What does the speech-to-speech benchmark measure?

It measures complete realtime voice models on live agent calls, including reliability, task success, saved-data accuracy, how often the agent went quiet, response time, and cost.

How is speech-to-speech reliability ranked?

The leaderboard uses pass³, the share of scenarios where all three retained calls passed every required check. A scenario must work repeatedly to count.

Are the models tested under the same conditions?

Yes. Every model runs the same voice agent, prompt, tools, connection, caller scenarios, and scoring rules. Only the realtime model changes.

How does a native speech-to-speech model compare with a cascaded voice agent?

Native speech-to-speech models handle audio directly. A cascade connects speech-to-text, an LLM, and text-to-speech. Cekura runs both through the same agent scenarios so reliability and response-time differences remain comparable.

Can I compare two realtime voice models directly?

Yes. The comparison tool provides head-to-head reliability, accuracy, latency, failure, and cost results for any two tested models.

Bench

Book a demo

Speech-to-speech model benchmarks

LeaderboardAgent code

Reliability vs. response time

Higher and further left is better

Best quadrantBest picks

Reliability, pass³ (%)

Median response time (s)

For the wider field, see our pick of the best voice AI APIs for real-time audio.

Who wins what

The leader on each measure

Most reliablePhonic v190.2%pass³
Fastest replyPrism1.34smedian
CheapestHiggs Realtime$0.0061per min

Full leaderboard

AI voice models handled live audio calls over WebRTC. This is who got them right.

