Roark Live Bench · Suite v1.3 · Sweep of 30 October 2026
Voice AI models, ranked on live phone calls.
Every speech-to-text, language model and text-to-speech model, run in every combination on the same real call to the same agent. A model is ranked on every stack it runs in; the stacks are the evidence.
Our picks
Roark scoreTask completionMedian responseCost per minute- Best overallUniversal-3.5 Pro + GPT-6.1 Sol + Turbo v2.570 / 10096%4.0 s$0.12
- Best valueUniversal-3.5 Pro + GPT-6.1 Sol + Coda69 / 10098%4.0 s$0.040
- FastestGrok Voice 3.066 / 10087%1.9 s$0.086
- Best speech-to-speechGPT Live 1 (Sol)65 / 10085%2.3 s$0.061
- Best in customer supportGrok Voice 3.085 / 10095%1.7 s$0.083
- Best in healthcareUniversal-3.5 Pro + Gemini 3.8 Flash + Flash v2.575 / 10097%4.3 s$0.10
- Best in home servicesGPT Live 1 (Luna)79 / 10095%2.1 s$0.052
- Best in insuranceUniversal-3.5 Pro + GPT-4o + Sonic-3.676 / 10098%3.2 s$0.042
The read
- 430 of 642 cascades complete the task in at least 90% of calls, and the top 10 sit within 1 point of each other. What separates the top stacks is speed and cost.
- Median response time runs from 2.0 s to 6.2 s on the same agent and the same caller, a spread of 4.2 s that comes entirely from the models behind it.
- No stack wins everything. The fastest, Speechmatics Linden-1 (own VAD) + Anthropic Claude Haiku 4.5 + ElevenLabs Turbo v2.5 at 2.0 s, ranks 559 of 642 on task completion; the most complete is Speechmatics Linden-1 (own VAD) + OpenAI GPT-6.1 Sol + OpenAI TTS-1 HD at 99%.
- The language model matters most: swapping it moves task completion by 7 points, against 3 for the speech engine and 2 for the transcriber.
- The fastest speech-to-speech model answers 0.1 s sooner than the fastest cascade, and its completion rate sits 12 points below the best cascade.
Cascade vs Speech-to-Speech
The first choice to make. A cascade chains three models: one transcribes, one decides what to say, one speaks. A speech-to-speech model does all three at once. Both run the same agent on the same calls here.
- Cascades lead on task completion, tool use, time to first word, barge-in handling and cost per task; speech-to-speech leads on response time. Neither wins outright, so the right choice depends on what your calls need.
- Speech-to-speech costs 1.6 times as much per completed task: $0.24 for speech-to-speech against $0.15 for cascades.
- The biggest difference is response time: 3.8 s for cascades against 2.1 s for speech-to-speech.
| Measure | Cascade | Speech-to-speech |
|---|---|---|
| Task completion | 94% (leads) | 85% |
| Tool use | 78% (leads) | 78% |
| Time to first word | 0.8 s (leads) | 1.1 s |
| Response time | 3.8 s | 2.1 s (leads) |
| Barge-in handling | 100% (leads) | 96% |
| Cost per task | $0.15 (leads) | $0.24 |
The three strongest of each, averaged. Cascades: AssemblyAI Universal-3.6 Pro + OpenAI GPT-6.1 Sol + Rime Coda; Speechmatics Linden-1 (own VAD) + OpenAI GPT-4.1 + Rime Coda; AssemblyAI Universal-3.6 Pro + Google Gemini 3.8 Flash + Rime Coda. Speech-to-speech: grok-voice-think-fast-2.0; gpt-live-1 (backend gpt-6-luna); grok-voice-3.0. Each measure is scored 0 to 100 on its own scale, further out is better; the table gives the same figures in their own units.
Which to use
- Choose speech-to-speech when the pause matters more than the outcome: short transactional calls, callers who interrupt, queues where every half second costs you callers. Expect to pay more, and choose the model carefully (grok-voice-3.0 leads here).
- Choose a cascade when the call has to complete, when cost matters, or when you need to see which stage to fix. It is 12 points ahead on task completion here.
- Costs are what each stack’s providers charged for these calls, at the rates in effect when they ran. Telephony and hosting are not included.
Speech-to-Text
The transcriber turns the caller's speech into text. It decides what the agent thinks it heard, and nothing after it can start until it finishes.
- AssemblyAI Universal-3.6 Pro, AssemblyAI Universal-3.5 Pro, Deepgram Nova-3 and Speechmatics Enhanced (realtime) are level: none leads by more than the margin of error. Of them, AssemblyAI Universal-3.6 Pro costs least, at $0.030 a call. Avoid Speechmatics Linden-1 (own VAD): last on semantic word error rate (2.2%).
- Speechmatics Linden-1 (own VAD) delivers its transcript in 0.49 s, OpenAI Whisper-1 in 1.93 s: 1.44 s added to every turn of the call.
- AssemblyAI Universal-3.6 Pro is the most accurate: it changes the meaning of 0.9% of the caller's words, against 2.2% for Speechmatics Linden-1 (own VAD).
- In 341 of 648 cascades the transcriber is the slowest step of the turn, whichever model runs it.
