--- title: Telnyx STT vs ElevenLabs Scribe Benchmark on Phone Audio canonical: https://telnyx.com/resources/telnyx-stt-vs-elevenlabs-scribe-benchmark --- # Telnyx STT vs ElevenLabs Scribe Benchmark on Phone Audio [Back to blog](https://telnyx.com/resources) **STT API** # Telnyx STT against ElevenLabs Scribe on real phone audio We tested Telnyx's speech-to-text engines against ElevenLabs Scribe on 8 kHz phone audio. Telnyx's fastest streaming engines returned first words sooner, and most Telnyx engines handled one-word replies better, while Scribe was more accurate on recorded consultations and on German, French and Turkish. By [Abhishek Sharma](https://telnyx.com/authors/abhishek-sharma) ElevenLabs Scribe comes up in a lot of our sales conversations, so we tested it against Telnyx's speech-to-text engines on phone audio at 8 kHz, which is what our customers send us. Telnyx's fastest streaming engines returned first words sooner, and most Telnyx engines handled one-word replies better. Scribe was more accurate on recorded consultations and on German, French and Turkish. On medical phone calls the two were about level. We ran the tests on October 1, 2026\. Each engine received the same audio, the same language setting and default options, and every transcript was scored against a human transcript or a read script. No AI model judged accuracy. The code and results are in \[TODO: link to the GitHub repo once it is public\], so you can rerun any number below on your own audio. ## The short version * **On live calls, Telnyx's fastest engines return words sooner.** Median time to first words was 1.0 seconds for Flux and 1.3 seconds for Nova-3, against 2.1 seconds for Scribe v2 Realtime. Three engines served through Telnyx start slower than Scribe, at 3.3 to 4.0 seconds. * **One-word replies go better on most Telnyx engines.** On 147 short replies such as "yes", "no", single letters and numbers, Flux got 85% right and Scribe v2 Realtime got 67%. * **Medical phone calls are close to even.** Nova-3 Medical got 16.4 words wrong per 100 and Scribe v2 Realtime got 17.3, on 90 lines. * **Scribe is more accurate on consultation audio.** On recorded consultations, Scribe v2 got 4.0 wrong words per 100 and the best Telnyx file engine got 5.7. * **Scribe v2 is more accurate in German, French and Turkish files.** Live, Telnyx's preview engine Linden-1 was ahead of Scribe v2 Realtime in French and level in Turkish. ## What we tested We used three audio sets and converted all of them to 8 kHz phone quality before any engine heard them. * **Phone calls.** Five people on iPhone, Android and speakerphone read a script over a real phone line. It has names and spelled-out items, one-word replies, and medical calls with drug names, doses and pharmacies. That is 339 clips and 28 minutes of speech. * **Consultations.** PriMock57, a public set of mock primary-care visits. We used 5.2 hours for the file test and a 93-minute slice for the live test. * **Read speech.** Google FLEURS sentences in English, German, French and Turkish, 400 clips and 77 minutes in all. Every table reports **wrong words per 100**, where lower is better. A 6.1 means about 6 words in every 100 were wrong. Timing columns are in seconds. Every engine runs on Telnyx except the rows marked ElevenLabs. Linden-1 is in preview and is not in Telnyx's public model list yet. Parakeet v3 and Omi Med STT v1 return only finished transcripts, so their time to first words is the time the finished text arrives. ## Live calls: Flux and Nova-3 answer sooner On live phone audio, the question is how long a caller waits. Time to first words is how long after the audio starts until the engine returns any text. Time to final transcript is how long after the speaker stops until the finished text arrives, shown here for the slowest 1 in 10 turns. | Engine | Time to first words, median (s) | Time to final transcript, slowest 10% (s) | One-word replies right | Wrong words per 100, medical calls | | ----------------------------------- | ------------------------------- | ----------------------------------------- | ---------------------- | ---------------------------------- | | Flux (Deepgram) | 1.0 | 2.36 | 85% | 20.3 | | Azure Fast | 4.0 | 1.56 | 83% | 17.4 | | Grok STT (xAI) | 3.8 | 1.47 | 82% | 21.0 | | Universal 3.5 Pro (AssemblyAI) | 3.3 | 