Catalog
Voice AI models and benchmarks
Every model Speko has measured, and the 73 of 73 catalog entries a request can pin today. Accuracy, latency and price as each board published them — nothing on this page is estimated to fill a column.
Provider
66 models
Speech-to-text19
#ModelProviderId
1PulseSmallest AI5.1%8.0%178ms~$0.0050
2Nova-3Deepgram12.0%12.9%106ms$0.0048
3FluxDeepgramnot measured6.6%406ms$0.0065
4Ink-2Cartesia11.0%9.9%102ms$0.0090
5Qwen3-ASRAlibaba2.8%4.0%424ms$0.0054
6Chirp 3Google3.9%7.4%581ms$0.0160
7Scribe v2 RealtimeElevenLabsnot measured3.4%233ms$0.0065
8Realtime STT-1Inworld3.3%3.6%139ms$0.0025
9Universal-3.5 ProAssemblyAI2.0%2.0%66ms$0.0075
10Velma 2Modulate4.4%5.4%1.11s$0.0010
11stt-rt-v5Soniox7.5%7.3%78ms$0.0020
12GPT-4o TranscribeOpenAI2.3%5.8%572ms$0.0060
13GPT Live TranscribeOpenAInot measured4.5%1.12s$0.0170
14GPT-4o-mini TranscribeOpenAI2.7%6.4%460ms$0.0030
15Grok STTxAI4.8%10.9%305ms$0.0033
16Gradium ASRGradium8.4%11.7%334ms$0.0104
17Solaria-1Gladia5.0%11.4%596ms$0.0125
18Muse Voice TranscribeMetanot measured4.6%not measured$0.0030
19Gemini 3.5 Transcribe LiveGooglenot measurednot measured1.22s~$0.0090
LLM14
#ModelProviderId
1gpt-5.6-lunaOpenAI94.21%659ms70%0%$1.20
2gpt-4.1OpenAI91.40%640ms59%10%$8.00
3gpt-5.6-terraOpenAI87.37%701ms66%0%$12.00
4gpt-4.1-miniOpenAI87.19%708ms40%57%$1.60
5gpt-oss-120b (Baseten)Baseten86.32%393ms71%27%$0.50
6gpt-oss-120b (Cerebras)Cerebras85.00%195ms68%30%$0.75
7Claude Haiku 4.5Anthropic82.46%532ms2%0%$5.00
8Claude Sonnet 5Anthropic81.72%1.21s40%0%$10.00
9DeepSeek-V4-Flash-0731Baseten80.18%361ms5%10%$0.26
10gemma-4-31bCerebras77.72%192ms53%0%$1.49
11gemini-3.8-flashGemini77.37%not measured69%0%$3.75
12GLM-4.7Baseten72.46%275ms2%17%$2.20
13Llama-3.3-70BTogether64.21%698ms74%0%$1.04
14inkling-smallBaseten57.72%177ms72%0%$1.20
Text-to-speech23
#ModelProviderId
1eleven_v3_conversationalElevenLabs1590266ms$100.0
2gemini-3.1-flash-tts-previewGemini1591978ms~$33.3
3aura-2Deepgram1584125ms$30.0
4fluxDeepgram~1550106ms$45.0
5sonic-3.5Cartesia1574121ms$50.0
6sonic-3.6Cartesia~1578120ms$50.0
7tts-rt-v1Soniox1569362ms~$13.0
8tts-rt-v2Soniox~1606381ms~$13.0
9simba-3.2Speechify1573345ms$10.0
10inworld-tts-2Inworld1561116ms$25.0
11inworld-tts-2-flashInworld~150182ms$15.0
12s2.1-proFish Audio~1566185ms$15.0
13lightning_v3.1Smallest1544173ms$25.0
14grok-ttsxAI Grok1495272ms$15.0
15Gradium TTSGradium~1585272ms$57.8
16gpt-4o-mini-ttsOpenAI1424691ms~$20.0
17arcanav3Rime1429238ms$40.0
18speech-2.8-hdMiniMax1431294ms$100.0
19octave-2Hume1377448ms$100.0
20qwen3-tts-flashQwen1310472ms$10.0
21palabra-tts-v1Palabra~152072ms$30.0
22bland-speechBland~1569303ms$15.0
23Maya 2 NativeMayanot measured100ms~$4.0
Speech-to-speech10
#ModelProviderIdPriceas published
1grok-voice-think-fast-2.0xAI0.800.67820ms$0.08 · min
2gemini-3.1-flash-liveGemini0.770.781.11s$3.00 / $12.00 · 1M tok
3gpt-realtimeOpenAI0.760.83494ms$32 / $64 · 1M tok
4gpt-realtime-2.1-miniOpenAI0.750.831.01s$10 / $20 · 1M tok
5grok-voice-fastxAI0.720.441.11s$0.05 · min
6gpt-realtime-2OpenAI0.680.871.10s$32 / $64 · 1M tok
7gpt-realtime-2.1OpenAI0.630.561.10s$32 / $64 · 1M tok
8gpt-realtime-miniOpenAI0.620.70614ms$10 / $20 · 1M tok
9grok-voice-think-fast-1.0xAI0.560.221.01s$0.05 · min
10gpt-4.1-nano (cascade)Inworld0.451.00not measurednot measured
Measured per language
24 studies across 12 languages.
