Skip to content

Roark Live Bench · Suite v1.3 · Sweep of 15 November 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.

1176 stacks published23,380 scored callsPublished at 20 scored calls

Our picks

The read

  • 496 of 890 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 1.2 s to 6.2 s on the same agent and the same caller, a spread of 5.0 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 + AssemblyAI TTS-1 at 1.2 s, ranks 832 of 890 on task completion; the most complete is Speechmatics Linden-1 (own VAD) + OpenAI GPT-6.1 Sol + OpenAI TTS-1 HD at 99%.

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 all 6 measures compared here.
  • Speech-to-speech costs 3.7 times as much per completed task: $0.24 for speech-to-speech against $0.07 for cascades.
CascadeSpeech-to-speech
Task completionTool useTime to first wordResponse timeBarge-in handlingCost per task255075100
MeasureCascadeSpeech-to-speech
Task completion
92% (leads)85%
Tool use
87% (leads)78%
Time to first word
0.6 s (leads)1.1 s
Response time
2.1 s (leads)2.1 s
Barge-in handling
100% (leads)96%
Cost per task
$0.07 (leads)$0.24

The three strongest of each, averaged. Cascades: Deepgram Flux + OpenAI GPT-6.1 Terra + AssemblyAI TTS-1; Deepgram Flux + Meta Muse Ember 1.0 + AssemblyAI TTS-1; Speechmatics Linden-1 (own VAD) + OpenAI GPT-6.1 Terra + AssemblyAI TTS-1. 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.

  • Pick Soniox v6: it leads on semantic word error rate (0.2%). If speed matters more, Deepgram Flux is the quickest, at 0.38 s.
  • Deepgram Flux delivers its transcript in 0.38 s, OpenAI Whisper-1 in 1.93 s: 1.54 s added to every turn of the call.
  • Soniox v6 is the most accurate: it changes the meaning of 0.2% of the caller's words, against 2.1% for Speechmatics Linden-1 (own VAD).
  • In 574 of 1,161 cascades the transcriber is the slowest step of the turn, whichever model runs it.

Roark score, best first

100Soniox v691Universal-3.6 Pro83Nova-382Universal-3.5 Pro72Enhanced (realtime)70Flux68Nova-268Linden-167Scribe v3 Live64Standard (realtime)56Linden-1 (own VAD)54Whisper-150GPT Transcribe
Every transcriber on every figure, sorted by roark score. Column headers sort the table.
#
1Soniox v6Soniox100 / 100. Semantic word error rate 0.2%, 71.4 of 71 points; Time to final transcript 0.85 s, 28.6 of 29 points(among the best)0.2%(among the best)0.85 s(among the best)$0.008(among the best)
2Universal-3.6 ProAssemblyAI91 / 100. Semantic word error rate 0.8%, 62.9 of 71 points; Time to final transcript 0.90 s, 28.6 of 29 points(among the best)0.8%(among the best)0.90 s$0.029
3Nova-3Deepgram83 / 100. Semantic word error rate 1.1%, 54.3 of 71 points; Time to final transcript 0.88 s, 28.6 of 29 points(among the best)1.1%(among the best)0.88 s$0.031
4Universal-3.5 ProAssemblyAI82 / 100. Semantic word error rate 1.1%, 53.3 of 71 points; Time to final transcript 0.90 s, 28.6 of 29 points1.1%(among the best)0.90 s$0.030
5Enhanced (realtime)Speechmatics72 / 100. Semantic word error rate 1.5%, 43.9 of 71 points; Time to final transcript 0.98 s, 28.6 of 29 points1.5%0.98 s$0.054
6FluxDeepgram70 / 100. Semantic word error rate 1.5%, 41.9 of 71 points; Time to final transcript 0.38 s, 28.6 of 29 points1.5%0.38 s(among the best)$0.031
7Nova-2Deepgram68 / 100. Semantic word error rate 1.6%, 39.7 of 71 points; Time to final transcript 0.89 s, 28.6 of 29 points1.6%0.89 s$0.023
8Linden-1Speechmatics68 / 100. Semantic word error rate 1.6%, 39.7 of 71 points; Time to final transcript 0.98 s, 28.6 of 29 points1.6%0.98 s$0.020
9Scribe v3 LiveElevenLabs67 / 100. Semantic word error rate 1.7%, 38.0 of 71 points; Time to final transcript 0.88 s, 28.6 of 29 points1.7%0.88 s$0.015(among the best)
10Standard (realtime)Speechmatics64 / 100. Semantic word error rate 1.8%, 35.1 of 71 points; Time to final transcript 0.96 s, 28.6 of 29 points1.8%0.96 s$0.031
11Linden-1 (own VAD)Speechmatics56 / 100. Semantic word error rate 2.1%, 27.1 of 71 points; Time to final transcript 0.46 s, 28.6 of 29 points2.1%0.46 s(among the best)$0.020
12Whisper-1OpenAI54 / 100. Semantic word error rate 1.5%, 43.0 of 71 points; Time to final transcript 1.93 s, 10.9 of 29 points1.5%1.93 s$0.025
13GPT TranscribeOpenAI50 / 100. Semantic word error rate 1.9%, 30.6 of 71 points; Time to final transcript 1.47 s, 19.5 of 29 points1.9%1.47 s$0.018(among the best)

