German vs English Voice AI: What Our Benchmark Shows

A benchmark published October 13, 2026 by voice AI platform Famulor found that German does not slow down voice agents: across nine pipeline stacks, median confirmation time was 1249 ms in German versus 1308 ms in English, and seven of the nine stacks were faster in German. The main language gap is coverage and accuracy rather than latency — only 15 of 23 speech recognition models understand German, and model rankings shift by language. The test ran real WebRTC calls in production with LiveKit Agents 1.8.6, holding the language model (Gemma 4 31B IT) and speech recognition (Soniox STT v5) constant while varying only the voice provider.

Summarize Content With: German vs English Voice AI: What Our Benchmark Shows Short answer: German does not make your voice agent slower. In our benchmark from October 13, 2026, all nine pipeline stacks passed the four test tasks in German and in English. Seven of the nine even confirmed slightly faster in German. The real differences sit elsewhere. Only 15 of 23 speech recognition models understand German. The ranking of the best models changes by language. And after a German interruption, the agent went quiet later than after an English one. This article takes the same measurements as our Voice AI Benchmark October 2026 https://www.famulor.io/blog/voice-ai-benchmarkchoosing-oct-2026 and splits them by language. You get the numbers per language, the limits of the sample and a plan for your own German acceptance test. Terms and test setup Each number below has a precise definition and clear limits. Confirmation time We measure the time from the end of the caller's utterance to the agent's first audible audio after the test tool's receipt has arrived. Network, tool call and the continued answer are all included. It is a proxy for "the caller hears the confirmation", not a timestamp of the first meaningful word. Word error rate WER Word error rate counts all substitutions, deletions and insertions and divides them by the number of reference words. We use micro WER across the whole corpus. The German corpus has 161 reference words, the English one 124. A single error therefore moves the German result by 0.62 percentage points and the English one by 0.81. For the basics, see our glossary entry on speech-to-text https://www.famulor.io/glossary/stt-speech-to-text . Time to silence For the interruption test, the caller talks into a long, running answer. We measure from the start of the interruption to the agent's first silent audio frame. Whether the agent then gives the right, short answer is checked separately. How the test ran All calls ran through real WebRTC rooms in production, with a worker in the EU and LiveKit Agents 1.8.6. There were no phone network calls and no real customer data. Each language and configuration got four fictional tasks: a returns question, an appointment https://www.famulor.io/use-cases/appointment-booking-faqs confirmation, a pricing https://www.famulor.io/pricing question and a router error. An interruption with recovery followed. The pipeline stacks always used the same language model Gemma 4 31B IT and the same speech recognition Soniox STT v5 . Only the voice provider changed. Speech recognition had its own test with eight FLEURS clips per language, 136.7 seconds of read speech in total. Test setup: identical tasks in both languages, only the voice changes per pipeline. Latency: does German cost time? No, at least not in this sample. Across all nine pipeline stacks, the median confirmation time was 1249 ms in German and 1308 ms in English. Seven stacks were faster in German, two marginally slower. Rime had the largest gap: 1232 ms in German versus 1469 ms in English. With ElevenLabs https://www.famulor.io/alternatives/elevenlabs-alternative plus 10 ms and xAI plus 27 ms , German was marginally slower. With four calls, gaps like these are within chance. The full values, including the slowest call: | Voice provider | DE P50 ms | EN P50 ms | Difference ms | DE P95 ms | EN P95 ms | |---|---|---|---|---|---| | Gradium | 1128 | 1209 | −81 | 1205 | 1272 | | Deepgram | 1178 | 1308 | −130 | 1201 | 1425 | | Rime | 1232 | 1469 | −237 | 1430 | 1499 | | Cartesia | 1239 | 1292 | −53 | 1319 | 1371 | | Inworld | 1249 | 1269 | −20 | 1318 | 1392 | | ElevenLabs | 1257 | 1247 | +10 | 1318 | 1326 | | Soniox | 1400 | 1487 | −87 | 1478 | 1616 | | Fish Audio | 1481 | 1539 | −58 | 1590 | 1610 | | xAI | 1582 | 1555 | +27 | 3054 | 1664 | One outlier: the slowest German call with xAI took 3054 ms. A single call, not a pattern, and exactly what more test calls catch. All nine stacks passed all four tasks in both languages. For your planning this means: you do not need an extra latency buffer for German. Plan your budget as you would for English, then measure with your own sentences. Native speech models: a different picture Native audio models process speech without a separate speech recognition step. For them, the language comparison looked different. The half-cascade variants of GPT Realtime 2 and 2.1 passed all four tasks in German and three of four in English. | Configuration | DE checks | DE P50 ms | EN checks | EN P50 ms | |---|---|---|---|---| | Half-Cascade GPT Realtime 2 | 4/4 | 1940 | 3/4 | 2444 | | Half-Cascade GPT Realtime 2.1 | 4/4 | 1780 | 3/4 | 1990 | | Half-Cascade GPT Realtime 2.1 Mini | 3/4 | 2260 | 3/4 | 1899 | | Realtime GPT Realtime 2 | 3/4 | 2954 | 3/4 | 2476 | | Realtime GPT Realtime 2.1 | 4/4 | 2658 | 4/4 | 2616 | | Realtime GPT Realtime 2.1 Mini | 3/4 | 2157 | 3/4 | 3026 | Together, the six native configurations passed 21 of 24 checks in German and 19 of 24 in English. The median sat roughly between 1.8 and 3 seconds in both languages, well above the pipeline