Last updated Nov 3, 2026Methodology by Dileep Chagam

Leaderboard

#ModelPass³

Success rate
95.1%
Data accuracy
97.5%
Went quiet
0.0%
Response time
1.63s
Cost / min
$0.0600
What we saw
One of the quickest to reply, and never left a caller waiting.
Passes the most Medicare calls of any realtime model, run after run.
Hangs up without acknowledging a caller who says they will call back later.
Now and then saves a Medicare handoff note or routing detail wrong.
Compare with Cascade See the calls
Everything on Phonic v1
More figures
Scenarios passed every run
74 of 82
Agent responses
97.2%
Interruption
4.88 of 5
Response time, p90
1.81s
Success, Booking calls
97.7%
Success, Medicare calls
88.4%
Setup
Model
phonic_v1
Setting
Intelligence high
Voice
sabrina
Turn taking
Service turn detection
Success rate
86.6%
Data accuracy
87.3%
Went quiet
0.0%
Response time
1.34s
Cost / min
$0.0800
What we saw
Among the quickest to reply, level with Rime Prism, and never left a caller waiting.
Passes more scenarios on every run than any other realtime model, and passed every noisy call, every call with audio glitches and every call where a tool failed.
On Medicare calls it often leaves out a detail it was given, most often the route ID in the handoff note or the ZIP code.
Well behind GPT Realtime 2.1 on Medicare calls, and one booking failed on every run.
Compare with Phonic v1 See the calls
Everything on Grok Voice 3.0
More figures
Scenarios passed every run
67 of 82
Agent responses
97.2%
Interruption
4.98 of 5
Response time, p90
1.62s
Success, Booking calls
97.7%
Success, Medicare calls
58.0%
Setup
Model
grok-voice-3.0
Setting
Reasoning high
Voice
eve
Turn taking
Service turn detection
Success rate
89.0%
Data accuracy
92.0%
Went quiet
3.3%
Response time
1.94s
Cost / min
$0.0853
What we saw
Almost never fails, and rarely twice in a row.
One of the strongest realtime models on Medicare calls.
Struggles with noisy audio.
Now and then saves a phone number wrong, or forgets something the caller said earlier.
Compare with Phonic v1 See the calls
Everything on GPT Realtime 2.1
More figures
Scenarios passed every run
65 of 82
Agent responses
99.2%
Interruption
4.98 of 5
Response time, p90
2.46s
Success, Booking calls
92.1%
Success, Medicare calls
81.2%
Setup
Model
gpt-realtime-2.1
Setting
Reasoning high
Voice
marin
Turn taking
Service turn detection
Success rate
83.7%
Data accuracy
89.0%
Went quiet
12.2%
Response time
2.02s
Cost / min
$0.0553
What we saw
Near perfect on bookings, run after run.
The conversation always reaches the right outcome.
On Medicare calls it sends the caller to the wrong team or skips a step.
Goes quiet the most, says goodbye without hanging up, and has the lowest voice score.
Compare with Phonic v1 See the calls
Everything on GPT-Live 1 + GPT-6 Sol
More figures
Scenarios passed every run
62 of 82
Agent responses
100.0%
Interruption
4.89 of 5
Response time, p90
2.42s
Success, Booking calls
97.2%
Success, Medicare calls
49.3%
Setup
Model
gpt-live-1
Setting
Sol on low thinking
Voice
marin
Turn taking
Service turn detection
Success rate
84.1%
Data accuracy
86.1%
Went quiet
4.1%
Response time
1.66s
Cost / min
$0.0798
What we saw
Among the best when a caller corrects themselves.
Passed every interrupted call and every call where a tool failed, and is among the quickest to reply.
On Medicare calls it often leaves out an ID it was given, most often in the handoff note.
Background noise and calls with several requests trip it up more than most, and it now and then goes quiet on a booking call.
Compare with Phonic v1 See the calls
Everything on Azure Realtime
More figures
Scenarios passed every run
58 of 82
Agent responses
98.8%
Interruption
4.96 of 5
Response time, p90
2.22s
Success, Booking calls
91.0%
Success, Medicare calls
66.7%
Setup
Model
azure-realtime
Setting
No reasoning setting
Voice
ava
Turn taking
Service turn detection
Success rate
81.7%
Data accuracy
84.5%
Went quiet
0.0%
Response time
1.34s
Cost / min
n/a
What we saw
Quickest median reply of the field, level with Grok Voice 3.0, and never left a caller waiting.
Passes more Medicare calls than most realtime models.
Medicare handoff notes often leave out the route ID, and some calls skip the consent, qualification or routing step.
Behind most models on bookings, where it now and then looks the caller up by the wrong phone number.
Compare with Phonic v1 See the calls
Everything on Prism
More figures
Scenarios passed every run
56 of 82
Agent responses
91.1%
Interruption
4.91 of 5
Response time, p90
1.57s
Success, Booking calls
88.7%
Success, Medicare calls
63.8%
Setup
Model
prism
Setting
No reasoning setting
Voice
astra
Turn taking
Service turn detection
Success rate
76.0%
Data accuracy
78.9%
Went quiet
4.5%
Response time
1.62s
Cost / min
$0.0800
What we saw
Shortest replies of the field, and steady on bookings.
On Medicare calls it drops or garbles details from earlier in the call.
Every time it went quiet was on a Medicare call.
Compare with Phonic v1 See the calls
Everything on Grok Voice Think Fast 2.0
More figures
Scenarios passed every run
55 of 82