Open a model to see where it ranks and the stacks it works best in.

Cost against qualitySemantic word error rate · Cost per call
$0$0.02$0.04$0.06$0.080%1%2%3%Cheap and accurateAccurate, pricierCheap, less accuratePricier and less accurateSoniox v6Scribe v3 LiveGPT TranscribeLinden-1 (own VAD)Linden-1Nova-2Whisper-1Universal-3.6 ProUniversal-3.5 ProStandard (realtime)FluxNova-3Enhanced (realtime)Cost per call · ← betterSemantic word error rate · better →
Speed against qualitySemantic word error rate · Time to final transcript
0 s1 s2 s3 s0%1%2%3%Fast and accurateAccurate, slowerFast, less accurateSlower and less accurateFluxLinden-1 (own VAD)Soniox v6Scribe v3 LiveNova-3Nova-2Universal-3.5 ProUniversal-3.6 ProStandard (realtime)Linden-1Enhanced (realtime)GPT TranscribeWhisper-1Time to final transcript · ← betterSemantic word error rate · better →

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.

  • OpenAI GPT-6.1 Terra and OpenAI GPT-6.1 Sol are level: neither leads by more than the margin of error. Of the two, OpenAI GPT-6.1 Terra costs least, at $0.013 a call. Avoid OpenAI GPT-5.1: last on tool use (42%) and instruction following (4.1 / 5).
  • Anthropic Claude Haiku 4.5 starts answering in 0.36 s, Google Gemini 3.8 Flash in 1.50 s: 1.15 s added to every turn of the call.
  • OpenAI GPT-6.1 Terra uses its tools most reliably: it handles its tools without a mistake on 86% of calls, against 42% for OpenAI GPT-5.1.
  • In 339 of 1,161 cascades the language model is the slowest step of the turn, whichever model runs it.

Roark score, best first

64GPT-6.1 Terra63GPT-6.1 Sol57Claude Sonnet 5.557Muse Ember 1.056Claude Haiku 5.549Gemini 3.8 Flash48GPT-4.145Claude Haiku 4.543GPT-4o mini42GPT-4o41GPT-5.1

Not ranked here for missing data: Groq GPT-OSS 120B (no task completion or tool use or instruction following or time to first token).