stacks. The Mini variants had long outliers in both languages: 9358 ms in German and 12913 ms in English for Realtime 2.1 Mini. The Gemini Live variants and GPT Realtime Mini did not complete the test protocol in either language in this run. We did not grade GPT Live 1 because its transcripts were incomplete. Both are therefore left out of the language comparison. Speech recognition: who understands German best? This is where the most important difference for German calls shows up. The model that leads in English is not automatically the best one for German. The comparison shows every model that completed the full corpus in both languages. Three observations stand out. ElevenLabs Scribe v2 Realtime led in German 0.62%, a single error and ranked fifth in English. Google Gemini 3.5 Transcribe Live reached 1.24% in German and 7.26% in English. The AssemblyAI U3 family stayed among the three best values in both languages. That the German values are lower almost everywhere does not mean German is easier to recognize. The clips differ, the German corpus has male speakers only, and a single error is worth 0.62 percentage points. Read the result as a ranking within one language, not as a comparison between languages. | STT model | German WER | English WER | Status | |---|---|---|---| | ElevenLabs Scribe v2 Realtime | 0.62% | 3.23% | both complete | | Google Gemini 3.5 Transcribe Live | 1.24% | 7.26% | both complete | | AssemblyAI U3 RT Pro | 1.86% | 2.42% | both complete | | AssemblyAI Universal 3.5 Pro | 1.86% | 2.42% | both complete | | AssemblyAI Universal 3.6 Pro | 1.86% | 2.42% | both complete | | Speechmatics Linden-1 | 2.48% | 8.06% | both complete | | Soniox STT v5 | 3.73% | 8.06% | both complete | | xAI STT 1 | 3.73% | not graded | EN only 6 of 8 clips | | xAI STT 2 | 3.73% | not graded | EN only 6 of 8 clips | | Deepgram Nova-3 | 4.35% | 11.29% | both complete | | Deepgram Flux General Multi | 6.21% | 9.68% | both complete | | Deepgram Nova-2 | 7.45% | 16.13% | both complete | | Cartesia Ink Whisper | 8.07% | not graded | EN only 6 of 8 clips | | Gladia Solaria-1 | not graded | not graded | stream closed early | | AssemblyAI Universal Streaming Multilingual | not graded | not graded | duplicate final results | Incomplete runs never count as zero errors. They are listed separately. These models understand English only Eight of the 23 STT models in the catalog are English-only models. For a German or bilingual agent they drop out, no matter how well they do in English: - Cartesia Ink-2 2.42% in English, strong, but no German - Deepgram Flux General EN 6.45% - Deepgram Nova-3 Pharma 12.90% - Deepgram Nova-2 Conversational AI 13.71% - Deepgram Nova-2 Medical, Nova-2 Phonecall and Nova-3 Medical 14.52% each - AssemblyAI Universal Streaming not graded, duplicate final results 28 of 29 voice models also ran in German in the test, only Deepgram Flux TTS ran in English only. We tested all 33 language models in both languages. Their language quality cannot be compared from this test, though, because the tasks differed in difficulty per language. To see which languages Famulor covers, visit German at Famulor https://www.famulor.io/supported-languages/de . Interruptions: German takes longer to go quiet This was the clearest language difference in the whole test. The caller asked for a detailed explanation and then interrupted with "Stopp. Ich brauche nur den ersten Schritt, bitte ganz kurz." In English the line was "Stop. I only need the first step, please keep it brief." At the median, the agent went quiet 1451 ms after a German interruption and 848 ms after an English one. Deepgram was the exception at 56 ms in German. Importantly, in all 18 pipeline cells the agent then passed the recovery check and briefly named the first step. So the interruption worked, it just took longer in German. We cannot tell the cause from one interruption per cell. Plausible factors are the length of German words, recognition of the first word and the moment the agent treats the input as an interruption. What matters for you is practice: test interruptions in German explicitly. An agent that stops instantly in English does not necessarily do so in German. Implementation: how to test your agent in German The numbers above are a starting point. Decide with your own calls: 1. Set the languages. Decide whether your agent speaks German only, English only or both. With two languages, the eight English-only speech recognition models drop out. 2. Pick speech recognition by its German result. Start with a shortlist of ElevenLabs Scribe v2 Realtime, Gemini 3.5 Transcribe Live, the AssemblyAI U3 family and Speechmatics Linden-1. Then test them with your own phone recordings, including dialects and background noise. 3. Pick a voice per language. Use a native German voice for German, not the English voice in German mode. In the voice library https://www.famulor.io/voices you can filter voices by language and listen to them. 4. Translate test cases, do not just copy them. German callers say dates differently: "am zwanzigsten Oktober um zehn Uhr". Add spelled-out numbers, umlauts in names, compound words and the formal or informal form of address to your tests. 5. Test interruptions in German. Try "Stopp", "Moment mal" and "Nein, warte" in the middle of a long answer. Measure when the agent goes quiet and whether it then gives the right short answer. 6. Measure and report per language. Evaluate P50 and P95 separately per language. A mixed average hides exactly the differences your callers notice. 