a person talks over the agent.

The trade-off is honest work: once you host the models, keeping them running, updated and fast is your job, which is the part a managed platform was doing at the start. For a client, that job is ours.

What we would tell anyone building a voice agent

  • Start managed, then look at the bill. A managed platform is the quickest way to find out whether callers accept an agent. Once they do, the per-minute fee is the number to watch.
  • Check what hardware you already own. If it can run the models, self-hosting can take the AI cost per minute close to zero.
  • The voice is not the product. Callers forgive a slightly robotic voice. They do not forgive an agent that cannot book, cannot find their order, or will not pass them to a person.
  • Handoff rules need designing. Every agent needs a clear rule for when it stops and transfers the caller, and the person who picks up needs to know what was already said.
  • Test with scripted calls. Bookings, cancellations, interruptions, background noise and after-hours calls, before a real customer hears it. Our guide to setting up and testing an AI receptionist has the list we use.

What this means if you want a voice agent

You do not need to repeat our three stages. For most businesses the right first step is one agent on one kind of call, such as bookings or missed calls, built from a template and connected to the system you already book into. Self-hosting makes sense when call volume makes per-minute fees a real cost, or when you already have hardware that can run the models. We will tell you plainly which side of that line you are on, and what each option would cost you per minute.

The same team that built Voice Ninja builds voice agents for clients, including the self-hosted phone agents described in our home care agency case study. You can try the product at Voice Ninja, or book a free 30-minute review and we will tell you what we would build first.

Frequently asked questions

How much can self-hosting a voice agent save?

It depends on call volume and hardware. In our case, moving Voice Ninja onto models we deploy ourselves cut its pricing by almost 60%. For clients who already own capable hardware, we run the models on their systems, and the AI cost per minute falls close to zero because they are no longer paying a provider for each minute. Phone numbers and any premium voices are still billed. A home care agency we work with can run its phone agents the same way: hardware paid for once, with phone numbers and any premium voices the remaining costs.

Can a voice agent run on our own servers?

Yes, if the hardware is good enough to run the speech and language models. We deploy the models on the client's own systems, so with the standard voices a call is transcribed, understood and answered on equipment they own.

Should I use a managed voice platform or self-host?

Start managed if you are new to voice agents: it is the quickest way to find out whether callers accept one. Self-hosting pays once call volume makes per-minute fees a real cost, or when you already own hardware that can run the models. We went through that sequence ourselves.

What is Voice Ninja?

Voice Ninja is Wope's own platform for building AI voice agents: phone agents that answer and make calls, and voice widgets for websites. You choose the language model, voice and language, connect the agent to your own systems with tool calls, and pay with credits for what you use.

Can you build a custom voice agent for my business?

Yes. We build voice agents on Voice Ninja or on a stack that suits your case, connect them to your booking, CRM or scheduling systems, and test them on scripted calls before they take real ones.

Can an AI voice agent book appointments and look things up?

Yes. Through tool calls the agent can book, look up and update records in your own systems during the conversation, and transfer the caller to a person when it should.

How is Voice Ninja priced?

On credits. You add a balance and credits are used only while your agents run, with no monthly commitment. Current prices are on voiceninja.ai.