Every language model on every figure, sorted by roark score. Column headers sort the table.
#
1GPT-6.1 TerraOpenAI64 / 100. Task completion 91%, 19.2 of 36 points; Tool use 86%, 19.6 of 27 points; Instruction following 4.3 / 5, 11.5 of 18 points; Time to first token 0.83 s, 14.2 of 18 points(among the best)91%(among the best)4.3 / 5(among the best)86%(among the best)0.83 s67%$0.013
2GPT-6.1 SolOpenAI63 / 100. Task completion 92%, 22.2 of 36 points; Tool use 82%, 17.4 of 27 points; Instruction following 4.3 / 5, 11.7 of 18 points; Time to first token 1.00 s, 12.2 of 18 points(among the best)92%(among the best)4.3 / 5(among the best)82%(among the best)1.00 s72%$0.019
3Claude Sonnet 5.5Anthropic57 / 100. Task completion 91%, 20.2 of 36 points; Tool use 75%, 13.5 of 27 points; Instruction following 4.2 / 5, 11.0 of 18 points; Time to first token 0.94 s, 12.8 of 18 points(among the best)91%(among the best)4.2 / 575%(among the best)0.94 s–$0.115
4Muse Ember 1.0Meta57 / 100. Task completion 90%, 18.4 of 36 points; Tool use 66%, 8.9 of 27 points; Instruction following 4.3 / 5, 11.8 of 18 points; Time to first token 0.56 s, 17.5 of 18 points(among the best)90%4.3 / 5(among the best)66%0.56 s85%(among the best)$0.005(among the best)
5Claude Haiku 5.5Anthropic56 / 100. Task completion 90%, 17.5 of 36 points; Tool use 68%, 10.0 of 27 points; Instruction following 4.3 / 5, 12.2 of 18 points; Time to first token 0.63 s, 16.5 of 18 points90%4.3 / 5(among the best)68%0.63 s–$0.008(among the best)
6Gemini 3.8 FlashGoogle49 / 100. Task completion 92%, 21.5 of 36 points; Tool use 69%, 10.2 of 27 points; Instruction following 4.3 / 5, 11.5 of 18 points; Time to first token 1.50 s, 6.0 of 18 points92%(among the best)4.3 / 5(among the best)69%1.50 s–$0.015
7GPT-4.1OpenAI48 / 100. Task completion 90%, 17.3 of 36 points; Tool use 52%, 1.1 of 27 points; Instruction following 4.2 / 5, 11.1 of 18 points; Time to first token 0.52 s, 18.0 of 18 points90%4.2 / 552%0.52 s82%(among the best)$0.016
8Claude Haiku 4.5Anthropic45 / 100. Task completion 88%, 14.4 of 36 points; Tool use 51%, 0.8 of 27 points; Instruction following 4.3 / 5, 11.4 of 18 points; Time to first token 0.36 s, 18.2 of 18 points88%4.3 / 5(among the best)51%0.36 s(among the best)–$0.053
9GPT-4o miniOpenAI43 / 100. Task completion 86%, 10.3 of 36 points; Tool use 56%, 3.1 of 27 points; Instruction following 4.2 / 5, 11.1 of 18 points; Time to first token 0.50 s, 18.2 of 18 points86%4.2 / 556%0.50 s(among the best)82%(among the best)$0.001(among the best)
10GPT-4oOpenAI42 / 100. Task completion 86%, 11.2 of 36 points; Tool use 53%, 1.8 of 27 points; Instruction following 4.2 / 5, 11.2 of 18 points; Time to first token 0.45 s, 18.2 of 18 points86%4.2 / 553%0.45 s(among the best)–$0.022
11GPT-5.1OpenAI41 / 100. Task completion 87%, 13.1 of 36 points; Tool use 42%, 0.0 of 27 points; Instruction following 4.1 / 5, 10.4 of 18 points; Time to first token 0.59 s, 17.1 of 18 points87%4.1 / 542%0.59 s74%$0.026

Open a model to see where it ranks and the stacks it works best in.

Cost against qualityTask completion · Cost per call
$0$0.05$0.1$0.1580%85%90%95%Cheap and reliableReliable, pricierCheap, less reliablePricier and less reliableGPT-4o miniMuse Ember 1.0Claude Haiku 5.5GPT-6.1 TerraGemini 3.8 FlashGPT-4.1GPT-6.1 SolGPT-4oGPT-5.1Claude Haiku 4.5Claude Sonnet 5.5Cost per call · ← betterTask completion · better →
Speed against qualityTask completion · Time to first token
0 s0.5 s1 s1.5 s2 s80%85%90%95%Fast and reliableReliable, slowerFast, less reliableSlower and less reliableClaude Haiku 4.5GPT-4oGPT-4o miniGPT-4.1Muse Ember 1.0GPT-5.1Claude Haiku 5.5GPT-6.1 TerraClaude Sonnet 5.5GPT-6.1 SolGemini 3.8 FlashTime to first token · ← betterTask completion · better →

Text-to-Speech

The speech engine is the voice the caller hears, and the last step before they hear anything.