Roark score, best first
| # | |||||
|---|---|---|---|---|---|
| 1 | Universal-3.6 ProAssemblyAI | 90 / 100. Semantic word error rate 0.9%, 61.2 of 71 points; Time to final transcript 0.89 s, 28.6 of 29 points(among the best) | 0.9%(among the best) | 0.89 s(among the best) | $0.030 |
| 2 | Universal-3.5 ProAssemblyAI | 84 / 100. Semantic word error rate 1.1%, 55.7 of 71 points; Time to final transcript 0.90 s, 28.6 of 29 points(among the best) | 1.1%(among the best) | 0.90 s | $0.030 |
| 3 | Nova-3Deepgram | 82 / 100. Semantic word error rate 1.1%, 53.5 of 71 points; Time to final transcript 0.89 s, 28.6 of 29 points(among the best) | 1.1%(among the best) | 0.89 s(among the best) | $0.031 |
| 4 | Enhanced (realtime)Speechmatics | 82 / 100. Semantic word error rate 1.1%, 53.3 of 71 points; Time to final transcript 0.98 s, 28.6 of 29 points(among the best) | 1.1%(among the best) | 0.98 s | $0.054 |
| 5 | Linden-1Speechmatics | 79 / 100. Semantic word error rate 1.3%, 50.0 of 71 points; Time to final transcript 0.99 s, 28.6 of 29 points | 1.3% | 0.99 s | $0.020(among the best) |
| 6 | Nova-2Deepgram | 71 / 100. Semantic word error rate 1.5%, 42.2 of 71 points; Time to final transcript 0.89 s, 28.6 of 29 points | 1.5% | 0.89 s(among the best) | $0.023(among the best) |
| 7 | Standard (realtime)Speechmatics | 66 / 100. Semantic word error rate 1.7%, 37.5 of 71 points; Time to final transcript 0.96 s, 28.6 of 29 points | 1.7% | 0.96 s | $0.031 |
| 8 | Whisper-1OpenAI | 56 / 100. Semantic word error rate 1.4%, 45.5 of 71 points; Time to final transcript 1.93 s, 10.9 of 29 points | 1.4% | 1.93 s | $0.025 |
| 9 | Linden-1 (own VAD)Speechmatics | 52 / 100. Semantic word error rate 2.2%, 23.1 of 71 points; Time to final transcript 0.49 s, 28.6 of 29 points | 2.2% | 0.49 s(among the best) | $0.020(among the best) |
Open a model to see where it ranks and the stacks it works best in.
Language model
The language model decides what the agent says and does. It has the most influence on whether the task gets done, and usually takes the largest share of each turn.
- Pick OpenAI GPT-6.1 Sol: it leads on task completion (95%), tool use (77%) and instruction following (4.2 / 5). If speed matters more, Anthropic Claude Haiku 4.5 is the quickest, at 0.36 s, and completes 91%. Avoid OpenAI GPT-4o: last on task completion (88%).
- Anthropic Claude Haiku 4.5 starts answering in 0.36 s, Google Gemini 3.8 Flash in 1.47 s: 1.11 s added to every turn of the call.
- OpenAI GPT-6.1 Sol uses its tools most reliably: it handles its tools without a mistake on 77% of calls, against 38% for OpenAI GPT-5.1.
- In 161 of 648 cascades the language model is the slowest step of the turn, whichever model runs it.
Roark score, best first
| # | ||||||||
|---|---|---|---|---|---|---|---|---|
| 1 | GPT-6.1 SolOpenAI | 66 / 100. Task completion 95%, 27.0 of 36 points; Tool use 77%, 15.0 of 27 points; Instruction following 4.2 / 5, 11.3 of 18 points; Time to first token 0.97 s, 12.4 of 18 points(among the best) | 95%(among the best) | 4.2 / 5(among the best) | 77%(among the best) | 0.97 s | 73% | $0.019(among the best) |
| 2 | Claude Sonnet 5.5Anthropic | 61 / 100. Task completion 94%, 25.1 of 36 points; Tool use 72%, 11.8 of 27 points; Instruction following 4.2 / 5, 10.8 of 18 points; Time to first token 0.90 s, 13.3 of 18 points(among the best) | 94%(among the best) | 4.2 / 5(among the best) | 72%(among the best) | 0.90 s | – | $0.138 |
| 3 | Gemini 3.8 FlashGoogle | 51 / 100. Task completion 95%, 26.5 of 36 points; Tool use 62%, 6.7 of 27 points; Instruction following 4.2 / 5, 11.1 of 18 points; Time to first token 1.47 s, 6.4 of 18 points(among the best) | 95%(among the best) | 4.2 / 5(among the best) | 62%(among the best) | 1.47 s | – | $0.016(among the best) |
| 4 | GPT-4.1OpenAI | 50 / 100. Task completion 92%, 21.2 of 36 points; Tool use 48%, 0.0 of 27 points; Instruction following 4.2 / 5, 10.8 of 18 points; Time to first token 0.52 s, 18.0 of 18 points | 92% | 4.2 / 5(among the best) | 48% | 0.52 s | 82%(among the best) | $0.022 |
| 5 | Claude Haiku 4.5Anthropic | 49 / 100. Task completion 91%, 19.4 of 36 points; Tool use 47%, 0.0 of 27 points; Instruction following 4.2 / 5, 11.0 of 18 points; Time to first token 0.36 s, 18.2 of 18 points | 91% | 4.2 / 5(among the best) | 47% | 0.36 s(among the best) | – | $0.062 |
| 6 | GPT-4o miniOpenAI | 47 / 100. Task completion 90%, 17.4 of 36 points; Tool use 51%, 0.7 of 27 points; Instruction following 4.2 / 5, 10.8 of 18 points; Time to first token 0.45 s, 18.2 of 18 points | 90% | 4.2 / 5(among the best) | 51% | 0.45 s(among the best) | 81%(among the best) | $0.002(among the best) |
| 7 | GPT-5.1OpenAI | 44 / 100. Task completion 89%, 17.0 of 36 points; Tool use 38%, 0.0 of 27 points; Instruction following 4.1 / 5, 10.1 of 18 points; Time to first token 0.59 s, 17.1 of 18 points | 89% | 4.1 / 5 | 38% | 0.59 s | 74%(among the best) | $0.031 |
| 8 | GPT-4oOpenAI | 44 / 100. Task completion 88%, 15.0 of 36 points; Tool use 50%, 0.0 of 27 points; Instruction following 4.2 / 5, 10.9 of 18 points; Time to first token 0.45 s, 18.2 of 18 points | 88% | 4.2 / 5(among the best) | 50% | 0.45 s(among the best) | – | $0.027 |
Open a model to see where it ranks and the stacks it works best in.