Spanish speech-to-text
#ModelWERbatch
1GPT-4o TranscribeOpenAI3.8%
2Universal-3.5 ProAssemblyAI4.0%
3Qwen3-ASRAlibaba4.4%
4GPT-4o-mini TranscribeOpenAI4.6%
5Chirp 3Google5.4%
6Gradium ASRGradium6.7%
7Ink-WhisperCartesia7.3%
8stt-rt-v5Soniox7.3%
9Nova-3Deepgram8.4%
10Pulse ProSmallest AI8.4%
11Nova-2Deepgram11.3%
12Scribe v2 RealtimeElevenLabsnot measured
13Qwen3-ASR RealtimeAlibabanot measured
14PulseSmallest AInot measured
German speech-to-text
#ModelWERbatch
1GPT-4o TranscribeOpenAI2.1%
2Universal-3.5 ProAssemblyAI2.6%
3Qwen3-ASRAlibaba3.0%
4GPT-4o-mini TranscribeOpenAI4.1%
5Scribe v2 RealtimeElevenLabsnot measured
6Chirp 3Google5.0%
7Qwen3-ASR RealtimeAlibabanot measured
8Gradium ASRGradium5.7%
9Ink-WhisperCartesia7.4%
10PulseSmallest AInot measured
11Pulse ProSmallest AI9.2%
12Nova-3Deepgram11.2%
13Nova-2Deepgram12.0%
14stt-rt-v5Soniox13.4%
French speech-to-text
#ModelWERbatch
1Universal-3.5 ProAssemblyAI3.3%
2GPT-4o TranscribeOpenAI4.0%
3Qwen3-ASR RealtimeAlibabanot measured
4Qwen3-ASRAlibaba4.4%
5Scribe v2 RealtimeElevenLabsnot measured
6GPT-4o-mini TranscribeOpenAI6.8%
7GPT Live TranscribeOpenAInot measured
8Chirp 3Googlenot measured
9stt-rt-v5Soniox8.6%
10Nova-3Deepgram9.9%
11Gradium ASRGradium10.0%
12Ink-WhisperCartesia10.4%
13PulseSmallest AInot measured
14Pulse ProSmallest AI11.8%
15Nova-2Deepgram12.9%
Arabic speech-to-text
#ModelCERbatch
1GPT-4o TranscribeOpenAI2.7%
2Universal-3.5 ProAssemblyAI3.3%
3Qwen3-ASRAlibaba3.5%
4Qwen3-ASR RealtimeAlibabanot measured
5Scribe v2 RealtimeElevenLabsnot measured
6Grok STTxAInot measured
7GPT-4o-mini TranscribeOpenAI4.7%
8S3Hamsa5.0%
9stt-rt-v5Soniox5.1%
10Nova-3Deepgram7.1%
11Ink-WhisperCartesia10.2%
Filipino speech-to-text
#ModelWERbatch
1GPT-4o TranscribeOpenAI6.1%
4GPT-4o-mini TranscribeOpenAI10.8%
5Scribe v2 RealtimeElevenLabsnot measured
6GPT Live TranscribeOpenAInot measured
7Velma 2 multilingualModulate12.5%
8Qwen3-ASRAlibaba18.8%
9Qwen-ASRAlibaba18.8%
10Ink-WhisperCartesia20.7%
11stt-rt-v5Soniox21.9%
12Nova-3Deepgram25.6%
13Chirp 3Google27.3%
Norwegian speech-to-text
#ModelWERbatch
2Scribe v1ElevenLabs5.7%
3GPT-4o TranscribeOpenAI6.4%
5Scribe v2 RealtimeElevenLabsnot measured
6Velma 2 multilingualModulate8.4%
7stt-rt-v5Soniox8.6%
8Universal-3.5 ProAssemblyAI10.1%
9Chirp 3Google11.2%
10Qwen3-ASRAlibaba12.0%
11Qwen-ASRAlibaba12.0%
12GPT-4o mini TranscribeOpenAI14.4%
13Qwen3-ASR RealtimeAlibabanot measured
14Ink-WhisperCartesia14.5%
15Nova-3Deepgram14.6%
16Nova-2Deepgram19.1%
Hindi speech-to-text
#ModelWERbatch
1Chirp 3Google6.1%