  • Deepgram Flux TTS, AssemblyAI TTS-1, Cartesia Sonic-3.6 and Cartesia Sonic-3.5 are level: none leads by more than the margin of error. Of them, AssemblyAI TTS-1 costs least, at $0.054 a call. Avoid OpenAI TTS-1 HD: last on readback accuracy (81%), word error rate (7.6%) and time to first audio byte (2.09 s).
  • AssemblyAI TTS-1 starts speaking in 0.04 s, OpenAI TTS-1 HD in 2.09 s: 2.04 s added to every turn of the call.
  • AssemblyAI TTS-1 is also the most accurate: it reads codes, amounts and dates back correctly 90% of the time, against 81% for OpenAI TTS-1 HD.
  • Rime Coda sounds the most human: 3.5 / 5, against 3.2 / 5 for ElevenLabs Turbo v2.5.
  • In 119 of 1,161 cascades the speech engine is the slowest step of the turn, most often with OpenAI TTS-1 (65 of them).

Roark score, best first

83Flux TTS82TTS-182Sonic-3.682Sonic-3.582v4 Turbo78Sonic-375Turbo v2.574Flash v2.573Coda57TTS-143TTS-1 HD
Every speech engine on every figure, sorted by roark score. Column headers sort the table.
#
1Flux TTSDeepgram83 / 100. Readback accuracy 89%, 27.4 of 38 points; Word error rate 4.8%, 17.9 of 25 points; Time to first audio byte 0.13 s, 25.0 of 25 points; Voice consistency 100%, 12.2 of 13 points(among the best)4.8%(among the best)89%(among the best)3.2 / 5100%(among the best)0.13 s$0.119
2TTS-1AssemblyAI82 / 100. Readback accuracy 90%, 28.0 of 38 points; Word error rate 5.3%, 16.8 of 25 points; Time to first audio byte 0.04 s, 25.0 of 25 points; Voice consistency 100%, 12.3 of 13 points(among the best)5.3%90%(among the best)3.3 / 5100%(among the best)0.04 s(among the best)$0.054(among the best)
3Sonic-3.6Cartesia82 / 100. Readback accuracy 89%, 27.2 of 38 points; Word error rate 5.1%, 17.3 of 25 points; Time to first audio byte 0.12 s, 25.0 of 25 points; Voice consistency 100%, 12.2 of 13 points(among the best)5.1%89%(among the best)3.5 / 5(among the best)100%(among the best)0.12 s(among the best)$0.160
4Sonic-3.5Cartesia82 / 100. Readback accuracy 88%, 26.5 of 38 points; Word error rate 5%, 17.6 of 25 points; Time to first audio byte 0.12 s, 25.0 of 25 points; Voice consistency 100%, 12.5 of 13 points(among the best)5%88%3.4 / 5(among the best)100%(among the best)0.12 s(among the best)$0.157
5v4 TurboElevenLabs82 / 100. Readback accuracy 88%, 26.6 of 38 points; Word error rate 4.7%, 18.2 of 25 points; Time to first audio byte 0.11 s, 25.0 of 25 points; Voice consistency 99%, 11.7 of 13 points(among the best)4.7%(among the best)88%3.4 / 5(among the best)99%0.11 s(among the best)$0.455
6Sonic-3Cartesia78 / 100. Readback accuracy 87%, 25.2 of 38 points; Word error rate 5.7%, 15.8 of 25 points; Time to first audio byte 0.16 s, 25.0 of 25 points; Voice consistency 100%, 12.2 of 13 points5.7%87%3.2 / 5100%(among the best)0.16 s$0.153
7Turbo v2.5ElevenLabs75 / 100. Readback accuracy 81%, 20.1 of 38 points; Word error rate 4.7%, 18.2 of 25 points; Time to first audio byte 0.15 s, 25.0 of 25 points; Voice consistency 99%, 11.7 of 13 points4.7%(among the best)81%3.2 / 599%0.15 s$0.439
8Flash v2.5ElevenLabs74 / 100. Readback accuracy 81%, 19.9 of 38 points; Word error rate 5%, 17.5 of 25 points; Time to first audio byte 0.12 s, 25.0 of 25 points; Voice consistency 99%, 11.9 of 13 points5%81%3.3 / 599%0.12 s(among the best)$0.408
9CodaRime73 / 100. Readback accuracy 84%, 22.5 of 38 points; Word error rate 6.8%, 13.1 of 25 points; Time to first audio byte 0.16 s, 25.0 of 25 points; Voice consistency 99%, 12.1 of 13 points6.8%84%3.5 / 5(among the best)99%0.16 s$0.111
10TTS-1OpenAI57 / 100. Readback accuracy 83%, 21.1 of 38 points; Word error rate 5.5%, 16.2 of 25 points; Time to first audio byte 1.12 s, 7.3 of 25 points; Voice consistency 100%, 12.5 of 13 points5.5%83%3.4 / 5(among the best)100%(among the best)1.12 s$0.032(among the best)
11TTS-1 HDOpenAI43 / 100. Readback accuracy 81%, 19.5 of 38 points; Word error rate 7.6%, 10.9 of 25 points; Time to first audio byte 2.09 s, 0.0 of 25 points; Voice consistency 100%, 12.5 of 13 points7.6%81%3.4 / 5(among the best)100%(among the best)2.09 s$0.054(among the best)

