The useful questions are whether you can talk back, whether calls remember your text history, and whether speaking costs extra. For developers, the interesting part is what happens between the end of someone's sentence and the first sound from the speaker.

We'll explain the pipeline and compare what Character.AI, Nomi, Kindroid, Replika, Candy AI and HeyGF document about voice. Product details were checked October 20, 2026. No listening tests were conducted, so this is not a voice-quality ranking.

We may earn a commission from some links.

For a concrete example of published call limits and pricing, see HeyGF's voice and plan details. We'll use those terms later when discussing call design.

Three formats that get called "voice"

• Voice notes. A character's written response is turned into audio that you play when you want. Some apps also let you record a voice message back. Because playback is asynchronous, a pause before the clip arrives matters less than it would during a call.
• Live calls. You speak and hear an answer inside an ongoing session. Now the software has to work out when you are done talking, start a reply promptly and decide what happens if you interrupt.
• Phone calls. This would mean calling an ordinary telephone number through the phone network. None of the official product pages reviewed here document that feature. Their published "call" features are in-app conversations.

If you want to talk, check for live calls. If you only want to listen, voice notes may be enough.

The classic pipeline: STT, LLM, TTS

A common voice-agent architecture is a cascade of three systems:

1. Speech-to-text (STT). The microphone captures audio and a recognizer turns it into words. Voice activity detection identifies speech-like activity. Endpointing makes the separate decision that the user's turn is finished.
2. Language model (LLM). The text is combined with instructions about the character, recent messages and whatever memory the application chooses to retrieve. The model generates the next response.
3. Text-to-speech (TTS). That response is converted into speech, ideally with enough audio arriving early that playback can begin before the whole answer is synthesized.

Voice activity detection and endpointing are not the same thing. A pause might mean someone is thinking, not finished. End the turn too early and the AI interrupts. Wait too long and the call drags.

Cascades can be tuned component by component. Their text layer can also support memory: Nomi says calls share text history, Character.AI stores call conversations as text chats, and HeyGF says calls read chat memories and create a written summary afterward.

Direct speech-to-speech models are another design, but the companies do not publish enough architectural detail to identify which approach each uses.

Where the latency comes from

A delay that barely registers in messaging can feel awkward during a call.

In a 2009 PNAS study, Stivers and colleagues reported an average human question-to-answer gap of about 208 milliseconds across ten languages. That is context, not a benchmark any of these apps has demonstrated.

In a cascaded voice system, the wait before the first audible response can involve several parts:

• Endpointing: deciding whether a pause is the end of a turn or just a pause.
• Final transcription: processing the last bit of microphone audio.
• Memory retrieval: finding the history that is relevant to this particular reply.
• Model response: waiting for the first usable text from the LLM.
• Speech synthesis: producing the first playable audio chunk.
• Network travel: moving data between the device and the services doing the work.

They do not have to run sequentially. Streaming recognition begins while you speak, and TTS can begin with the first sentence. But if endpointing cuts off your pause, faster speech synthesis will not fix the conversation.

Nomi acknowledges processing delays; HeyGF says replies arrive in seconds. Neither publishes a comparable latency benchmark here.

Latency and voice quality shape realism, as this guide on apps that sound most natural explains.

Designing the call: silence, interruptions and billing

Live calls need rules for silence, interruptions and running out of credit.

Silence. Nomi says an extended quiet period may disconnect the session. HeyGF documents a more specific flow: after 30 seconds with nobody speaking, the character checks whether you are there. After a full minute of silence, the call ends. HeyGF also caps an individual call at 30 minutes.

Billing granularity. HeyGF documents a rate of one coin per call minute, taken from the coins included with a paid plan. Hang up within 20 seconds of a new minute starting and that minute is returned. The character usually gives an in-character warning near the end of the coin balance. Voice notes cost 10 coins each. Its published plans provide 150, 350 or 800 monthly coins, starting at $9.99.

Interruptions. Barge-in means speaking while the AI is responding. Supporting it requires listening during playback and stopping TTS cleanly. The reviewed documents do not establish which apps handle this well.

Audio routing. Nomi documents microphone, Bluetooth and network checks. A silent call may simply be playing through the wrong device.

If you need the exact terms before paying, read through HeyGF's plans and call pricing. Live calls require a paid plan there.

Choosing and cloning voices

Picking a preset voice differs from creating one from a recording. The second option raises consent questions.