2Qwen3-ASRAlibaba9.8%
3stt-rt-v5Soniox12.3%
4GPT-4o-mini TranscribeOpenAI12.7%
5Pulse ProSmallest AI13.4%
6Universal-3.5 ProAssemblyAI17.2%
7Nova-3Deepgram20.3%
8GPT-4o TranscribeOpenAI21.0%
9Nova-2Deepgram25.1%
10Ink-WhisperCartesia41.9%
Tamil speech-to-text
#ModelCERbatch
1Chirp 3Google7.4%
2GPT-4o TranscribeOpenAI10.1%
3Pulse ProSmallest AI12.1%
4Nova-3Deepgram14.2%
5stt-rt-v5Soniox14.6%
6GPT-4o-mini TranscribeOpenAI16.1%
7Ink-WhisperCartesia30.6%
Telugu speech-to-text
#ModelCERbatch
1Chirp 3Google3.2%
2GPT-4o TranscribeOpenAI7.5%
3Pulse ProSmallest AI8.5%
4stt-rt-v5Soniox9.8%
5GPT-4o-mini TranscribeOpenAI10.2%
6Nova-3Deepgram13.4%
Korean speech-to-text
#ModelCERstream
1stt-rt-v5Sonioxnot measured
2Scribe v2 RealtimeElevenLabsnot measured
3Grok STTxAI
direct benchmark
not measured4Qwen3-ASRAlibaba
direct benchmark
not measured5GPT-4o TranscribeOpenAInot measured
6Solaria-1Gladianot measured
8Nova-3Deepgramnot measured
Chinese (Mandarin) speech-to-text
#ModelCERstream
1stt-rt-v5Sonioxnot measured
2Qwen3-ASRAlibaba
direct benchmark
not measured3Scribe v2 RealtimeElevenLabsnot measured
4Grok STTxAI
direct benchmark
not measured6GPT-4o TranscribeOpenAInot measured
7Nova-3Deepgramnot measured
8Solaria-1Gladianot measured
Japanese speech-to-text
#ModelCERbatch
1GPT-4o TranscribeOpenAI1.6%
3stt-rt-v5Soniox2.0%
4Scribe v1ElevenLabs2.1%
5Qwen3-ASRAlibaba2.1%
6Qwen-ASRAlibaba2.1%
8GPT-4o mini TranscribeOpenAI2.8%
9Universal-3.5 ProAssemblyAI
direct benchmark
not measured10Scribe v2 RealtimeElevenLabsnot measured
11Velma 2 multilingualModulate
direct benchmark
4.8%12Pulse ProSmallest AI5.5%
13PulseSmallest AI
direct benchmark
not measured14GPT Live TranscribeOpenAI
direct benchmark
not measured15Solaria-1Gladia11.5%
16Nova-3Deepgram20.7%
17Ink-WhisperCartesia30.3%
18Grok STTxAI42.9%
Spanish text-to-speech
#ModelNaturalnessElo
1eleven_v3_conversationalElevenLabs1951.00
2sonic-3.5Cartesia1751.00
3tts-rt-v2Soniox1696.00
4grok-ttsxAI Grok1634.00
5inworld-tts-2Inworld1598.00
6octave-2Hume1568.00
7tts-rt-v1Soniox1525.00
8lightning_v3.1Smallest1487.00
9speech-2.8-hdMiniMax1460.00
10qwen3-tts-flashQwen1437.00
11gpt-4o-mini-ttsOpenAI1408.00
12arcanav3Rime1172.00
13aura-2Deepgram1055.00
German text-to-speech
#ModelNaturalnessElo
1eleven_v3_conversationalElevenLabs1798.00
2grok-ttsxAI Grok1762.00
3tts-rt-v1Soniox1717.00
4tts-rt-v2Soniox1700.00
5octave-2Hume1687.00
6sonic-3.5Cartesia1665.00
7inworld-tts-2Inworld1576.00
8qwen3-tts-flashQwen1573.00
9speech-2.8-hdMiniMax1528.00
10gpt-4o-mini-ttsOpenAI1497.00
11arcanav3Rime1257.00
