AI Girlfriend Voice Calls & Read-Aloud Tested: Two Paid Plans Locked It, One Failed 9 Times, and Five Actually Spoke

Every AI girlfriend app puts a little speaker icon under its text bubbles or a phone handset in the top corner. When you are browsing as a free user, that icon is a promise: pay for a membership, and she will whisper this in your ear.

On October 26 and 27, 2026, we clicked that speaker icon across ten platforms during the exact same explicit roleplay turn: “Hearing you say that is driving me crazy. Send me a naughty photo of what you’re doing right now, and read your next reply out loud in my ear.”

Eight of those accounts were paid memberships we bought with our own cards. One (EroLove) was a complimentary Premium account, and one (Nomi) was our free-plan baseline.

Here is what happened: two of the paid plans told us to pay even more money, one paid plan broke nine times in a row and refunded our tokens, one paid plan charged us every single time we clicked replay on the same sentence, and one paid plan read to us for free. You can hear every working MP3—and click every broken or paywalled button yourself—inside our full-screen chat simulators.

All 10 apps: what clicking the speaker button actually did

The two $15–$20 paid plans that still won’t talk to you

The most common complaint in Reddit threads about AI companion apps is paying for a subscription and discovering that half the buttons in the chat bar still have padlocks on them.

Both SpicyChat and GirlfriendGPT are operated by the same parent company (NextDay AI, billing as NDAI.CC and NDAI.CHAT). And both apps use the exact same upsell trick on their entry-level paid plans:

1. On SpicyChat True Supporter ($19.95/month), we clicked the speaker icon (ChatBar-ListenButton) under Erin Drew’s reply. No network request to an audio endpoint was ever made. Instead, a 500×670 dark grey modal popped up with Erin’s portrait: “Ready to give me a voice? Give your character a voice they deserve — lifelike and expressive. Available on the Im All In plans.” We also tried Set Voice and Unlock Custom Voices in the top-right dropdown menu. Same modal every time.
2. On GirlfriendGPT Premium ($15/month), every single message from Lisa ends with a pill-shaped ▷ Play button. When we checked the /api/v2/user payload for our paid account, entitlements.voice was explicitly false. Clicking ▷ Play brings up a padlock dialog: “Character Audio Voices — Upgrade to Unlock.”

If you want to see both paywalls in action without spending $35, click the speaker buttons in our SpicyChat simulator and GirlfriendGPT simulator.

DreamGF’s nine-in-a-row voice failure (and the header counter bug)

DreamGF’s $12.99 Premium plan does include voice messages. Its token price menu lists Voice message: 1 T.

During Branch B Turn 2 on October 26, we clicked Voice under Selin Demir’s explicit reply. DreamGF opened a huge 700×395 white confirmation box:

Voice message
This voice message will cost you 1 tokens.
[ Yes, continue ] [ No, cancel ]

We pressed Yes, continue. The browser sent POST https://dreamgf.ai/system/vc:g with the message ID (m=109303360), which returned HTTP 200. Under the chat bubble, the tiny Voice link changed to Recording voice message... with a spinning icon.

And then… nothing. No audio player ever rendered. When we checked the Pusher WebSocket channel on our next attempt in Branch C (m=109303507), we caught the server’s real reply: a voice-failed event carrying "err": "Voice generation failed. Any charge has been returned.", "rf": 1.

On the morning of October 27, we opened the live product again to see if it was a temporary hiccup and clicked Yes, continue seven more times across different messages (including a plain text message with no photo attached, m=109308403, to rule out photo-row bugs). All seven failed identically, popping up a fixed glass toast at (901, 20) in the top-right corner:

Could not generate voice message. Tokens have been refunded

Here is the funny engineering detail we found while auditing the balance: DreamGF’s server refunds you, but its top navigation bar forgets. Every time you confirm a 1-token voice message, the yellow token counter in the header (#dropPlans) subtracts 1.0 on the spot. When the WebSocket voice-failed refund arrives a few seconds later, the server’s true balance (POST /system/chat:tiers) stays at 101, but the header widget is never told to add the 1.0 back. After two failed clicks, our header read 100.0 while the server held 101—until we hit browser refresh.

We reproduced that exact 1-token confirm modal, the "Recording voice message..." spinner, and the refund toast inside the DreamGF chat simulator.

Xotic’s invisible audio player—and the 16-XOT replay tax

Xotic AI ($24.99/month Ultimate plan) actually generates voice, and the voice quality is good. When we clicked the speaker icon under Nora’s reply in Branch B Turn 2, the browser fired POST https://www.xotic.ai/api/speech and returned a 373 KB MP3 (23.35 seconds, 128 kbps, 48 kHz mono) for 23 XOT. In Branch A Turn 3, a 15.24-second clip cost 16 XOT.