Open a model to see where it ranks and the stacks it works best in.

Cost against qualityReadback accuracy · Cost per call
$0$0.2$0.4$0.675%80%85%90%95%Cheap and accurateAccurate, pricierCheap, less accuratePricier and less accurateTTS-1TTS-1TTS-1 HDCodaFlux TTSSonic-3Sonic-3.5Sonic-3.6Flash v2.5Turbo v2.5v4 TurboCost per call · ← betterReadback accuracy · better →
Speed against qualityReadback accuracy · Time to first audio byte
0 s1 s2 s3 s75%80%85%90%95%Fast and accurateAccurate, slowerFast, less accurateSlower and less accurateTTS-1v4 TurboSonic-3.6Flash v2.5Sonic-3.5Flux TTSTurbo v2.5Sonic-3CodaTTS-1TTS-1 HDTime to first audio byte · ← betterReadback accuracy · better →

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%.
  • Phonic 1 answers in 1.5 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

66Grok Voice 3.066GPT Live 1 (Sol)65GPT Live 1 (Luna)65Grok Voice Think Fast 2.060Phonic 153GPT Realtime 2.153Gemini 3.8 Live48GPT Realtime 248Realtime v1 Max49Ultravox v0.748Gemini 3.1 Flash Live (preview)46Nova 2 Sonic46Realtime v1 Mini45GPT Realtime 2.1 mini36Gemini 2.5 Flash native audio

Whiskers show the 95% interval. The shaded models tie for first: these calls cannot separate them.