12lightning_v3.1Smallest1254.00
13Gradium TTSGradium1245.00
14aura-2Deepgram1062.00
French text-to-speech
#ModelNaturalnessElo
1eleven_v3_conversationalElevenLabs1775.00
2tts-rt-v2Soniox1695.00
3grok-ttsxAI Grok1692.00
4inworld-tts-2Inworld1684.00
5octave-2Hume1671.00
6tts-rt-v1Soniox1628.00
7sonic-3.5Cartesia1622.00
8gpt-4o-mini-ttsOpenAI1535.00
9qwen3-tts-flashQwen1529.00
10speech-2.8-hdMiniMax1419.00
11lightning_v3.1Smallest1416.00
12arcanav3Rime1316.00
13Gradium TTSGradium1233.00
14aura-2Deepgram1114.00
Arabic text-to-speech
#ModelNaturalnessElo
1grok-ttsxAI Grok1736.00
2sonic-3.5Cartesia1644.00
3tts-rt-v2Soniox1612.00
4eleven_v3_conversationalElevenLabs1564.00
5inworld-tts-2Inworld1454.00
Filipino text-to-speech
#ModelNaturalnessElo
1gemini-3.1-flash-tts-previewGemini2012.00
2tts-rt-v2Soniox1790.00
3eleven_v3_conversationalElevenLabs1693.00
4gpt-4o-mini-ttsOpenAI1676.00
5grok-ttsxAI Grok1657.00
6speech-2.8-hdMiniMax1598.00
7sonic-3.5Cartesia1597.00
8inworld-tts-2Inworld1251.00
Norwegian text-to-speech
#ModelNaturalnessElo
1sonic-3.5Cartesia1836.00
2gemini-3.1-flash-tts-previewGemini1705.00
3eleven_v3_conversationalElevenLabs1628.00
4tts-rt-v2Soniox1621.00
5Azure AI Speech (native)Azure
direct benchmark
1467.006gpt-4o-mini-ttsOpenAI1420.00
7speech-2.8-hdMiniMax1386.00
Hindi text-to-speech
#ModelNaturalnessElo
1sonic-3.5Cartesia1736.00
2gemini-3.1-flash-tts-previewGemini1563.00
3speech-2.8-hdMiniMax1522.00
4eleven_v3_conversationalElevenLabs1490.00
5Maya 2 NativeMaya1478.00
6chirp-3-hdGoogle Chirp1468.00
7Azure AI Speech (native)Azure
direct benchmark
1412.008simba-multilingualSpeechify1310.00
Tamil text-to-speech
#ModelNaturalnessElo
1gemini-3.1-flash-tts-previewGemini1676.00
2sonic-3.5Cartesia1620.00
3Maya 2 NativeMaya1603.00
4chirp-3-hdGoogle Chirp1555.00
5eleven_v3_conversationalElevenLabs1543.00
6Azure AI Speech (native)Azure
direct benchmark
1462.007simba-multilingualSpeechify1144.00
Telugu text-to-speech
#ModelNaturalnessElo
1sonic-3.5Cartesia1639.00
2gemini-3.1-flash-tts-previewGemini1625.00
3chirp-3-hdGoogle Chirp1548.00
4Maya 2 NativeMaya1540.00
5Azure AI Speech (native)Azure
direct benchmark
1419.006simba-multilingualSpeechify1269.00
Chinese (Mandarin) text-to-speech
#ModelNaturalnessElo
1inworld-tts-2Inworld1688.00
2eleven_v3_conversationalElevenLabs1675.00
3gemini-3.1-flash-tts-previewGemini1649.00
4sonic-3.5Cartesia1564.00
5Azure AI Speech (native)Azure
direct benchmark
1552.006speech-2.6-hdMiniMax1467.00
7gpt-4o-mini-ttsOpenAI1365.00
Korean text-to-speech
#ModelNaturalnessElo
