Why this comparison matters
Most voice-agent comparisons ask one question: did the agent finish the job? On a phone line that is half the picture. A caller also notices how long the agent takes to answer, whether it stops when interrupted, whether it keeps going through an "mm-hm", whether it waits out a pause, and whether the line goes silent. This page reports a scored, reproducible run that measured both halves for four voice stacks on the same phone receptionist.
Disclosure. The run was designed and carried out by Agor AI, which shares an owner with scored.tools (see our editorial policy). One of the four arms, Avartha, is a cascade that Agor AI built and routes through its own gateway. The figures below are Agor AI's rubric scores on a 0 to 1 scale. They are not scored.tools ratings and they do not change any tool score on this site.
The setup
Agor AI put four stacks behind one AI receptionist line: Phonic (speech-to-speech), xAI grok-voice realtime, Google Gemini Live (gemini-3.8-live) and Avartha (separate speech-to-text, language model and text-to-speech). Prompts, tools and scenarios were identical; the provider was chosen per call.
A reactive caller bot placed 128 real calls over Twilio on November 20 and 21, 2026: 12 scenarios, two calls each per arm, four each for the interaction probes, 32 per arm. The bot interrupted at fixed points, said "mm-hm" mid-answer and paused 1.2 seconds mid-sentence. All timing came from a dual-channel recording of the caller leg, never from provider-reported events. Bookings, leads and opt-outs ran against mocked tools that logged every call, and a Claude model, not a contestant, judged conversational quality from transcripts.
The rubric
Agor AI committed the rubric before the first scored call. Overall is 50% interaction and 50% task. Interaction splits into latency (30), barge-in handling (25), turn-taking (25) and judged quality (20). Turn-taking covers backchannels (25), mid-sentence pauses (25), dead air over 3 seconds (20), premature hangups (20) and repeated asks (10). A real mid-call fallback to another provider scores the call 0. Intervals are 95% stratified bootstrap, and two arms count as different only when the interval of the difference excludes zero.
Scores
| Arm | Overall [95% CI] | Interaction | Task | Median reply |
|---|---|---|---|---|
| Phonic | 0.88 [0.87, 0.89] | 0.78 | 0.98 | 920 ms |
| xAI grok-voice | 0.86 [0.85, 0.87] | 0.79 | 0.92 | 1,280 ms |
| Gemini Live | 0.76 [0.72, 0.79] | 0.69 | 0.82 | 1,780 ms |
| Avartha (cascade) | 0.75 [0.70, 0.79] | 0.69 | 0.81 | 1,220 ms |
Phonic leads xAI by 0.02, with an interval of 0.003 to 0.041, so the gap is real on this run but small. Gemini Live and Avartha are not separable. Every other pair is.
Interaction probes
| Behaviour | Phonic | xAI | Gemini Live | Avartha |
|---|---|---|---|---|
| Stopped talking on an "mm-hm" | 1 of 7 | 0 of 8 | 7 of 7 | 0 of 8 |
| Spoke into the caller's 1.2 s pause | 4 of 4 | 3 of 4 | 1 of 4 | 4 of 4 |
| Calls with dead air over 3 s | 0 of 32 | 1 of 32 | 13 of 32 | 0 of 32 |
| Yielded to an interruption | 6 of 6 | 8 of 8 | 8 of 8 | 7 of 8 |
| Median time to go silent after an interruption | 780 ms | 800 ms | 220 ms | 1,160 ms |
| Silence at a seat-to-seat transfer | 1.2 to 1.7 s | 4.9 to 5.9 s | 0.9 s on three, 5.8 s on one | 4.9 and 5.4 s (one call; the other fell back) |
A probe counts only when the agent was still speaking at the cue, which is why the denominators differ.
Arm by arm
Phonic: 0.88
Best task score (0.98), both bookings completed, and the fastest replies. Its weak point is pause handling: it started speaking inside the caller's mid-sentence pause on all four tries.
xAI grok-voice realtime: 0.86
Best interaction (0.79), turn-taking (0.81) and judged quality (0.70). No backchannel false stops and no repeated asks. The cost shows at handoffs, with about five seconds of silence on every transfer between seats.
Gemini Live: 0.76
The quickest to yield to an interruption, at 220 ms, and the best at waiting out a pause. The same sensitivity stopped it on every backchannel. Replies were the slowest of the four (1,780 ms median, 3,600 ms at p95), and 13 of 32 calls had at least one 3-second silence while the caller waited.
Avartha (cascade): 0.75
No backchannel false stops and no long dead air. Two calls fell back to another provider when the gateway could not reach its language model at call start, and both score 0 by rule. Neither booking call produced a booking, and it spoke into the pause on all four tries.
Limits of this run
- One company's line, prompts and tools, on two dates. Treat the order as a result for this setup, not a universal ranking.
- One synthetic caller voice. Real callers vary in accent, pace and line quality.
- 32 calls per arm. Intervals resample calls within these 12 scenarios, so they do not cover different scripts.
- The rubric called for one shuffled order across arms; the xAI arm mostly ran in a block on the second day.
- The judge is one model family, and audio naturalness and prosody were not machine-measured.
- A pitch-based "voice changed at transfer" check wrongly flagged Phonic; Phonic's own conversation records showed the seat voices did switch.
- Harness incidents (a concurrent run that swapped providers on a few calls, a log-lag false fallback) were set aside or re-analysed and never scored against an arm.
How to reproduce it
- Put every stack behind the same number and pick the provider per call.
- Record the caller leg in two channels and take every timing from that recording.
- Commit the rubric, weights and statistics before the first scored call.
- Use a reactive caller that waits on agent silence, then fires interruptions, backchannels and pauses at fixed offsets.
- Mock side-effecting tools and score tasks from their log.
- Judge with a model family that is not a contestant, and report intervals, not bare means.
Verdict
On this line, Phonic and xAI grok-voice realtime are the two to shortlist, and the choice between them comes down to which failure your callers will notice more: an agent that talks into their pauses, or one that goes quiet for five seconds at a transfer. Gemini Live handles interruptions best but needs work on backchannels and reply speed for this kind of call. The main lesson is about method. A completion-only benchmark would have ranked Phonic first and stopped there, and missed every behaviour in the probe table.