Agent responses
96.7%
Interruption
4.92 of 5
Response time, p90
1.94s
Success, Booking calls
96.0%
Success, Medicare calls
24.6%
Setup
Model
grok-voice-think-fast-2.0
Setting
Reasoning high
Voice
eve
Turn taking
Service turn detection
Success rate
76.0%
Data accuracy
79.2%
Went quiet
3.3%
Response time
2.88s
Cost / min
$0.0703
What we saw
Reliable on bookings.
Slowest to reply, and after several tool calls replies lag or go missing.
Hands Medicare callers over with the wrong details.
Compare with Phonic v1 See the calls
Everything on Gemini 3.1 Flash Live Preview
More figures
Scenarios passed every run
51 of 82
Agent responses
95.9%
Interruption
4.98 of 5
Response time, p90
4.00s
Success, Booking calls
93.2%
Success, Medicare calls
31.9%
Setup
Model
models/gemini-3.1-flash-live-preview
Setting
Thinking level high
Voice
Charon
Turn taking
Shared local turn detection
Success rate
70.7%
Data accuracy
78.5%
Went quiet
6.1%
Response time
2.31s
Cost / min
$0.0224
What we saw
Every plain booking and every call with several requests passed.
Often skips a step on Medicare calls, most often routing the caller or writing the handoff note.
Stops when cut off during the Medicare disclosure, but doesn't go back to finish it.
The weakest on silences, coughs and packet loss, and goes quiet more often than most on Medicare calls.
Compare with Phonic v1 See the calls
Everything on Deepslate Opal
More figures
Scenarios passed every run
48 of 82
Agent responses
91.1%
Interruption
4.94 of 5
Response time, p90
2.88s
Success, Booking calls
84.7%
Success, Medicare calls
34.8%
Setup
Model
opal
Setting
No reasoning setting
Voice
ElevenLabs 21m00Tcm4TlvDq8ikWAM
Turn taking
Service turn detection
Success rate
64.6%
Data accuracy
66.2%
Went quiet
1.2%
Response time
1.94s
Cost / min
$0.0061
What we saw
The cheapest model of the field, and its reply time barely changes from one turn to the next.
Passed every call with an accented caller.
Says it is saving, routing or transferring without making the call: it skips the consent record and never transfers, so it fails almost every Medicare call.
Rarely hangs up on its own, and goes quiet when its service returns an error.
Compare with Phonic v1
Everything on Higgs Realtime
More figures
Scenarios passed every run
46 of 82
Agent responses
90.2%
Interruption
5.00 of 5
Response time, p90
2.02s
Success, Booking calls
89.3%
Success, Medicare calls
1.4%
Setup
Model
higgs-realtime
Setting
No reasoning setting
Voice
default
Turn taking
Service turn detection
Success rate
76.0%
Data accuracy
84.4%
Went quiet
5.3%
Response time
2.07s
Cost / min
n/a
What we saw
One of the strongest realtime models on Medicare calls.
Weakest on plain bookings: every tool call right, outcome still missed.
Unpredictable run to run, and can go silent under background conversation.
Compare with Phonic v1 See the calls
Everything on Gemini 3.8 Live Extended Thinking
More figures
Scenarios passed every run
43 of 82
Agent responses
88.6%
Interruption
4.95 of 5
Response time, p90
2.74s
Success, Booking calls
78.5%
Success, Medicare calls
69.6%
Setup
Model
gemini-3.8-live-extended-thinking
Setting
Thinking level high
Voice
Charon
Turn taking
Service turn detection
Success rate
59.3%
Data accuracy
59.9%
Went quiet
1.2%
Response time
2.18s
Cost / min
n/a
What we saw
Best voice score of the field.
Misses caller names and phone numbers, and rarely passes Medicare calls.
Repeats tool calls more than any other model, with the longest replies.
Compare with Phonic v1 See the calls
Everything on Amazon Nova 2 Sonic
More figures
Scenarios passed every run
41 of 82
Agent responses
93.9%
Interruption
4.99 of 5
Response time, p90
2.76s
Success, Booking calls
80.2%
Success, Medicare calls
5.8%
Setup
Model
amazon.nova-2-sonic-v1:0
Setting
Endpointing medium
Voice
matthew
Turn taking
Shared local turn detection
Success rate
61.4%
Data accuracy
64.5%
Went quiet
2.5%
Response time
2.08s
Cost / min
$0.0202
What we saw
Gets most bookings right, at a fraction of the cost.
Forgets earlier details: not one Medicare record was exact.
Mishears phone numbers, and rarely recovers from a tool error.
Compare with Phonic v1 See the calls
Everything on GPT Realtime 2.1 Mini
More figures
Scenarios passed every run
41 of 82
Agent responses
93.5%
Interruption
4.99 of 5
Response time, p90
2.75s
Success, Booking calls
83.6%
Success, Medicare calls
4.3%
Setup
Model
gpt-realtime-2.1-mini
Setting
Reasoning high
Voice
marin
Turn taking
Service turn detection
Success rate
69.1%
Data accuracy
70.5%
Went quiet
2.0%
Response time
1.91s
Cost / min
n/a
What we saw
Replies sooner, and more bookings reach the outcome.
Medicare records are wrong far more often.
Weakest when a caller corrects themselves, and the least predictable.
Compare with Phonic v1 See the calls
Everything on Gemini 3.8 Live
More figures
Scenarios passed every run
34 of 82
Agent responses
94.3%
Interruption
4.98 of 5
Response time, p90
2.23s
Success, Booking calls
75.7%
Success, Medicare calls
52.2%
Setup
Model
gemini-3.8-live
Setting
No thinking level
Voice
Charon
Turn taking
Service turn detection