Text-to-Speech
The speech engine is the voice the caller hears, and the last step before they hear anything.
- Cartesia Sonic-3.6, ElevenLabs v4 Turbo and Cartesia Sonic-3.5 are level: none leads by more than the margin of error. Of them, Cartesia Sonic-3.5 costs least, at $0.165 a call. Avoid OpenAI TTS-1 HD: last on readback accuracy (80%), word error rate (8.7%) and time to first audio byte (2.09 s).
- ElevenLabs v4 Turbo starts speaking in 0.11 s, OpenAI TTS-1 HD in 2.09 s: 1.97 s added to every turn of the call.
- Cartesia Sonic-3.6 is the most accurate: it reads codes, amounts and dates back correctly 87% of the time, against 80% for OpenAI TTS-1 HD.
- Cartesia Sonic-3.6 sounds the most human: 3.8 / 5, against 3.4 / 5 for ElevenLabs Flash v2.5.
- In 99 of 648 cascades the speech engine is the slowest step of the turn, most often with OpenAI TTS-1 HD (54 of them).
Roark score, best first
| # | ||||||||
|---|---|---|---|---|---|---|---|---|
| 1 | Sonic-3.6Cartesia | 77 / 100. Readback accuracy 87%, 25.6 of 38 points; Word error rate 6.2%, 14.5 of 25 points; Time to first audio byte 0.12 s, 25.0 of 25 points; Voice consistency 99%, 12.1 of 13 points(among the best) | 6.2% | 87%(among the best) | 3.8 / 5(among the best) | 99% | 0.12 s(among the best) | $0.168 |
| 2 | v4 TurboElevenLabs | 77 / 100. Readback accuracy 87%, 25.1 of 38 points; Word error rate 5.7%, 15.8 of 25 points; Time to first audio byte 0.11 s, 25.0 of 25 points; Voice consistency 98%, 11.2 of 13 points(among the best) | 5.7%(among the best) | 87%(among the best) | 3.6 / 5 | 98% | 0.11 s(among the best) | $0.470 |
| 3 | Sonic-3.5Cartesia | 77 / 100. Readback accuracy 86%, 24.6 of 38 points; Word error rate 6.2%, 14.6 of 25 points; Time to first audio byte 0.13 s, 25.0 of 25 points; Voice consistency 100%, 12.4 of 13 points(among the best) | 6.2% | 86%(among the best) | 3.8 / 5(among the best) | 100%(among the best) | 0.13 s | $0.165 |
| 4 | Sonic-3Cartesia | 73 / 100. Readback accuracy 85%, 23.3 of 38 points; Word error rate 6.9%, 12.8 of 25 points; Time to first audio byte 0.17 s, 25.0 of 25 points; Voice consistency 99%, 12.1 of 13 points | 6.9% | 85% | 3.4 / 5 | 99% | 0.17 s | $0.165 |
| 5 | Turbo v2.5ElevenLabs | 71 / 100. Readback accuracy 80%, 19.0 of 38 points; Word error rate 5.7%, 15.6 of 25 points; Time to first audio byte 0.15 s, 25.0 of 25 points; Voice consistency 98%, 11.5 of 13 points | 5.7%(among the best) | 80% | 3.4 / 5 | 98% | 0.15 s | $0.451 |
| 6 | Flash v2.5ElevenLabs | 70 / 100. Readback accuracy 80%, 18.5 of 38 points; Word error rate 6%, 14.9 of 25 points; Time to first audio byte 0.12 s, 25.0 of 25 points; Voice consistency 98%, 11.5 of 13 points | 6%(among the best) | 80% | 3.4 / 5 | 98% | 0.12 s(among the best) | $0.437 |
| 7 | CodaRime | 68 / 100. Readback accuracy 83%, 21.7 of 38 points; Word error rate 8.1%, 9.8 of 25 points; Time to first audio byte 0.18 s, 25.0 of 25 points; Voice consistency 99%, 11.8 of 13 points | 8.1% | 83% | 3.7 / 5(among the best) | 99% | 0.18 s | $0.116(among the best) |
| 8 | TTS-1OpenAI | 55 / 100. Readback accuracy 82%, 21.0 of 38 points; Word error rate 6.1%, 14.8 of 25 points; Time to first audio byte 1.13 s, 7.1 of 25 points; Voice consistency 100%, 12.5 of 13 points | 6.1% | 82% | 3.6 / 5 | 100%(among the best) | 1.13 s | $0.036(among the best) |
| 9 | TTS-1 HDOpenAI | 39 / 100. Readback accuracy 80%, 18.3 of 38 points; Word error rate 8.7%, 8.4 of 25 points; Time to first audio byte 2.09 s, 0.0 of 25 points; Voice consistency 100%, 12.4 of 13 points | 8.7% | 80% | 3.6 / 5 | 100%(among the best) | 2.09 s | $0.065(among the best) |
Open a model to see where it ranks and the stacks it works best in.
Speech-to-Speech models
One model listens, reasons and speaks, so there is nothing to swap and no stages to split. Each is judged on the whole call, with the same calls and the same scoring as every cascade.
- The model matters as much as the architecture: xAI Grok Voice 3.0 completes 87% of tasks, ElevenLabs Realtime v1 Mini 42%.
- xAI Grok Voice 3.0 answers in 1.9 s against 4.2 s for Google Gemini 2.5 Flash native audio.
- Price varies as much as quality: $0.06 per completed task for Amazon Nova 2 Sonic, $0.69 for OpenAI GPT Realtime 2.
