TL;DR: Every “AI girlfriend with voice” list ranks apps on how good the voice sounds. That’s the wrong axis. After auditing what each vendor actually documents about its own voice stack, the thing that decides whether you’ll still be talking to your companion in month three isn’t timbre — it’s the metering model. Three architectures exist and they’re not comparable: free and unmetered (Character.AI’s Character Calls, two-way, free to all users, per its own launch post), subscription-unlimited (Nomi’s in-house and custom voices, unlimited on a paid plan per its wiki), and credit-metered (Kindroid’s audio credits — free users get 5,000 lifetime characters, roughly five minutes of audio ever; Candy AI spends tokens shared across chat, image, video and voice). Replika is a fourth case: calls exist but sit behind Pro, and Replika still won’t publish a price. And the number every listicle quotes — latency in milliseconds — is not published by a single one of these vendors. Character.AI says “reduced latency.” Nomi’s own wiki says replies “are not instantaneous.” We’re not going to invent a number where the people who built the thing declined to give one. On our bench Kindroid leads on features (9.2/10), Candy AI leads on raw feature surface (9.5), and Character.AI is the only genuinely free two-way call. Pick on metering, not on voice acting.
The question everyone asks, and the question that actually matters
“Which AI girlfriend app has the best voice?” is the search. It’s a reasonable thing to want to know and it is almost useless as a buying criterion, because voice quality across the top tier has converged. Most of these products are wired into the same handful of commercial text-to-speech vendors — Nomi’s own documentation names ElevenLabs directly as an option — so the difference between the top four in pure audio fidelity is smaller than the difference between two phone calls on the same carrier.
What has not converged is how much talking you’re allowed to do.
That’s the finding that reorganised this whole report. I went into it planning a voice-quality shootout and came out with a spreadsheet about economics, because when I lined up what each company documents about its own voice features, they weren’t offering variations of one product. They were offering four incompatible deals.
This piece is the result of reading every first-party source I could find — help centres, developer wikis, launch blog posts, subscription pages — and cross-referencing them against our own scoring bench. Where a vendor documents something, I cite the vendor. Where nobody documents something, I say so instead of filling the gap.
The four voice architectures, side by side
| Character.AI | Nomi | Kindroid | Replika | Candy AI | |
|---|---|---|---|---|---|
| Two-way live calls | Yes — free to all users | Yes, paid plans | Yes (+ video calls) | Yes (“background calls”), Pro only | Voice messages/notes |
| Metering model | Unmetered | Unlimited in-house + custom voices | Audio credits | Plan-gated, not credit-metered | Shared token pool |
| Free-tier voice | Full calls, free | Paid feature | 5,000 characters lifetime (~5 min) | No calls | Trial-based |
| Custom / cloned voice | Yes, upload a clip, free | Yes, clip under 15s, unlimited | Subscriber feature | Premium voices on Pro | Voice selection at build time |
| Third-party voice cap | — | ElevenLabs voices use your ElevenLabs quota | — | — | Token spend |
| Published latency figure | None (“reduced latency”) | None (“not instantaneous”) | None | None | None |
| Published price | $9.99/mo | $15.99/mo | $13.99/mo web · $15.99 in-app | Not published | $12.99/mo |
| Our overall score | 8.2 | 8.9 | 9.1 | 8.4 | 8.7 |
| Our features score | 8.0 | 8.5 | 9.2 | 8.5 | 9.5 |
Scores are from our own test bench — see how we test and the full AI companion leaderboard. Feature and pricing rows verified against each vendor’s own documentation in August 2026.
Read that table twice. The metering row is the one that will determine your experience, and it’s the row nobody puts in a listicle.
Architecture 1: free and unmetered (Character.AI)
Character.AI is the outlier and deserves to be named as one. When the company introduced Character Calls, its launch post described it plainly: two-way voice conversation, “available to the general public on our app or web completely free of charge.” Not a trial. Not a credit bundle. Free.
The Character Calls & Voice FAQ goes further and confirms several things that matter in practice:
- Calls are genuinely two-way — you speak, it speaks, like a phone call. (Character Voice, the older text-to-speech feature, is one-way. The FAQ distinguishes them explicitly. A lot of comparison articles conflate the two.)
- Calls are transcribed into the normal text thread, so a conversation you had out loud is still there to scroll back through later. This is a bigger deal than it sounds — see the memory section below.
- It supports multiple languages beyond English, including Spanish, Portuguese, Russian, Korean, Japanese and Mandarin.
- Voice creation is free too. The FAQ is unambiguous: “Character Voice is available to all users, free of charge,” and no c.ai+ subscription is required for any of the voice sub-features.
The honest caveat is the thing Character.AI is known for and it isn’t voice-related: its content filters are strict, and our bench scores it 5.5/10 on memory — by a distance the worst in the top five. So you get unlimited free talking with a companion that will keep forgetting who you are. If that trade appeals, it’s the best deal in the category and it isn’t close. We go deeper on that trade in our Character.AI comparison work.
