Live two-way call
You speak and the companion responds in real time. Check latency, interruption handling and whether the call continues the same conversation.
AI GIRLFRIEND VOICE · EVIDENCE-FIRST
Voice can mean four different things.
Check which one you are paying for.
Some services offer live two-way calls. Others offer generated voice notes, text-to-speech playback or a mix. LarkMetric separates the mode, metering, continuity and privacy evidence before any commercial route is considered.
Evidence checked: 18 Sep 2026 · primary sources · hands-on voice quality not claimed
FIRST QUESTION
What does “voice” mean here?
Do not compare unlike features under one label.
Voice notes
Generated audio messages arrive inside chat. This can be useful, but it is not a live call.
Text-to-speech
Existing text is read aloud. It may sound natural while offering no spoken turn-taking.
Hybrid voice
A service may offer both calls and messages with different limits or credit costs.
CURRENT DOCUMENTED SNAPSHOT
Two services, two different voice economics.
These are documentation-backed facts, not a subjective “best voice” ranking.
Candy.ai
VERIFIEDLive call is explicit — and metered.
- Mode: AI Live Phone Call.
- Access: Premium feature.
- Current call meter: 3 tokens/min, shown as beta pricing.
- Subscription: paid plans include 100 tokens/month; plan prices vary by billing term and locale.
- Privacy gap: exact live-call audio-retention behavior is not yet verified from the current public policy.
Secrets.ai
VERIFIEDLive calls are tied directly to chat continuity.
- Mode: real-time audio calls are documented.
- Audio handling: raw voice is transcribed in real time and immediately deleted; text transcription is retained.
- Continuity: retained transcription supports later companion memory.
- Billing: audio calls are currently documented at 100 Moments/min; the Store also exposes Premium subscription pricing.
- Control: account deletion and memory controls are documented separately from subscription cancellation.
DECISION RULE
Choose the voice system, not the marketing word.
Confirm live two-way audio is present and identify its usage meter before paying.
Check whether the call becomes chat context or persistent memory after the session ends.
Look for a current statement about raw audio, transcripts, deletion and retention. UNKNOWN stays UNKNOWN.
Model subscription plus credits/tokens/minutes. A low monthly headline can hide expensive voice usage.
EVIDENCE MATRIX
What is actually documented today.
Green means the requirement is supported by current primary-source evidence. Partial means the feature exists but the relevant detail is incomplete.
| Criterion | Candy.ai | Secrets.ai |
|---|---|---|
| Real-time live call | VerifiedAI Live Phone Call; beta | VerifiedReal-time audio calls documented |
| Current usage meter | Verified3 tokens/min shown in call UI | Verified100 Moments/minute documented in the current Audio Calls guide |
| Raw audio retention | UnknownNo current precise call-retention statement verified | VerifiedRaw audio immediately deleted; transcript retained |
| Persistent continuity | PartialMemory claims documented; call-specific mechanics incomplete | VerifiedTranscript + memory continuity documented |
| Billing clarity | VerifiedMultiple subscription terms + 100 tokens/month | VerifiedCurrent Store exposes subscription pricing |
NEXT
Compare beyond voice.
See the adult AI companion comparison, the memory-focused comparison, or compare privacy, deletion and retention evidence.
Commercial route is separate from fit.
Affiliate disclosure: LarkMetric may earn an affiliate commission from an eligible referral. Route availability, GEO and payout do not change the evidence matrix above.
Checking whether the current supported referral route is available for your location…
SOURCE DISCIPLINE
Primary sources first. UNKNOWN when necessary.
LarkMetric did not perform hands-on latency or naturalness testing for this page. Current facts were checked against provider-owned product, billing and privacy surfaces. Provider claims are labeled as provider evidence, not independent performance proof.
For the broader voice + memory decision checklist, see AI companion voice and memory.