For reference, not ranked: speech-to-text, then an LLM, then text-to-speech

Success rate
93.1%
Data accuracy
94.1%
Went quiet
0.8%
Response time
2.17s
Cost / min
n/a
What we saw
Most scenarios pass every run, and none fails every run.
Almost never leaves the caller waiting.
Not the quickest to reply.
Now and then forgets something the caller said earlier in a Medicare call.
Compare with Phonic v1 See the calls
More figures
Scenarios passed every run
68 of 82
Agent responses
99.2%
Interruption
4.96 of 5
Response time, p90
2.51s
Success, Booking calls
97.2%
Success, Medicare calls
82.6%
Setup
Model
flux-general-en → gpt-4.1 → eleven_flash_v2_5
Setting
No reasoning setting
Voice
ElevenLabs 21m00Tcm4TlvDq8ikWAM
Turn taking
Speech-to-text turn detection

Open a model for success rate, data accuracy, how often it went quiet, response time, cost and notes from the calls.

How to read this table

Reliability (pass³) · Passed all 3 runs: Scenarios the agent passed on every one of its 3 runs.
Success rate · Passed every check: Calls where the agent said the right things and saved the right details.
Data accuracy · Saved exactly right: Calls where every detail the agent saved matched. Calls that save nothing are left out.
Went quiet · Went quiet 10 s or more: Calls where the agent went quiet for 10 seconds or more after the caller spoke. Lower is better.
Response time · Typical reply time: Median time from the caller finishing to the agent starting to reply. Lower is better.
Cost / min · Published price: One minute of call at the vendor's published rates. Blank where no rate is published.

Results by call

Booking calls are easy, and most models get them right almost every time. Medicare calls ask for far more, and about half get through half of them.

Booking callsMedicare calls

97%
74%
Phonic v1
86%
61%
GPT Realtime 2.1
93%
57%
Cascade
81%
43%
Azure Realtime
97%
43%
Grok Voice 3.0
78%
43%
Prism
92%
35%
GPT-Live 1 + GPT-6 Sol
59%
35%
Gemini 3.8 Live Extended Thinking
53%
13%
Gemini 3.8 Live
83%
9%
Gemini 3.1 Flash Live Preview
92%
4%
Grok Voice Think Fast 2.0
68%
4%
Amazon Nova 2 Sonic
68%
4%
GPT Realtime 2.1 Mini
80%
4%
Deepslate Opal
78%
0%
Higgs Realtime

Scenarios that passed on all 3 runs. Cascade: the older three-step setup, shown for reference and not ranked.

No model wins on everything

A caller wants the right answer and a quick reply. You want a low bill. No single model gives all three, so pick the measure that matters most to you.

Scenarios that passed on all 3 runs, pass³. Higher is better.

#ModelReliability

01Phonic v1Intelligence high90.2%
02Grok Voice 3.0Reasoning high81.7%
03GPT Realtime 2.1Reasoning high79.3%
04GPT-Live 1 + GPT-6 SolSol on low thinking75.6%
05Azure RealtimeNo reasoning setting70.7%
06PrismNo reasoning setting68.3%
07Grok Voice Think Fast 2.0Reasoning high67.1%
08Gemini 3.1 Flash Live PreviewThinking level high62.2%
09Deepslate OpalNo reasoning setting58.5%
10Higgs RealtimeNo reasoning setting56.1%
11Gemini 3.8 Live Extended ThinkingThinking level high52.4%
12Amazon Nova 2 SonicEndpointing medium50.0%
13GPT Realtime 2.1 MiniReasoning high50.0%
14Gemini 3.8 LiveNo thinking level41.5%

refCascadeNot ranked82.9%

How each model fails

Two models can share a score and fail for different reasons. Some slip on the details, some go quiet, and some pass a call one run and fail it the next.