Roark score, best first
Whiskers show the 95% interval. The shaded models tie for first: these calls cannot separate them.
| # | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1= (tied) | GPT Live 1 (Sol)OpenAI | 65 / 100±3.8. Task completion 85%, 6.0 of 23 points; Response time 2.3 s, 16.6 of 18 points; Instruction following 4.2 / 5, 8.2 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 87%, 10.2 of 14 points; Time to first word 1.4 s, 3.9 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points(among the best) | 85%(among the best) | 4.2 / 5 | 87%(among the best) | 2.3 s(among the best) | 1.4 s | 77% | 3.1 / 5 | 100%(among the best) | $0.24 |
| 2= (tied) | Grok Voice 3.0xAI | 66 / 100±4.7. Task completion 87%, 8.0 of 23 points; Response time 1.9 s, 18.2 of 18 points; Instruction following 4.5 / 5, 10.5 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 66%, 4.4 of 14 points; Time to first word 0.9 s, 4.5 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points(among the best) | 87%(among the best) | 4.5 / 5(among the best) | 66% | 1.9 s(among the best) | 0.9 s | 94% | 3.0 / 5 | 100%(among the best) | $0.26 |
| 3= (tied) | GPT Live 1 (Luna)OpenAI | 65 / 100±3.5. Task completion 85%, 5.7 of 23 points; Response time 2.3 s, 16.6 of 18 points; Instruction following 4.2 / 5, 8.4 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 86%, 9.9 of 14 points; Time to first word 1.5 s, 3.8 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points(among the best) | 85%(among the best) | 4.2 / 5 | 86%(among the best) | 2.3 s(among the best) | 1.5 s | 93% | 3.1 / 5 | 100%(among the best) | $0.21 |
| 4= (tied) | Grok Voice Think Fast 2.0xAI | 65 / 100±3.7. Task completion 82%, 2.8 of 23 points; Response time 2.2 s, 17.4 of 18 points; Instruction following 4.6 / 5, 10.7 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 82%, 8.6 of 14 points; Time to first word 0.9 s, 4.5 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points(among the best) | 82% | 4.6 / 5(among the best) | 82%(among the best) | 2.2 s(among the best) | 0.9 s | 100%(among the best) | 3.6 / 5 | 100%(among the best) | $0.26 |
| 5= (tied) | Gemini 3.8 LiveGoogle | 53 / 100±4.8. Task completion 78%, 0.0 of 23 points; Response time 2.8 s, 14.6 of 18 points; Instruction following 4.3 / 5, 8.7 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 69%, 5.2 of 14 points; Time to first word 1.6 s, 3.6 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points | 78% | 4.3 / 5(among the best) | 69% | 2.8 s | 1.6 s | 100%(among the best) | 4.1 / 5(among the best) | 100%(among the best) | $0.14(among the best) |
| 6= (tied) | GPT Realtime 2.1OpenAI | 53 / 100±5.7. Task completion 78%, 0.0 of 23 points; Response time 2.6 s, 15.6 of 18 points; Instruction following 3.8 / 5, 5.6 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 75%, 6.8 of 14 points; Time to first word 0.8 s, 4.5 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 99%, 4.3 of 5 points | 78% | 3.8 / 5 | 75% | 2.6 s | 0.8 s(among the best) | 83% | 3.9 / 5(among the best) | 99% | $0.43 |
| 7= (tied) | Ultravox v0.7Ultravox | 48 / 100±5.0. Task completion 79%, 0.0 of 23 points; Response time 2.4 s, 16.2 of 18 points; Instruction following 3.7 / 5, 4.9 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 61%, 3.0 of 14 points; Time to first word 1.6 s, 3.6 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points | 79% | 3.7 / 5 | 61% | 2.4 s | 1.6 s | 94% | 3.6 / 5 | 100%(among the best) | $0.22 |
| 8= (tied) | Gemini 3.1 Flash Live (preview)Google | 48 / 100±5.1. Task completion 80%, 0.0 of 23 points; Response time 2.7 s, 15.1 of 18 points; Instruction following 4.2 / 5, 8.0 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 38%, 0.0 of 14 points; Time to first word 1.5 s, 3.8 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points | 80% | 4.2 / 5 | 38% | 2.7 s | 1.5 s | 100%(among the best) | 4.0 / 5(among the best) | 100%(among the best) | $0.15 |
| 9= (tied) | GPT Realtime 2OpenAI | 48 / 100±6.7. Task completion 76%, 0.0 of 23 points; Response time 2.9 s, 14.0 of 18 points; Instruction following 3.8 / 5, 5.4 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 64%, 3.7 of 14 points; Time to first word 0.8 s, 4.5 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points | 76% | 3.8 / 5 | 64% | 2.9 s | 0.8 s(among the best) | 91% | 3.7 / 5 | 100%(among the best) | $0.69 |
| 10= (tied) | Nova 2 SonicAmazon | 46 / 100±5.1. Task completion 69%, 0.0 of 23 points; Response time 3.0 s, 13.8 of 18 points; Instruction following 4.2 / 5, 7.9 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 48%, 0.0 of 14 points; Time to first word 1.5 s, 3.7 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points | 69% | 4.2 / 5 | 48% | 3.0 s | 1.5 s | 100%(among the best) | 3.5 / 5 | 100%(among the best) | $0.06(among the best) |
| 11= (tied) | Realtime v1 MaxElevenLabs | 47 / 100±7.5. Task completion 56%, 0.0 of 23 points; Response time 2.6 s, 15.6 of 18 points; Instruction following 4.0 / 5, 6.7 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 40%, 0.0 of 14 points; Time to first word 0.7 s, 4.5 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points | 56% | 4.0 / 5 | 40% | 2.6 s | 0.7 s(among the best) | 76% | 3.9 / 5(among the best) | 100%(among the best) | $0.65 |
| 12= (tied) | Realtime v1 MiniElevenLabs | 45 / 100±7.1. Task completion 42%, 0.0 of 23 points; Response time 2.7 s, 15.1 of 18 points; Instruction following 3.8 / 5, 5.1 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 34%, 0.0 of 14 points; Time to first word 0.7 s, 4.5 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points | 42% | 3.8 / 5 | 34% | 2.7 s | 0.7 s(among the best) | 94% | 3.4 / 5 | 100%(among the best) | $0.64 |
| 13= (tied) | GPT Realtime 2.1 miniOpenAI | 44 / 100±6.7. Task completion 70%, 0.0 of 23 points; Response time 2.8 s, 14.5 of 18 points; Instruction following 3.6 / 5, 4.4 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 51%, 0.3 of 14 points; Time to first word 1.0 s, 4.5 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points | 70% | 3.6 / 5 | 51% | 2.8 s | 1.0 s | 84% | 3.9 / 5(among the best) | 100%(among the best) | $0.14(among the best) |
| 14 | Gemini 2.5 Flash native audioGoogle | 36 / 100±5.4. Task completion 77%, 0.0 of 23 points; Response time 4.2 s, 8.2 of 18 points; Instruction following 3.9 / 5, 6.2 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.6 of 14 points; Tool use 40%, 0.0 of 14 points; Time to first word 3.4 s, 0.9 of 5 points; Readback accuracy 84% (field median, not measured), 2.7 of 5 points; Word error rate 5.9% (field median, not measured), 2.8 of 5 points; Voice consistency 100%, 4.5 of 5 points | 77% | 3.9 / 5 | 40% | 4.2 s | 3.4 s | 100%(among the best) | 3.7 / 5 | 100%(among the best) | $0.08(among the best) |
Open a model for its full report, every figure it was scored on and the calls behind them.