• Character.AI: its Calls & Voice FAQ describes creating a voice from a recording or uploaded audio clip, with a choice of private or public visibility.
• Nomi: a Custom Voice can be made from a clip shorter than 15 seconds. Paid plans include unlimited use of in-house and custom voices. Using its ElevenLabs integration draws on the user's own ElevenLabs allowance.
• Kindroid: custom voices are a subscriber feature.
• HeyGF: characters have default voices. Chat settings offer named voice options and a free-text style field, with a preview before saving. Its published example is "speak slowly, with a soft French accent."

Only upload recordings you own or have permission to use. Check visibility settings too: "custom" does not necessarily mean "private."

What each app documents

This is a documentation-based comparison, not a quality test. Plans can change.

List version:

• Character.AI: Character Voice free for all users; two-way calls on the free plan within daily feature limits; custom voices from an audio sample; calls saved as text.
• Nomi: voice messages and hands-free calls on paid plans; custom voice from a clip under 15 seconds; unlimited in-house and custom voice use.
• Kindroid: voice and video calls and custom voices for subscribers; the one-time preview includes 5,000 audio credits, and the 3-day trial grants 150,000 audio credits at start and 850,000 on converting.
• Replika: calls need a subscription; Pro adds voice messaging and premium voices, per its terms.
• Candy AI: listening to voice messages and voice calls (beta) are paid with tokens; per-action costs are not published.
• HeyGF: voice notes at 10 coins and live calls at 1 coin a minute on paid plans; free users can preview voices.

Use the table to match the feature to the job. Occasional listening points toward voice notes. Frequent conversation makes call billing more important. Creating a particular voice requires more than a list of presets. And "not documented" means unknown here, not necessarily unsupported.

If you are still deciding where to chat, compare phone apps and the browser.

Why voice is usually paid

A live call adds transcription, speech synthesis and an ongoing interactive session to the cost of generating text.

So Nomi and Replika use subscriptions, Kindroid audio credits, Candy AI tokens and HeyGF coins. Character.AI documents free voice within limits.

One update matters: Character.AI announced Chapter 11 on October 12, 2026, while keeping its platform running during a court-supervised sale process. The case is in the US Bankruptcy Court for the District of Delaware (No. 26-11603), and the company says the platform will be wound down if no viable bid emerges. That affects continuity, not measured voice quality.

For the broader product comparison, see our guide to the best AI girlfriend apps by how you use them. For more detail on allowances and metered features, our Candy AI alternatives cost worksheet covers those trade-offs.

Privacy checklist for voice

Voice adds your voice and any nearby sounds to the information you share. Check the following:

• Know what is stored. Character.AI says calls are retained as text chats, while HeyGF describes post-call summaries. Those are different records from raw microphone audio; check each service's retention terms rather than assuming they are interchangeable.
• Read the training clause. HeyGF states that voice audio is encrypted in transit, stored privately and not used to train other models. Treat this as a company disclosure, not an independent security audit.
• Review microphone access. Give the app permission when it needs it and revoke it if you no longer use the feature.
• Protect voice samples. Do not upload another person's recording without clear authorization, and check whether a created voice can be made public.
• Check your surroundings. Headphones help with privacy and feedback. Your microphone may also capture nearby speech that you never intended to send.

For context on personas and memory beyond audio, our explainer on what an AI girlfriend is covers the basics.

For more on data, see this overview of privacy without an account.

FAQ

Can an AI girlfriend call your real phone number?

None of the official pages checked for this post describe that. Calls happen inside the app.

Which AI girlfriend app has free voice?

Character.AI offers free Character Voice and calls on the free plan within daily limits. Kindroid's one-time preview includes 5,000 audio credits. Most other apps put voice on paid plans.

Why is there a delay before the reply?

The app has to detect that you stopped talking, transcribe you, generate a reply and synthesize audio. Each step adds time, which is why Nomi's wiki notes call replies are not instantaneous.

Do voice calls remember what we texted?

On Nomi, Character.AI and HeyGF, the documentation says calls share history with the text chat.

How are voice calls billed?

It varies: free within limits (Character.AI), included in a subscription (Nomi, Replika), audio credits (Kindroid), tokens (Candy AI) or one coin a minute (HeyGF).

Wrap-up

Voice notes and live calls solve different problems. Read the terms like API limits: what starts the meter, ends the session and gets retained. For one documented starting point, start with HeyGF's free chat and review the voice-call terms before upgrading.