Every speech-to-speech model on every figure, sorted by roark score. Column headers sort the table.
#
1= (tied)Grok Voice 3.0xAI66 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 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 s94%3.0 / 5100%(among the best)$0.26
2= (tied)GPT Live 1 (Sol)OpenAI66 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points(among the best)85%(among the best)4.2 / 587%(among the best)2.3 s1.4 s77%3.1 / 5100%(among the best)$0.24
3= (tied)GPT Live 1 (Luna)OpenAI65 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points(among the best)85%(among the best)4.2 / 586%(among the best)2.3 s1.5 s93%3.1 / 5100%(among the best)$0.21
4= (tied)Grok Voice Think Fast 2.0xAI65 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 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 s100%(among the best)3.6 / 5100%(among the best)$0.26
5= (tied)Phonic 1Phonic60 / 100±5.4. Task completion 86%, 6.4 of 23 points; Response time 1.5 s, 18.2 of 18 points; Instruction following 4.4 / 5, 9.2 of 14 points; Semantic word error rate 1.1% (field median, not measured), 10.2 of 14 points; Tool use 59%, 2.4 of 14 points; Time to first word 2.1 s, 2.9 of 5 points; Readback accuracy 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points86%(among the best)4.4 / 5(among the best)59%1.5 s(among the best)2.1 s–3.2 / 5100%(among the best)$0.28
6= (tied)GPT Realtime 2.1OpenAI53 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 99%, 4.3 of 5 points78%3.8 / 575%2.6 s0.8 s(among the best)83%3.9 / 5(among the best)99%$0.43
7= (tied)Gemini 3.8 LiveGoogle53 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points78%4.3 / 569%2.8 s1.6 s100%(among the best)4.1 / 5(among the best)100%(among the best)$0.14(among the best)
8= (tied)GPT Realtime 2OpenAI48 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points76%3.8 / 564%2.9 s0.8 s(among the best)91%3.7 / 5100%(among the best)$0.69
9= (tied)Realtime v1 MaxElevenLabs48 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points56%4.0 / 540%2.6 s0.7 s(among the best)76%3.9 / 5(among the best)100%(among the best)$0.65
10= (tied)Ultravox v0.7Ultravox49 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points79%3.7 / 561%2.4 s1.6 s94%3.6 / 5100%(among the best)$0.22
11= (tied)Gemini 3.1 Flash Live (preview)Google48 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points80%4.2 / 538%2.7 s1.5 s100%(among the best)4.0 / 5(among the best)100%(among the best)$0.15
12= (tied)Nova 2 SonicAmazon46 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points69%4.2 / 548%3.0 s1.5 s100%(among the best)3.5 / 5100%(among the best)$0.06(among the best)
13= (tied)Realtime v1 MiniElevenLabs46 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points42%3.8 / 534%2.7 s0.7 s(among the best)94%3.4 / 5100%(among the best)$0.64
14= (tied)GPT Realtime 2.1 miniOpenAI45 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points70%3.6 / 551%2.8 s1.0 s84%3.9 / 5(among the best)100%(among the best)$0.14(among the best)
15Gemini 2.5 Flash native audioGoogle36 / 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), 10.2 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 86% (field median, not measured), 3.0 of 5 points; Word error rate 5.3% (field median, not measured), 3.1 of 5 points; Voice consistency 100%, 4.5 of 5 points77%3.9 / 540%4.2 s3.4 s100%(among the best)3.7 / 5100%(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.

Cost against qualityTask completion · Cost per task
$0$0.5$120%40%60%80%100%Cheap and reliableReliable, pricierCheap, less reliablePricier and less reliableNova 2 SonicGemini 2.5 Flash native audioGPT Realtime 2.1 miniGemini 3.8 LiveGemini 3.1 Flash Live (preview)GPT Live 1 (Luna)Ultravox v0.7GPT Live 1 (Sol)Grok Voice 3.0Grok Voice Think Fast 2.0Phonic 1GPT Realtime 2.1Realtime v1 MiniRealtime v1 MaxGPT Realtime 2Cost per task · ← betterTask completion · better →
Speed against qualityTask completion · Response time
0 s1 s2 s3 s4 s5 s20%40%60%80%100%Fast and reliableReliable, slowerFast, less reliableSlower and less reliablePhonic 1Grok Voice 3.0Grok Voice Think Fast 2.0GPT Live 1 (Luna)GPT Live 1 (Sol)Ultravox v0.7Realtime v1 MaxGPT Realtime 2.1Gemini 3.1 Flash Live (preview)Realtime v1 MiniGemini 3.8 LiveGPT Realtime 2.1 miniGPT Realtime 2Nova 2 SonicGemini 2.5 Flash native audioResponse time · ← betterTask completion · better →