Hardest kind of call

Medicare compliance & routing

43%average pass rate across 13 models, 20 scenarios

1Phonic v192%
2GPT Realtime 2.178%
3Gemini 3.8 Live Extended Thinking70%
4Azure Realtime65%
5Grok Voice 3.060%
6GPT-Live 1 + GPT-6 Sol50%
7Gemini 3.8 Live48%
8Deepslate Opal33%
9Gemini 3.1 Flash Live Preview32%
10Grok Voice Think Fast 2.025%
11Amazon Nova 2 Sonic2%
12Higgs Realtime0%
13GPT Realtime 2.1 Mini0%

Kind of callAvg. pass rate

1. 1Grok Voice 3.096%
2. 2Azure Realtime92%
3. 3GPT-Live 1 + GPT-6 Sol88%
4. 4Phonic v183%
5. 5GPT Realtime 2.183%
6. 6Gemini 3.1 Flash Live Preview83%
7. 7Higgs Realtime79%
8. 8GPT Realtime 2.1 Mini79%
9. 9Deepslate Opal75%
10. 10Grok Voice Think Fast 2.071%
11. 11Gemini 3.8 Live Extended Thinking67%
12. 12Amazon Nova 2 Sonic63%
13. 13Gemini 3.8 Live38%
1. 1Phonic v1100%
2. 2Grok Voice 3.0100%
3. 3GPT Realtime 2.1100%
4. 4Azure Realtime100%
5. 5Grok Voice Think Fast 2.0100%
6. 6Higgs Realtime100%
7. 7GPT-Live 1 + GPT-6 Sol83%
8. 8Gemini 3.1 Flash Live Preview83%
9. 9GPT Realtime 2.1 Mini83%
10. 10Deepslate Opal67%
11. 11Gemini 3.8 Live Extended Thinking50%
12. 12Amazon Nova 2 Sonic50%
13. 13Gemini 3.8 Live50%
1. 1Phonic v1100%
2. 2Grok Voice 3.0100%
3. 3GPT-Live 1 + GPT-6 Sol100%
4. 4Grok Voice Think Fast 2.0100%
5. 5Gemini 3.1 Flash Live Preview86%
6. 6Higgs Realtime81%
7. 7Amazon Nova 2 Sonic81%
8. 8GPT Realtime 2.1 Mini81%
9. 9GPT Realtime 2.176%
10. 10Deepslate Opal71%
11. 11Azure Realtime67%
12. 12Gemini 3.8 Live Extended Thinking67%
13. 13Gemini 3.8 Live62%
1. 1Phonic v1100%
2. 2GPT Realtime 2.194%
3. 3Deepslate Opal94%
4. 4Azure Realtime89%
5. 5Grok Voice Think Fast 2.089%
6. 6Grok Voice 3.083%
7. 7GPT-Live 1 + GPT-6 Sol83%
8. 8Gemini 3.1 Flash Live Preview83%
9. 9Gemini 3.8 Live Extended Thinking83%
10. 10Gemini 3.8 Live83%
11. 11Higgs Realtime78%
12. 12GPT Realtime 2.1 Mini78%
13. 13Amazon Nova 2 Sonic67%
1. 1Grok Voice 3.0100%
2. 2Grok Voice Think Fast 2.0100%
3. 3Phonic v196%
4. 4GPT-Live 1 + GPT-6 Sol96%
5. 5Azure Realtime92%
6. 6GPT Realtime 2.188%
7. 7Higgs Realtime88%
8. 8GPT Realtime 2.1 Mini88%
9. 9Gemini 3.1 Flash Live Preview83%
10. 10Gemini 3.8 Live Extended Thinking83%
11. 11Gemini 3.8 Live79%
12. 12Amazon Nova 2 Sonic63%
13. 13Deepslate Opal54%
1. 1Phonic v1100%
2. 2Grok Voice 3.0100%
3. 3GPT Realtime 2.1100%
4. 4Azure Realtime100%
5. 5Gemini 3.8 Live100%
6. 6GPT-Live 1 + GPT-6 Sol92%
7. 7Grok Voice Think Fast 2.092%
8. 8Gemini 3.1 Flash Live Preview92%
9. 9Higgs Realtime92%
10. 10Amazon Nova 2 Sonic92%
11. 11Deepslate Opal83%
12. 12Gemini 3.8 Live Extended Thinking75%
13. 13GPT Realtime 2.1 Mini25%
1. 1GPT Realtime 2.1100%
2. 2GPT-Live 1 + GPT-6 Sol100%
3. 3Azure Realtime100%
4. 4Amazon Nova 2 Sonic100%
5. 5Gemini 3.8 Live Extended Thinking93%
6. 6Phonic v187%
7. 7Grok Voice 3.087%
8. 8Gemini 3.1 Flash Live Preview87%
9. 9GPT Realtime 2.1 Mini87%
10. 10Gemini 3.8 Live87%
11. 11Grok Voice Think Fast 2.080%
12. 12Higgs Realtime80%
13. 13Deepslate Opal73%
1. 1Phonic v1100%
2. 2GPT-Live 1 + GPT-6 Sol100%
3. 3Grok Voice Think Fast 2.0100%
4. 4Gemini 3.1 Flash Live Preview100%
5. 5Deepslate Opal100%
6. 6GPT Realtime 2.198%
7. 7Azure Realtime96%
8. 8Grok Voice 3.093%
9. 9Amazon Nova 2 Sonic89%
10. 10GPT Realtime 2.1 Mini89%
11. 11Higgs Realtime87%
12. 12Gemini 3.8 Live87%
13. 13Gemini 3.8 Live Extended Thinking76%
1. 1Phonic v1100%
2. 2Grok Voice 3.0100%
3. 3GPT Realtime 2.1100%
4. 4Gemini 3.1 Flash Live Preview100%
5. 5Deepslate Opal100%
6. 6GPT-Live 1 + GPT-6 Sol95%
7. 7Grok Voice Think Fast 2.095%
8. 8Higgs Realtime95%
9. 9Gemini 3.8 Live Extended Thinking95%
10. 10GPT Realtime 2.1 Mini90%
11. 11Azure Realtime86%
12. 12Gemini 3.8 Live86%
13. 13Amazon Nova 2 Sonic81%