Every stack
The evidence behind the model boards: each combination on its own calls, scored on the same recipe. Open a stack for its full report, every figure it was scored on and the calls behind them.
Roark score, top 15
Whiskers show the 95% interval. The shaded stacks tie for first: these calls cannot separate them.
Each score carries its 95% margin of error. A rank is the best place a stack could hold once every margin is allowed for, so stacks the calls cannot tell apart share one and are listed by the low end of their margin; hover a rank for the worst place it could hold. 85 stacks are level with the leader: their calls cannot tell them apart. Among them, 36 OpenAI GPT-6.1 Sol, 22 Anthropic Claude Sonnet 5.5, 19 Google Gemini 3.8 Flash, 4 speech-to-speech, 3 Anthropic Claude Haiku 4.5 and 1 OpenAI GPT-4o mini; 42 AssemblyAI, 20 Deepgram and 19 Speechmatics; 37 Cartesia, 36 ElevenLabs, 7 Rime and 1 OpenAI. A cascade's figures are 25% its own calls and 75% what every other stack shows about its transcriber, language model and voice. Not ranked here for missing data: Speechmatics Linden-1 + OpenAI GPT-6.1 Sol + OpenAI TTS-1 HD (no response time or tool use or readback accuracy or voice consistency).
By industry
Each industry ranked on its own calls: the same five callers, its own tasks and tools. A stack has a handful of calls in each, so a cascade is ranked on what its transcriber, language model and voice show across every stack in that industry, and a speech-to-speech model on its own calls there. Pairings show on the main board. Each row reads across: the main board's figure first, then every industry's on the same figure.
No stack matches these filters.
No stack matches
Each industry column is that industry's own board, every row it ranks. A dash is a row the industry does not rank on this figure; the arrow beside an industry opens its full board.
How it works
Every number on this page comes from real phone calls, placed and scored the same way for every stack.
Real calls, four industries, five callers
Roark places real phone calls over the public phone network to a reference agent for each of four industries: rescheduling a clinic appointment, booking an urgent home repair, opening an auto insurance claim and fixing an internet outage. Each agent has its own prompt and tools, identical for every stack. The tasks carry the details real calls trip on: a corrected date of birth, a policy number spelled out, a constraint added halfway through.
Every industry is called by the same five simulated callers: a quiet US English line, a Spanish-accented caller on a city street, an Indian English caller in a busy café, a flustered Southern caller with a baby crying, and a fast caller who interrupts over construction noise. They are equally clear about what they want, so a call fails because of the agent, not the caller. Only the models change, so any difference on the board comes from them.
Each caller rings each industry once for a cascade and four times for a speech-to-speech model. A cascade's figures also draw on its parts, heard in dozens of other stacks; a speech-to-speech model has only its own calls, so it takes more of them.
That makes industries comparable over the same callers, and callers over the same tasks, and the board can be re-ranked by either. One caller within one industry is too few calls to judge, so the board never shows it.
656
stacks
20
calls per stack
13,956
calls scored
Every model, in every combination
9 transcribers, 8 language models and 9 voices are run in every combination with each other, beside 14 speech-to-speech models. A model's own figures pool every stack it runs in, adjusted for its partners in each one, so every model is compared as if it ran with the same partners. A transcriber's figure rests on hundreds of calls with every language model and every voice, not on one lucky pairing.
That measures a model where it actually works: inside a running agent, on live phone audio, through interruptions, corrections and real turn-taking. A test that feeds a model clean audio files or one prompt at a time never sees any of that.
One score, weighted by what matters on a call
Each section ranks on a Roark score. Every figure scores 0 to 100 between two published points: full marks where doing better makes no difference to the caller, zero where the call breaks down. Each then counts at its weight, and the weights follow one order everywhere: whether the caller got what they rang for, then whether the agent got it right, then speed.
A stack, cascade or speech-to-speech, is scored on the outcome and the reply first, then on what each of its components got right on its own calls: the words its transcriber got wrong in a way that changes the meaning, whether its voice read codes and amounts back right and kept one steady voice, and whether its tools were handled. How human a voice sounds is shown on the boards and not scored. A component that leads its section lifts every stack that carries it, by the same recipe for every stack. A figure a stack cannot carry, such as a semantic word error rate for a speech-to-speech model with no transcript, is credited at the field median, so the stack is neither flattered nor punished for it.