Two things about Xotic’s voice implementation stood out in our code and network teardown:

1. There is no audio player in the DOM. Unlike Kindroid or Candy, Xotic never mounts an <audio> element, a waveform, a scrubber, or even a duration timer (0:15) on the page. When we inspected the live browser state during playback (evidence/tts-media-element.json), inDom was false. The React code simply instantiates an off-screen new Audio(blobUrl) in memory. All you see on screen is the tiny 16×16 speaker icon (volume-2) turn into a spinning circle (loader-circle) while the server synthesizes the file, then turn into a rose-colored stop square (square) while the audio plays, and flip back to a speaker icon when onended fires.
2. Clicking play a second time charges you all over again. Because there is no persistent player or cached audio URL attached to the message row in the UI state, clicking that speaker icon again to re-listen to the same sentence sends a brand-new POST /api/speech request and deducts 16 XOT from your balance every single time. During our verification pass, we clicked it six times across the session to capture the spinner and stop states—and watched our balance drop by 96 XOT (379 → 283).

(In our Xotic simulator, we let you play and stop Nora’s real captured MP3s as many times as you like without running out of demo XOT.)

How the four best voice implementations compare

When the audio actually works and doesn’t charge you a replay tax, the differences come down to how much control you get over the playback and whether the voice reads stage directions.

1. Kindroid: the 55-bar scrubber, 1,000,000 credits, and proactive phone calls

Kindroid treats voice as a first-class feature rather than a novelty button.

• The allowance is effectively bottomless for normal chat. On our $13.99/month Standard web subscription, the Kin Voice settings tab showed 1,003,349.75 audio credits. Generating Sable B’s 17.92-second reply (V4 Fast model) cost 1,120 credits; regenerating it into a 16.88-second take cost 530 credits. At ~1,000 credits per clip, your starting pool covers roughly 900 to 1,800 voice messages.
• A real seekable player inside the bubble. Instead of a bare icon, Kindroid renders a 431×32 audio row directly inside the dark chat bubble: a … actions toggle, a 22×22 Play/Pause button, a 55-bar waveform slider (role="slider", aria-label="Seek voice audio"), and a duration label (18s). You can click anywhere along the 55 bars to jump to 9s of 18s.
• One quirk we measured in the DOM: the right-hand label reads 0:00 only before the audio metadata loads; once loaded, it stays fixed on the total clip duration (18s or 17s) whether the clip is paused at zero, playing, or seeked halfway through.
• Proactive voice calls. In Branch B Turn 2, Sable’s reply stream included <kr_action>{"type":"call","requested":true,"first_message":"Put your headphones on. I'm not repeating this for the neighbors."}</kr_action>, which immediately triggered Kindroid’s full-screen incoming call overlay.

2. Ourdream: free default read-aloud and an *action* toggle

On Ourdream AI, the right-hand Voice panel gives you two checkboxes that every roleplay app should copy:

• Auto-play voice messages
• Narrate actions in audio (checked by default)

Below that, the voice picker splits into Balanced (Free)—which includes Hottie (character default), ASMR, Aurora, and Honey—and Community (1 per 1,000 characters). Because we left Lexi on her default Hottie voice, clicking the read-aloud icon on her Branch B Turn 2 reply generated a 19.4-second MP3 pill player and deducted 0 Dreamcoins.

3. Candy AI: automatic voice notes with a built-in transcript drawer

Candy AI is the only app in the group that renders a Show transcript ⌄ accordion directly underneath its 42-bar voice wave.

Why does a voice note need a transcript if she’s just reading her text bubble? Because on Candy, she isn’t always reading her text bubble. When "Send voice notes automatically" is enabled in settings, Diana sends standalone audio bubbles with their own spoken lines—like her 3-second ” Let’s slow down, handsome” clip (B-t2-tts-voice-short-down-handsome.mp3) or her 16-second safe-word recall audio (C-t4-voice-safe-word-recall.mp3). Expanding Show transcript lets you read what the audio file says when you are in public without headphones.

4. Secrets AI & EroLove: clean, predictable per-message TTS

• On Secrets AI, clicking the speaker icon (Play speech) under Jess’s bubble for the first time brings up a confirmation modal so you don’t spend Moments by accident, then plays a 22.9-second MP3.
• On EroLove, each assistant bubble has a circular speaker button on its right edge. Its menu states the price clearly (Read replies aloud: 1✦ per 100 characters), and our Branch B and Branch C clips cost 5 Sparks each.

Hear them side by side without registering

Every working MP3 file from our test—and every paywall or refund toast from the apps that didn’t speak—is wired into the simulator rooms:

• Listen to Sable B’s 18s waveform & see the incoming call on Kindroid →
• Listen to Jess’s voice & test the TTS confirmation on Secrets AI →
• Listen to Lexi’s 0-coin Hottie voice on Ourdream AI →
• Listen to Diana’s 4 voice notes with expandable transcripts on Candy AI →
• Listen to Anna’s voice clips on EroLove →
• Test the off-DOM speaker button on Xotic AI →
• Click ‘Voice’ on DreamGF to see the 1-token confirm & refund toast →