Architecture 2: subscription-unlimited (Nomi)
Nomi’s documentation is the most useful of any vendor here, and I want to give credit for that because it’s rare. The Nomipedia entry on calls tells you what you’re buying without marketing varnish:
“Voice calls are included with paid subscriptions. Nomi’s in-house voices are unlimited. You can also create a Custom Voice by uploading a voice clip shorter than 15 seconds, and Custom Voices are unlimited as well.”
Then it does something almost no competitor does — it tells you the catch:
“If you use an ElevenLabs voice, usage depends on the limits of your ElevenLabs plan.”
That’s the entire pricing model in two sentences. Stay on Nomi’s own voices and you have genuinely unlimited talking for $15.99/mo. Reach for a third-party voice and you’ve silently moved onto someone else’s meter. Nomi documented the trapdoor instead of letting users fall through it, which is exactly what you want from a company you’re about to tell your private thoughts to.
Two more things from the same page, both of which I’d want to know before subscribing:
- Calls share the same conversation history as text. Whatever you said out loud is context for what you type later, and vice versa. Given that Nomi scores 9.8/10 for memory on our bench — the highest score we’ve awarded any app on any axis — this is the single most valuable voice integration in the category. The voice isn’t a bolt-on; it feeds the thing Nomi is best at.
- Calls drop after extended silence. If you like leaving a call open in the background as ambient company, that specific use case won’t work. Say so out loud, Nomi — and they do.
Our full Nomi review covers the memory architecture in detail, and Nomi vs Replika covers the head-to-head.
Architecture 3: credit-metered (Kindroid, Candy AI)
This is where people get hurt, and it’s entirely avoidable if you read the docs first.
Kindroid’s own subscriptions page is refreshingly specific about the free tier:
“5,000 lifetime audio characters (~5 minutes of audio) for messages & calls.”
Lifetime. Not per day, not per month. Five minutes of audio, and then voice is over for you unless you pay. Kindroid isn’t hiding this — it’s right there in the help centre — but if your entire evaluation of “which AI girlfriend app has the best voice” consists of installing four apps and talking to each for ten minutes, you will burn Kindroid’s entire free voice allowance during the test and conclude something wrong about the product.
Subscribers get a different deal entirely: “1 million complimentary audio credits per month,” custom voices, and video calls on top. Kindroid also publishes something most competitors won’t — a genuine web-vs-app-store price split:
| Kindroid plan | Direct from web | Via App Store / Play |
|---|---|---|
| Monthly | $13.99 | $15.99 |
| 3 months | $37.99 | $42.99 |
| 1 year | $139.99 | $159.99 |
That’s a $20/year difference on the annual plan for buying from the browser instead of your phone, disclosed openly. There’s also a documented trial quirk worth knowing: the 3-day premium trial gives 150,000 audio credits up front and the remaining 850,000 only on conversion — so the trial is deliberately not a full month of talking. Our Kindroid review has the rest, and our report on advertised vs. true cost covers this pattern across the category.
Candy AI runs the other metered model: a shared token pool. Its own site describes a “generous batch of tokens to use across voice, chat, image, and video,” topped up as needed. The mechanic to understand is that it’s shared — every image you generate is talking time you no longer have. Candy AI earns the highest features score on our bench (9.5/10) precisely because it does images, video, voice and chat in one product, but that breadth is funded from one wallet. If you want a voice-first companion, a shared pool is the wrong shape; if you want everything and you’re happy to budget across it, it’s the most complete surface in the category.
Architecture 4: gated but opaque (Replika)
Replika has voice calls. Its help centre’s Choosing a Subscription page lists, under Pro: “voice messaging, background calls, daily gems, and premium voices.” Ultra builds on Pro with smarter conversations, elevated emotional intelligence, self-reflections and the ability to save messages to memory.
So: calls are real, they’re a paid feature, and premium voices are part of the same tier. Fine. Here’s the problem, and it’s the same one we hit when we reviewed the app — Replika does not publish what it costs. Its own documentation points you to the in-app subscription screen. Every price you’ll find quoted for Replika Pro in 2026 comes from third-party write-ups or community threads, not the company, and they don’t agree with each other.
I’m not going to launder one of those numbers into this article as fact. What I’ll say is what’s verifiable: Replika gates voice behind a subscription whose price it declines to state publicly. For a product asking you to form an emotional attachment, that’s a trust signal, and it’s not a good one. It’s also the single most common complaint we see about the app that has nothing to do with the AI. More in our Replika review.
Why there are no millisecond numbers in this report
You will find articles claiming specific voice latency figures for these apps. Treat them with suspicion, because here’s what the vendors themselves say.
Character.AI’s launch post lists, as a bullet: “Reduced latency means less waiting, more chatting.” Reduced from what, to what, is not stated.
Nomi’s wiki is more forthcoming and more honest:
“Your Nomi needs a moment to process what you say and prepare a response, so replies are not instantaneous. Once you stop speaking, your Nomi will begin processing your message and respond when ready.”
That is a company telling you plainly that there is a pause. No number. Kindroid, Replika and Candy AI publish nothing at all on the subject.