On a full sweep each cascade's figure is 25% what its own calls measured and 75% what every other stack shows about its transcriber, language model and voice, fitted across the whole sweep. The split is the same for every stack, so a lucky run does not lead and a stack's score does not move with how noisy its calls were. In an industry a stack has a handful of calls, so a cascade's figure there is what its transcriber, language model and voice show across every stack in that industry's calls, and a speech-to-speech model's is its own calls there. How a particular pairing beats or misses its parts shows on the main board, where each stack has all of its calls. The best overall, best value and best speech-to-speech picks, and the best stack in each industry, are each the one whose score is highest at the low end of its margin of error, so a few lucky calls cannot carry a stack past one whose score rests on hundreds. Stacks that share a rank are listed by score, so the pick can sit below a higher score it cannot be told from, and its reason says how many stacks are level with it.
That order comes from real-world voice agent calls, where getting the conversation right predicts how a call ends more strongly than speed does. The latency points come from the same calls: replying faster than about 2 seconds made no measurable difference, and calls went measurably worse once the agent took more than about 3 seconds to reply.
Stacks and speech-to-speech
| Figure | Weight | Full marks | Zero |
|---|---|---|---|
Task completionscenario_adherence_check | 23% | 100% | 80% |
| An agent that completes fewer than four calls in five is not fit to deploy. On a full sweep, a cascade's figure is 25% what its own calls (20 or more) measured and 75% what every other stack shows about its transcriber, language model and voice, the same split for every stack, so a lucky run does not lead. Where the sweep is too small for that, and for a speech-to-speech model, the figure is the stack's own. In an industry a stack has a handful of calls, so a cascade's figure there is what its transcriber, language model and voice show across every stack in that industry's calls, and a speech-to-speech model's is its own calls there. How a particular pairing beats or misses its parts shows on the main board, where each stack has all of its calls. | |||
Response timeresponse_time | 18% | 2.0 s | 6.0 s |
| In real calls, replying faster than 2 s made no measurable difference, and calls went worse past 3 s. | |||
Instruction followinginstruction_follow | 14% | 5.0 / 5 | 3.0 / 5 |
| Every model scores between 3 and 5 on the judge's 1 to 5 scale, so that is the range. | |||
Semantic word error ratestt_semantic_word_error_rate | 14% | 0.5% | 3% |
| At 0.5% the best transcribers change the meaning of one word in two hundred on phone audio; at 3% one word in thirty sends the agent the wrong way. | |||
Tool usetool_use | 14% | 100% | 50% |
| An agent that mishandles its tools on half its calls cannot be trusted with the systems it runs on. | |||
Time to first wordtime_to_first_word | 5% | 1.0 s | 4.0 s |
| A greeting within 1 s sounds natural; 4 s of silence sounds like a dead line. | |||
Readback accuracytts_entity_readback_accuracy | 5% | 100% | 60% |
| Reading a code, amount or date wrong breaks the task it was for; the weakest voices get two in five wrong. | |||
Word error ratetts_word_error_rate | 5% | 2% | 12% |
| At 2% the voice says what it was given; at 12% one word in eight comes out wrong. | |||
Voice consistencyvoice_consistency_check | 5% | 100% | 80% |
| A voice that changes speaker on one call in five is not one a business can put its name to. | |||
Shown, not scored: Turn budget, Barge-in handlingbarge_in_yield_appropriateness, Human-likenessvoice_human_likeness
Counted only with price factored in: Cost per task
Language model
| Figure | Weight | Full marks | Zero |
|---|---|---|---|
Task completionscenario_adherence_check | 36% | 100% | 80% |
| An agent that completes fewer than four calls in five is not fit to deploy. | |||
Tool usetool_use | 27% | 100% | 50% |
| An agent that mishandles its tools on half its calls cannot be trusted with the systems it runs on. | |||
Instruction followinginstruction_follow | 18% | 5.0 / 5 | 3.0 / 5 |
| Every model scores between 3 and 5 on the judge's 1 to 5 scale, so that is the range. | |||
Time to first tokenllm_time_to_first_token | 18% | 0.50 s | 2.00 s |
| Under 0.5 s is not noticeable; at 2 s the first token alone takes two thirds of the 3 s past which real calls go worse. | |||
Shown, not scored: Prompt cache hitsllm_prompt_cache_hit_rate
Counted only with price factored in: Cost per call
Speech-to-Text
| Figure | Weight | Full marks | Zero |
|---|---|---|---|
Semantic word error ratestt_semantic_word_error_rate | 71% | 0.5% | 3% |
| At 0.5% the best transcribers change the meaning of one word in two hundred on phone audio; at 3% one word in thirty sends the agent the wrong way. | |||
Time to final transcriptstt_time_to_final_transcript | 29% | 1.00 s | 2.50 s |
| The fastest transcribers finish in about 1 s; past 2.5 s the transcript alone uses most of a natural reply time. | |||
Shown, not scored: Word error ratestt_word_error_rate
Counted only with price factored in: Cost per call
Text-to-Speech
| Figure | Weight | Full marks | Zero |
|---|---|---|---|
Readback accuracytts_entity_readback_accuracy | 38% | 100% | 60% |
| Reading a code, amount or date wrong breaks the task it was for; the weakest voices get two in five wrong. | |||
Word error ratetts_word_error_rate | 25% | 2% | 12% |
| At 2% the voice says what it was given; at 12% one word in eight comes out wrong. | |||
Time to first audio bytetts_time_to_first_byte | 25% | 0.20 s | 1.50 s |
| Under 0.2 s is not noticeable; 1.5 s is half the point where calls go worse. | |||
Voice consistencyvoice_consistency_check | 13% | 100% | 80% |
| A voice that changes speaker on one call in five is not one a business can put its name to. | |||
Shown, not scored: Human-likenessvoice_human_likeness
Counted only with price factored in: Cost per call
Factor in price and cost takes 25% of the score; the other weights keep their proportions. Price is scored on a log scale, so each doubling costs the same points.