7. Back it up with more calls. Four calls show a direction. For a release you need far more calls per language, spread across different times of day. Checklist for your German acceptance test. Best practices and common mistakes Common mistakes - Copying the English ranking. Cartesia Ink-2 is among the best models in English but does not understand German. Gemini 3.5 Transcribe Live ranked seventh in English and second in German. Check the German ranking separately. - An English-only model in a bilingual agent. Eight models do not understand German. A test catches that right away, live traffic only with the first German caller. - Adding a latency buffer for German instead of measuring. In this test German was not slower. A flat surcharge costs you conversation quality for no reason. - Not testing numbers and dates. Spelled-out dates and prices are longer and more error-prone in German. They belong in every test case. Best practices - Maintain pronunciation. Product names, street names and technical terms often need help. Our guide to a custom pronunciation dictionary https://www.famulor.io/blog/custom-vocabulary-for-ai-voice-agents-pronunciation-guide shows how. - Short answers. Short sentences make interruptions less painful and cut the time to the first answer. - Separate reports per language. That way you see right away when a model update makes only one language worse. - Retest regularly. Providers change models and routes. A result from October 2026 can look different in December. Industry examples Dental practice in Munich A patient calls and confirms "Tuesday the twentieth at two p.m." The agent has to recognize date and time exactly and confirm the appointment without a double booking. In the benchmark, every pipeline stack passed the appointment confirmation in German. For the practice, recognizing names with umlauts matters too, so "Müller" or "Größer" belong in the test. Online shop with customers in Germany and the UK Customers ask about the return window, in German or English. A bilingual agent with speech recognition that handles both languages pays off here. The answer "14 days" came out right for every pipeline stack in both languages. What matters is a voice that sounds natural in both languages. IT service provider with a hotline A customer reports router error E17, and the agent starts a long walkthrough. After two sentences the customer says "Stop, just the first step". The benchmark tested exactly this scenario. Plan short answer steps here so that a German interruption takes effect quickly. Estimate your ROI from automating calls See how much your business could save by switching to AI-powered voice agents. ROI Result ROI 0% Get started https://app.famulor.io/register No credit card required Conclusion German is no brake on voice AI. In our benchmark, the pipeline stacks were just as reliable in German and at the median even slightly faster than in English. The decisions that really count are speech recognition and interruptions: pick an STT model by its German result, not by the English ranking, and test interruptions explicitly in German. With Famulor you choose speech recognition, language model and voice per agent yourself and can compare test calls in both languages directly. You will find all supported languages under Languages https://www.famulor.io/feature/languages . Next step: book a demo and hear your agent in German. Try our AI Assistant Experience how natural our AI phone assistant sounds. Enter your details and receive a call from our AI agent within seconds. Agent is trained to discuss Famulor services and book appointments. Demo AI agent Famulor representative FAQ Is voice AI slower in German than in English? Not in our benchmark. Across nine pipeline stacks, the median confirmation time was 1249 ms in German and 1308 ms in English, with four calls per language and stack. Which STT model recognizes German best? In our short corpus, ElevenLabs Scribe v2 Realtime led with a 0.62% word error rate, ahead of Gemini 3.5 Transcribe Live 1.24% and the AssemblyAI U3 family 1.86% . Test the shortlist with your own phone recordings. Is German easier to recognize than English? This test cannot tell. The clips differ, the German corpus has male speakers only, and one error is worth 0.62 percentage points. Compare models only within one language. Do all models support German? No. 15 of 23 STT models and 28 of 29 voice models ran in German. All 33 language models were tested in both languages. Eight STT models understand English only. How do I test interruptions in German? Interrupt the agent in the middle of a long answer with "Stopp" or "Moment mal". Measure the time to silence and check that the next answer is short and correct. In the benchmark this took a median of 1451 ms in German. Are native speech models the better choice for German? Not automatically for short service tasks. Native configurations had medians of about 1.8 to 3 seconds, pipelines about 1.1 to 1.6 seconds. The half-cascade variants of GPT Realtime 2 and 2.1 did pass all four tasks in German, though. How current are these results? We measured on October 13, 2026. Providers change models and routes all the time. Use the numbers as a starting point and test again before you go live. Can I rebuild the test with Famulor? Yes. In Famulor you choose speech recognition, language model and voice per agent, run test calls in German and English and compare the conversations in the call history. Sources and method All numbers come from the Famulor Labs benchmark of October 13, 2026: 44 language and configuration cells in production WebRTC rooms, LiveKit Agents 1.8.6, EU worker, plus 304 human-speech requests to STT models. The speech corpus comes from the FLEURS https://huggingface.co/datasets/google/fleurs dataset CC BY 4.0 . You will find the full tables across all models in the Voice AI Benchmark October 2026 https://www.famulor.io/blog/voice-ai-benchmarkchoosing-oct-2026 . Writer at Famulor