1gemini-3.1-flash-tts-previewGemini1917.00
2sonic-3.5Cartesia1859.00
3inworld-tts-2Inworld1722.00
4eleven_v3_conversationalElevenLabs1628.00
5gpt-4o-mini-ttsOpenAI1498.00
6Azure AI Speech (native)Azure
direct benchmark
1438.007speech-2.6-hdMiniMax1407.00
8codaRime1187.00
9generativeAWS Polly953.00
Japanese text-to-speech
#ModelNaturalnessElo
1gemini-3.1-flash-tts-previewGemini1756.00
2eleven_v3ElevenLabs1646.00
3sonic-3.5Cartesia1600.00
4inworld-tts-2Inworld1597.00
5DragonHDLatestNeuralAzure1539.00
6tts-rt-v2Soniox1447.00
7speech-2.6-hdMiniMax1441.00
8chirp-3-hdGoogle Chirp 3 HD1407.00
9gpt-4o-mini-ttsOpenAI1338.00
10octave-2Hume1228.00
Published stacks
What benchmarks.speko.ai picks for each job off the boards above, and the measurement that decided it. A stack whose legs are identical to another’s is one row: the publisher distinguishes four jobs and its boards distinguish two stacks. A stack with a leg Router cannot currently call is not shown at all — a measurement may stay visible in that state, an instruction you would copy may not.
Use case
Real-time phone agentdecided on 66ms · 192ms · 116ms
Speech-to-textUniversal-3.5 Proassemblyai:universal-3-5-pro
LLMgemma-4-31bcerebras:gemma-4-31b
Text-to-speechinworld-tts-2inworld:inworld-tts-2
Accuracy-criticaldecided on 2.0% WER · 0% fabrication · 0.93 robustnessTool-heavy agentdecided on 2.0% WER · 2.7% tool silence · 116ms
Speech-to-textUniversal-3.5 Proassemblyai:universal-3-5-pro
LLMClaude Haiku 4.5anthropic:claude-haiku-4-5
Text-to-speechinworld-tts-2inworld:inworld-tts-2
Natural conversationdecided on 2.0% WER · 1.6% dead-air · MOS 1590
Speech-to-textUniversal-3.5 Proassemblyai:universal-3-5-pro
LLMClaude Haiku 4.5anthropic:claude-haiku-4-5
Text-to-speecheleven_v3_conversationalelevenlabs:eleven_v3_conversational
How to read this
Every model id is copyable. Click an id to copy it. An id offered by the current Router catalog can be sent as model; other ids identify benchmark evidence without claiming that the route is currently available. A direct benchmark label means the provider published no Router model id for that row.
Not measured is not zero. A dash in a column of word error rates reads as a perfect score, so an unmeasured cell says so in words. A leading ~ is the board’s own mark for an estimate.
Nothing here re-ranks a board. The # column indexes the order you sorted into. It is not a verdict.
The catalog endpoint did not answer this request, so the routable column comes from the bundled snapshot. The boards are unaffected.