A figure listed as shown, not scored is on the section’s boards and outside its score. Where a section has not collected a scored figure yet, the others share its weight.
Every call recorded and judged
Every call is recorded. Judge models score each transcript for the outcome and the conversation, and the timings come from the recording. Every figure links to the calls behind it.
Roark is on both sides of every call: it plays the caller, so it keeps the exact text its caller said, and it runs the agent under test, so it sees what the agent’s transcriber heard and what its model wrote for its voice. The transcript and voice figures come from lining those up, with no hand-made transcript in between. Each carries a margin from the synthetic voice or Roark’s transcriber in the loop. Every model faces the same ones under the same conditions, so the figures compare models on the same calls through the same voice and the same ear, though those slips need not fall evenly on each model. The rates will not match a vendor’s figure on a curated set; read them against each other.
An example: one call from this sweep
Speechmatics Linden-1 (own VAD) + OpenAI GPT-6.1 Sol + OpenAI TTS-1 HD, 30 October 2026. Highlighted turns are the ones that affected a score. Every stack's report links to its calls.
See how your own agent compares
Roark runs the same kind of calls against your agent, on your scenarios, with the same scoring.
Rank on Roark score
#333of 655 stacks
48 / 100
Leader AssemblyAI Universal-3.5 Pro + OpenAI GPT-6.1 Sol + ElevenLabs Turbo v2.5, 70 / 100
Where it ranks
Its models
Task completion
Task completion
How much of the caller’s task the agent completed: the share of the scenario’s checks a judge model found met in the transcript, averaged over calls. Before v1.3, the share of calls judged 4 or better out of 5.
#422 of 662 on the board · best 99%
Flow adherence
How closely the agent followed the intended conversation flow, judged per call. 5 = fully followed.
#100 of 662 on the board · best 5.0 / 5
Instruction following
How well the agent stuck to the instructions in its own prompt. 5 = followed throughout.
#13 of 662 on the board · best 5.0 / 5
Turn-taking
Barge-in handling
When the caller interrupts, how often the agent made the right call: yielding to a real correction, and holding the floor through a filler.
Not scored by this sweep.
Incorrect interruptions
How often the agent interrupted the caller when it should have kept listening.
#1 of 662 on the board · best 0%
Overtalk
Share of the conversation where both sides were speaking over each other.
#334 of 662 on the board · best 0.0%
Agent self-cutoffs
Times the agent cut its own sentence off mid-word.
#1 of 662 on the board · best 0.0 per call
Understanding & effort
Caller effort
How hard the caller had to work: repeating themselves, correcting, re-explaining. 1 = effortless.
Not scored by this sweep.
Repeated-info loops
Times per call the agent re-asked for something the caller had already given.
#1 of 494 on the board · best 0.0 per call
Speech-to-Text
Semantic word error rate
The caller's words the agent's transcriber got wrong in a way that changes the meaning, as a judge reads each difference.
#164 of 648 on the board · best 0.0%
Word error rate
The caller's words the agent's transcriber missed, changed or invented, against what the caller actually said.
#242 of 648 on the board · best 1.2%
Text-to-Speech
Readback accuracy
Share of the codes, amounts, dates and phone numbers the agent wrote that its voice read out correctly. One wrong digit is a miss.
#5 of 603 on the board · best 100%
Word error rate
The share of the words the agent wrote that its voice dropped or changed, read back from Roark’s transcription of the recording; the same transcriber hears every voice.
#398 of 648 on the board · best 1.9%
Voice
Human-likeness
How human the agent's voice sounds, scored by an audio model on intonation, rhythm and expressiveness. 5 = hard to tell from a person.
#246 of 660 on the board · best 4.3 / 5
Voice consistency
Share of calls where every line the agent spoke was the same speaker, with no change of speaker, gender or language voice partway. It checks who is speaking, not how they sound.
Not scored by this sweep.
#1 of 660 on the board · best 100%
Latency
Response time
The silence between the caller finishing and the agent starting to speak, measured from the recording on every turn of the call.
#574 of 661 on the board · best 1.85 s
Time to first word
How long the caller waits before the agent first speaks after pickup.
#649 of 662 on the board · best 0.45 s
Latency spikes
Turns where the response took far longer than the rest of the conversation.
#225 of 662 on the board · best 0.0 per call
By industry
| Industry | Task completion | Latency P50 | Barge-in handling |
|---|---|---|---|
| Healthcare | 88% | 5.21 s | — |
| Home services | 83% | 5.05 s | — |
| Insurance | 98% | 4.75 s | — |
| Customer support | 93% | 4.72 s | — |
By caller
| Caller | Task completion | Latency P50 | Barge-in handling |
|---|---|---|---|
| US English, quiet line | 92% | 4.38 s | — |
| Spanish-accented English, city street | 91% | 4.86 s | — |
| Indian English, busy café | 91% | 4.74 s | — |
| Southern US, flustered, crying baby | 88% | 5.01 s | — |
| Fast and interrupting, construction outside | 89% | 5.38 s | — |
The calls behind the numbers
Customer support
AGENTBrightwave Internet support, this is Alex. What’s going on with your service?
CUSTOMERYeah, listen, my internet's been down for 2 hours. I work from home and there's a road crew digging up the street right outside. I already rebooted the gateway twice and checked every cable. The internet light is blinking orange. Can you check if the road crew's done yet? roadwork is the cause and get this fixed. I really don't have time, so can we do this quickly?
AGENTSure thing.
CUSTOMERGreat! My account number is BW883017. Sorry, last digit is 1. Can you look up what's going on?
AGENTI've sent a remote reset to your gateway, and it's now registered and online. Your devices should reconnect within a couple of minutes.
CUSTOMERI see the internet light just went solid white. How long until my devices reconnect? And I need a credit for the downtime.