Average pass rate across the ranked models for each kind of call. Open one to see every model.

Put two models head to head

Pick any two models and see them side by side. Every pair has its own page with the full numbers.

Compare

MeasurePhonic v1Cascade

Reliability90.2%82.9%
Success rate95.1%93.1%
Data accuracy97.5%94.1%
Went quiet0.0%0.8%
Response time1.63s2.17s
Cost / min$0.0600n/a
Measures won5 of 60 of 6

All suites. The check marks the better value on each measure.All 105 comparisons

The most reliable model against the next six

Every comparison

How we tested

Every model ran the same agent on the same calls and was scored the same way.

Step 1Only the model changesSame Pipecat agent, prompt and tools for every model. Only the realtime model is swapped.
Step 2Live calls, three runs eachA simulated caller runs all 82 scenarios, three times. Every call counts.
Step 3Words and data both countA call passes only if the outcome is right and every saved field matches.
Step 4Every call is publicEach model links to its calls, with transcripts, tool calls and scores.

Reference agent Suites and tools Test cases

Speech-to-speech benchmark FAQ

What is the best speech-to-speech model in 2026?

Phonic v1 is the most reliable, passing 90.2% of scenarios on every run. Rime Prism replies fastest, at 1.34s median. Higgs Realtime is the cheapest, at $0.0061 a minute of call. The right pick depends on which of the three your voice agent needs most, and all three are measured on the same live calls.

Phonic v1 is the most reliable, passing 90.2% of scenarios on every run. Rime Prism replies fastest, at 1.34s median. Higgs Realtime is the cheapest, at $0.0061 a minute of call. The right pick depends on which of the three your voice agent needs most, and all three are measured on the same live calls.
It measures complete realtime voice models on live agent calls, including reliability, task success, saved-data accuracy, how often the agent went quiet, response time, and cost.
The leaderboard uses pass³, the share of scenarios where all three retained calls passed every required check. A scenario must work repeatedly to count.
Yes. Every model runs the same voice agent, prompt, tools, connection, caller scenarios, and scoring rules. Only the realtime model changes.
Native speech-to-speech models handle audio directly. A cascade connects speech-to-text, an LLM, and text-to-speech. Cekura runs both through the same agent scenarios so reliability and response-time differences remain comparable.
Yes. The comparison tool provides head-to-head reliability, accuracy, latency, failure, and cost results for any two tested models.

Test your voice agent the same way.

Cekura runs these simulated calls on your agent and scores every one.

Book a demo Start free

Speech-to-speech model benchmarks

Reliability vs. response time

Leaderboard

Results by call

No model wins on everything

How each model fails

Medicare compliance & routing

Voice quality

Slow replies

Cost per success

Response time

Cost per minute

Interruption handling