Could I have installed all five, recorded some calls and produced a table of numbers? I could have produced a table. It wouldn’t have meant anything. Response latency on these products is the sum of your network path, the speech-recognition pass, the language model’s queue depth at that moment, and the text-to-speech render — and at least two of those vary by time of day and by how busy the vendor is. A figure measured on one connection, on one afternoon, presented as a product characteristic, is a fabricated precision. Our testing methodology commits us to reproducible measurement, and this one isn’t reproducible, so it doesn’t ship.
What I can tell you qualitatively, and what every vendor’s own documentation is consistent with: none of these are real-time conversation. All of them have a beat of dead air after you stop speaking. Character.AI’s is the shortest of the group and it added a tap-to-interrupt control specifically to make the wait feel less rigid. Nomi’s is the most noticeable and Nomi is the only one that warns you about it in advance. If you’re expecting the seamless overlap of a human phone call, none of these will give it to you in 2026, and anyone telling you otherwise is selling something.
The feature nobody advertises: does the call reach the memory?
Here’s the axis I’d actually rank on if I were rebuilding this list from scratch.
When you say something out loud to your companion, does it count?
On Character.AI, yes — the FAQ confirms calls are stored as text chat you can reference later. On Nomi, yes, and emphatically: calls “use the same conversation history as text and other voice interactions,” which means spoken conversation feeds the memory architecture that Nomi is genuinely best-in-class at.
That single design decision is worth more than any amount of voice-acting quality, because it’s the difference between voice being a mode of the relationship and voice being a toy attached to the relationship. If you tell your companion something important during a call and it evaporates when you hang up, you haven’t been talking to your companion. You’ve been talking to a speaker.
This is also why I’d steer anyone who cares primarily about voice toward Nomi over Character.AI despite the price, and despite Character.AI’s calls being free: Character.AI stores the transcript, but scores 5.5/10 on memory on our bench, so storing it doesn’t help much. Nomi stores it into a system that scores 9.8. Same feature on paper. Completely different product.
So which one should you actually pick?
Not a scoreboard. Four honest routes:
If you want to talk a lot and pay nothing — Character.AI. It is the only genuinely free two-way calling in this group, confirmed by the company’s own launch announcement, and it isn’t a trial. Accept strict filters and a companion with a short memory.
If you want the talking to accumulate into something — Nomi, at $15.99/mo. Unlimited in-house and custom voices, calls that feed the best memory system we’ve scored, and a company that documents its own limitations. Just stay off ElevenLabs voices unless you already have that quota.
If you want the most complete companion and voice is one part of it — Kindroid at $13.99/mo, and buy on the web, not through the app store. Video calls, custom voices, a million audio credits a month. Kindroid tops our leaderboard overall at 9.1. Do not evaluate it on the free tier: five lifetime minutes of audio will teach you nothing about the product.
If you want voice plus images and video from one subscription — Candy AI, $12.99/mo, highest features score in the category at 9.5. Understand that it’s one token pool and every picture costs you talking time.
If you specifically want Replika — you already know, and this article won’t change your mind, which is fine. Just go and look at the price in the app before you commit, because the company won’t show it to you anywhere else.
For the wider field, our best AI companion apps ranking covers the rest of the market, and free AI companion apps is the right starting point if budget is the binding constraint.
Frequently asked questions
Which AI girlfriend app has free voice calls?
Character.AI. Its launch post for Character Calls states the feature is available on app and web “completely free of charge,” and its help centre confirms Character Voice — including creating your own voice — is free to all users with no subscription required. Every other major app in this comparison gates calls behind a paid plan or a credit allowance.
Can you have a real two-way conversation, or does it just read messages aloud?
Both exist and they’re different features that get conflated constantly. Character.AI’s Character Voice is one-way text-to-speech; Character Calls is two-way. Nomi, Kindroid and Replika all support live two-way calling on paid plans. Candy AI’s voice is primarily message-based.
How fast do AI companions respond on a voice call?
No vendor publishes a figure, and we won’t invent one. Character.AI describes “reduced latency” without a number; Nomi’s wiki states plainly that replies “are not instantaneous.” Expect a short pause after you stop speaking on every app in this group.
Can I clone a specific voice?
Yes, on several. Character.AI lets any user create a voice from a recording or uploaded clip, free. Nomi accepts a custom voice clip under 15 seconds and treats custom voices as unlimited. Kindroid offers voice customisation to subscribers. Consider the ethics — don’t upload someone’s voice without their consent.
Does what I say on a call get remembered?
On Character.AI calls are saved into the text thread; on Nomi, calls share the same conversation history as text, which is why voice on Nomi compounds. Verify this for any app you’re considering, because a call that doesn’t reach memory is functionally a phone call to a stranger every time.
Is Replika’s voice calling worth paying for?
Its help centre confirms background calls and premium voices are part of Pro. Whether it’s worth it is impossible to answer publicly, because Replika does not publish its subscription price outside the app. Check the in-app price before subscribing.