AGENTYour devices should reconnect within about two minutes. I'll go ahead and apply a service credit for the downtime. One moment, please.
AGENTI'm not hearing anything from you, so I'll let you go now. Goodbye, and have a great day!
Healthcare
AGENTRiverbend Health Clinic, this is Alex. How can I help you today?
CUSTOMERHello, I'm having one dermatology appointment next Wednesday morning, but my work shift has changed. I need to move the appointment, kindly. My name is Arjun Mehta. Can you help with that?
AGENTI'd be happy to help you reschedule your dermatology appointment.
AGENTCould you please provide your date of birth so I can verify your information?
CUSTOMERMarch 17th, 1948. Sorry, no, 1984. March 17th, 1984.
AGENTYou have a dermatology appointment with Dr. Okafor next Wednesday at 09:15 AM at the Main Campus, Building B.
AGENTLet's look for a new time that works for you. Do you have any specific days or times in mind for rescheduling?
CUSTOMERYes, I can only come after 3 PM. What times do you have available in the afternoon?
AGENTDr. Okafor has the following afternoon slots available:
- Tuesday at 1545
- Thursday at 1610
Would either of these work for you?
CUSTOMERTuesdays do not work after all because I pick up my daughter. What about Friday? Is there any slot available then?
AGENTOn Friday, Dr. Okafor has an available slot at 11:30 AM. Unfortunately, that's before 3 PM. Would you like to take it, or should we stick with the Thursday time at 16:10 instead?
CUSTOMERLet's go with Thursday at 4:10 PM. That works for me.
AGENTYour appointment has been rescheduled to Thursday at 4:10 PM. Your new confirmation code is RVB-3M8P2. Please arrive 15 minutes early with a photo ID and your insurance card.
CUSTOMEROkay, I have a few quick questions. First, is there any fee for changing the appointment?
AGENTChanges made more than 24 hours before the appointment are free, so there will be no fee for your rescheduled appointment. Is there anything else you'd like to know?
CUSTOMERYes, where should I park when I come for the appointment?
AGENTYou can park in Lot C, which offers free patient parking. The entrance is on River Road. Anything else you need help with?
CUSTOMEROne last thing. Does Dr. Okafor accept Aetna insurance?
AGENTI don't have access to specific insurance details for individual providers. I recommend contacting our billing department directly to confirm whether Dr. Okafor accepts Aetna insurance. Is there anything else I can assist you with?
CUSTOMERNo, that covers everything. Thank you for your help. Goodbye.
AGENTYou're welcome! If you have any more questions in the future, feel free to reach out. Goodbye, and have a great day!
Home services
AGENTCopperline Home Services, this is Alex. What can I help you with?
CUSTOMERHello, yes, I'm having an emergency here. A pipe under my kitchen sink has burst and water is spreading across the floor heading toward a power strip. Can you send someone right away?
AGENTFirst, let's make sure everyone is safe. Please immediately shut off the water main to stop the flooding if you haven't already.
CUSTOMEROkay, I will. The water is off now. It has stopped spreading. What should I do next?
AGENTGreat job on shutting off the water.
CUSTOMERRight.
CUSTOMERAha, what else do you need me to check?
AGENTPlease make sure to stay clear of any electrical outlets or power strips that are near the water. Do not touch anything electrical until it's safe and dry. Now, let's get you some help. Could you please provide your phone number or address so I can find your customer information?
CUSTOMERMy address is 1480 Larkspur Avenue. I moved last month, so I'm in unit 4D now, not 4B. My phone number is 503- 555-0139— no, 0193.
AGENTI've booked an emergency visit for you today between 2:00 PM and 4:00 PM. As a Copperline Care member, your emergency callout fee is $99. Technician Dana will be coming, and you'll receive a text update 30 minutes before they arrive. Here's what I have for your service details:
- Service Address1480 Larkspur Avenue, Unit 4D
- Callback Number503-555-0193
- Ticket NumberCPL-58210
If anything changes or if you have further questions, don't hesitate to call us back. Stay safe and dry until Dana arrives!
Insurance
AGENTNorthgate Mutual claims, this is Alex. Are you somewhere safe right now?
CUSTOMERYes, I'm somewhere safe. I'm just standing next to my car.
AGENTGreat to hear you're in a safe place.
CUSTOMEROkay.
CUSTOMERNo, nobody is hurt. I'm fine.
AGENTThat's a relief.
CUSTOMERYeah, that's right. So I need to file a claim. Another driver hit the back of my car while I was stopped at a red light about 20 minutes ago. The bumper is crushed and And the rear wheel is rubbing, so I can't drive it. Police didn't come. My policy number is NG4471903. No, sorry, last one is 2. The other driver is Kevin Lund, insured with Pacific Shore. His plate is 7KX, like x-ray, R214. Can you tell me how much I have to pay for this?
AGENTI've filed your claim, and here's what you need to know:
- Claim NumberNGM-904417
- AdjusterPriya Shah will contact you within one business day. For your coverage:
- Collision Deductible$500
- Rental$40 per day for up to 30 days, starting today
- TowingCovered up to 25 miles
A tow truck from Eastside Towing is on its way and should arrive in about 45 minutes. They'll take your car to a Northgate partner repair shop at 9th and Morrison. Your policy number is NG4471902, and the other driver's plate is 7KXR214. If there's anything else you need, just let me know!
CUSTOMEROkay, got it. But he hit me, so I should not have to pay that deductible, right? And do you know if my premium is going to go up because of this?
AGENTI understand your concern.
CUSTOMERRight, and I
AGENTI can't say for certain how the deductible or your premium might be affected, as it depends on the outcome of the claim investigation and your specific policy terms. Generally, if you're not at fault, you might not have to pay the deductible, and your premium might not increase. However, these determinations are made after the claim is reviewed.
AGENTYour adjuster, Priya Shah, will be able to provide more specific information once they assess the details of the accident.