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 score; hover a rank for the worst place it could hold. 36 stacks are level with the leader: their calls cannot tell them apart. Among them, 13 OpenAI GPT-6.1 Sol, 8 Meta Muse Ember 1.0, 7 OpenAI GPT-6.1 Terra, 4 Anthropic Claude Haiku 5.5 and 4 Anthropic Claude Sonnet 5.5; 19 Soniox, 14 AssemblyAI and 3 Deepgram; 14 Cartesia, 12 ElevenLabs, 7 AssemblyAI, 2 Rime and 1 Deepgram. 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: Soniox v6 + OpenAI GPT-6.1 Sol + Cartesia Sonic-3 (no response time or readback accuracy or voice consistency); Speechmatics Linden-1 + OpenAI GPT-6.1 Sol + OpenAI TTS-1 HD (no response time or tool use or readback accuracy or voice consistency).

Loading every stack

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.

01

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.

905

stacks

20

calls per stack

18,950

calls scored

02

Every model, in every combination

13 transcribers, 11 language models and 11 voices are run in every combination with each other, beside 15 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.

03

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

FigureWeightFull marksZero
Task completionscenario_adherence_check23%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_time18%2.0 s6.0 s
In real calls, replying faster than 2 s made no measurable difference, and calls went worse past 3 s.
Instruction followinginstruction_follow14%5.0 / 53.0 / 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_rate14%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_use14%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_word5%1.0 s4.0 s
A greeting within 1 s sounds natural; 4 s of silence sounds like a dead line.
Readback accuracytts_entity_readback_accuracy5%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_rate5%2%12%
At 2% the voice says what it was given; at 12% one word in eight comes out wrong.
Voice consistencyvoice_consistency_check5%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

FigureWeightFull marksZero
Task completionscenario_adherence_check36%100%80%
An agent that completes fewer than four calls in five is not fit to deploy.
Tool usetool_use27%100%50%
An agent that mishandles its tools on half its calls cannot be trusted with the systems it runs on.
Instruction followinginstruction_follow18%5.0 / 53.0 / 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_token18%0.50 s2.00 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

FigureWeightFull marksZero
Semantic word error ratestt_semantic_word_error_rate71%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_transcript29%1.00 s2.50 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

FigureWeightFull marksZero
Readback accuracytts_entity_readback_accuracy38%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_rate25%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_byte25%0.20 s1.50 s
Under 0.2 s is not noticeable; 1.5 s is half the point where calls go worse.
Voice consistencyvoice_consistency_check13%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.

04

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, 15 November 2026. Highlighted turns are the ones that affected a score. Every stack's report links to its calls.

Transcript, verbatim.

Read the full methodology

See how your own agent compares

Roark runs the same kind of calls against your agent, on your scenarios, with the same scoring.

Test your agent

OpenAI GPT-6.1 Terra

Language model81 stacks

Rank on Roark score

#1of 11 language models

64 / 100

The leader

Where it ranks

Roark score with price73 / 1001 of 11Leads
Task completion91%4 of 11Leader OpenAI GPT-6.1 Sol, 92%; behind OpenAI GPT-6.1 Sol
Instruction following4.3 / 54 of 11Leader Anthropic Claude Haiku 5.5, 4.3 / 5; behind Anthropic Claude Haiku 5.5
Tool use86%1 of 11Leads
Time to first token0.83 s8 of 11Leader Anthropic Claude Haiku 4.5, 0.36 s; behind Anthropic Claude Haiku 4.5
Prompt cache hits67%6 of 6Leader Meta Muse Ember 1.0, 85%
Cost per call$0.0134 of 11Leader OpenAI GPT-4o mini, $0.001

Every language model, on roark score

Works best with

Stacks running OpenAI GPT-6.1 Terra, by task completion

Soniox v6 + OpenAI GPT-6.1 Terra + Cartesia Sonic-3.593%
Soniox v6 + OpenAI GPT-6.1 Terra + OpenAI TTS-192%
Deepgram Nova-3 + OpenAI GPT-6.1 Terra + Cartesia Sonic-3.592%
Speechmatics Linden-1 + OpenAI GPT-6.1 Terra + Cartesia Sonic-3 weakest

Column: task completion. Every stack here holds OpenAI GPT-6.1 Terra in the language model slot.

Every stack it runs in (81)

All language models · How these numbers are measured