1.61 | 81% | 23.5 | | Nova-3 (Deepgram) | 1.3 | 0.93 | 80% | 17.7 | | Soniox stt-rt-v4 | 1.3 | 1.51 | 77% | 19.4 | | Nova-3 Medical (Deepgram) | 1.3 | 0.97 | 73% | 16.4 | | Speechmatics Standard | 1.3 | 1.17 | 73% | 28.1 | | Omi Med STT v1 | 6.7 | 4.68 | 71% | 21.8 | | Linden-1 (Speechmatics, preview) | 1.3 | 0.77 | 67% | 15.7 | | **Scribe v2 Realtime (ElevenLabs)** | 2.1 | 1.32 | 67% | 17.3 | | Parakeet v3 (NVIDIA) | 6.7 | 4.55 | 51% | 26.6 | Flux starts fastest, at 1.0 seconds, and finishes slowest of the fast group, at 2.36 seconds for its slowest turns, because it waits for its own end-of-turn signal. Nova-3 starts at 1.3 seconds and finishes its slowest turns in 0.93 seconds, against 1.32 for Scribe v2 Realtime. The one-word replies column counts a reply as right when the whole line matched the script. We allowed one stray extra word at the start. Nine of the 11 Telnyx streaming engines scored higher than Scribe v2 Realtime on this set, Linden-1 tied it, and Parakeet v3 scored lower. Spelled names and addresses were hard for every engine. Across every streaming engine, between 36 and 59 words in 100 came back wrong in that set, so none of these results should be read as "solved". ## Medical audio: level on calls, Scribe ahead on consultations The medical phone-call lines are refill requests, drug names, doses, pharmacies and insurers. Nova-3 Medical got 16.4 wrong words per 100 and Scribe v2 Realtime got 17.3\. That gap is under one point on 90 lines, which is inside the noise. \[TODO: Telnyx's model docs list Nova-3 Medical as REST only, but it returned results over WebSocket in our runs. Confirm with product that live use is supported before this post publishes.\] When we gave both engines the same 21-term vocabulary list of drugs, pharmacies and insurers, the numbers moved to 15.4 and 16.2. Consultations are a different test, with longer, cleaner speech between a doctor and a patient, streamed live. | Engine | Wrong words per 100 (95% range) | Time to first words, median (s) | Time to final transcript, slowest 10% (s) | | ----------------------------------- | ------------------------------- | ------------------------------- | ----------------------------------------- | | **Scribe v2 Realtime (ElevenLabs)** | 6.1 (5.2 to 6.9) | 1.7 | 0.70 | | Parakeet v3 (NVIDIA) | 7.5 (6.5 to 8.5) | 4.8 | 4.15 | | Nova-3 Medical (Deepgram) | 8.5 (7.7 to 9.4) | 1.0 | 0.71 | | Omi Med STT v1 | 8.6 (7.7 to 9.5) | 4.8 | 4.00 | | Nova-3 (Deepgram) | 8.8 (7.9 to 9.8) | 1.0 | 0.71 | | Grok STT (xAI) | 8.9 (7.8 to 9.9) | 3.2 | 1.38 | | Soniox stt-rt-v4 | 10.0 (8.9 to 11.0) | 1.2 | 0.85 | | Linden-1 (Speechmatics, preview) | 10.0 (8.0 to 12.8) | 1.0 | 0.88 | | Flux (Deepgram) | 10.4 (9.1 to 11.6) | 0.8 | 1.43 | | Azure Fast | 11.1 (10.0 to 12.4) | 2.9 | 1.31 | | Universal 3.5 Pro (AssemblyAI) | 14.6 (13.6 to 15.6) | 2.7 | 1.27 | | Speechmatics Standard | 14.7 (12.6 to 17.4) | 1.1 | 1.00 | Scribe v2 Realtime has the lowest error rate, and its range of 5.2 to 6.9 does not overlap Nova-3 Medical's 7.7 to 9.4\. Parakeet v3 comes closest on accuracy, but it takes 4.8 seconds to return first words, which is slow for a live call. ## Recorded files and other languages: Scribe leads For recorded audio, where nobody is waiting on a live reply, accuracy is the number that matters. We tested Scribe v2 Medical on files only, so it appears here and not in the live tables. | Engine | Wrong words per 100 (95% range) | | ---------------------------------- | ------------------------------- | | **Scribe v2 (ElevenLabs)** | 4.0 (3.6 to 4.4) | | **Scribe v2 Medical (ElevenLabs)** | 4.8 (4.3 to 5.3) | | Parakeet v3 (NVIDIA) | 5.7 (5.0 to 6.4) | | Omi Med STT v1 | 5.8 (5.1 to 6.6) | | Nova-3 Medical (Deepgram) | 5.9 (5.2 to 6.7) | | Nova-3 (Deepgram) | 6.0 (5.2 to 6.8) | | Whisper large-v3-turbo (OpenAI) | 9.9 (8.8 to 11.1) | On German, French and Turkish read speech, Scribe v2 led every Telnyx file engine we tested: | Language | Scribe v2 | Scribe v2 Realtime | Best Telnyx file engine | Best Telnyx streaming engine | | -------- | --------- | ------------------ | ------------------------------------ | ------------------------------------- | | English | 4.6 | 5.3 | Parakeet v3 (NVIDIA), 5.2 | Linden-1 (Speechmatics, preview), 5.6 | | German | 4.7 | 6.5 | Whisper large-v3-turbo (OpenAI), 7.6 | Linden-1 (Speechmatics, preview), 6.9 | | French | 5.5 | 9.4 | Parakeet v3 (NVIDIA), 7.9 | Linden-1 (Speechmatics, preview), 7.1 | | Turkish | 5.7 | 9.6 | Whisper large-v3-turbo (OpenAI), 9.8 | Linden-1 (Speechmatics, preview), 9.5 | ## Which engine for which job * **A voice agent that has to take turns quickly and catch short confirmations:** Flux starts fastest and got the most one-word replies right, and Nova-3 finishes its slowest turns sooner. * **Live medical calls:** Nova-3 Medical and Scribe v2 Realtime were level on phone-call lines, so a vocabulary list is worth testing on your own audio. * **Recorded consultations where accuracy matters most:** Scribe v2 led in this test. * **German, French or Turkish files:** Scribe v2 led in this test. ## What this test does not show It is one set of runs, and engines change, so rerun it before you quote it. The samples are small in places, with five callers, 90 medical lines and 147 short replies, so treat gaps of a point or two as ties. We scored 339 of 375 phone lines and set 36 aside for a human listen. The phone recordings are not published, so that track cannot be rerun from the repo. Prices are not in these tables. ## Try it on your own audio Run the same test on your own recordings with the [Telnyx speech-to-text API](https://telnyx.com/products/speech-to-text-api?utm%5Fsource=telnyx%5Fblog&utm%5Fmedium=internal%5Flink&utm%5Fcampaign=202610%5FTOFU%5FSpeech%5Fto%5FText%5FTL%5FGLOBAL&utm%5Fcontent=telnyx-stt-vs-elevenlabs-scribe-benchmark), or read the [full list of our published benchmarks](https://telnyx.com/benchmarks?utm%5Fsource=telnyx%5Fblog&utm%5Fmedium=internal%5Flink&utm%5Fcampaign=202610%5FTOFU%5FSpeech%5Fto%5FText%5FTL%5FGLOBAL&utm%5Fcontent=telnyx-stt-vs-elevenlabs-scribe-benchmark). Share on Social Abhishek Sharma Sr Technical Product Marketing Manager Senior Technical Product Marketing Manager ## Jump to: [The short version](#the-short-version)[What we tested](#what-we-tested)[Live calls: Flux and Nova-3 answer sooner](#live-calls-flux-and-nova-3-answer-sooner)[Medical audio: level on calls, Scribe ahead on consultations](#medical-audio-level-on-calls-scribe-ahead-on-consultations)[Recorded files and other languages: Scribe leads](#recorded-files-and-other-languages-scribe-leads)[Which engine for which job](#which-engine-for-which-job)[What this test does not show](#what-this-test-does-not-show)[Try it on your own audio](#try-it-on-your-own-audio) ## Sign up for emails of our latest articles and news [Start building](https://telnyx.com/sign-up) Related articles * [**STT API**How I built a voice dictation app for macOS using Telnyx![Abhishek Sharma](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/Abishek_Sharma_8313622a2b.png?width=96&format=webp)By Abhishek Sharma8, Sep 2026](https://telnyx.com/resources/voice-dictation-app-macos-telnyx) * [**Insights and Resources**B300 GPU rental price per hour in 2026: what you actually pay![Eli Mogul](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/Eli_Mogul_1de9939dcb.png?width=96&format=webp)By Eli Mogul12, Oct 2026](https://telnyx.com/resources/b300-gpu-rental) * [**Messaging**SMS vs MMS Explained for Developers: Limits, Cost and One API![Serhii](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/_DSC8799-6.jpg?width=96&format=webp)By Serhii Omelchenko12, Oct 2026](https://telnyx.com/resources/sms-vs-mms) * [**Conversational AI**AI Agents in Financial Services: Use Cases in India![megha-profile](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/Untitled%20(400%20x%20400%20px)%20(1).png?width=96&format=webp)By Megha Sujanani12, Oct 2026](https://telnyx.com/resources/ai-agents-indian-financial-services) * [**Inference**NVIDIA B300 pricing in 2026: per hour, per system and per token![Eli Mogul](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/Eli_Mogul_1de9939dcb.png?width=96&format=webp)By Eli Mogul8, Oct 2026](https://telnyx.com/resources/nvidia-b300-pricing) * [**Inference**We tested 3 open-weight models for multi-turn agents![Sonam Gupta, PhD](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/sonam_gupta_phd_2ae3a78101.png?width=96&format=webp)By Sonam Gupta, PhD8, Oct 2026](https://telnyx.com/resources/choose-open-weight-model) ### Ask AI * [![GPT](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/chatgpt_icon-ea55e09d.jpeg?width=92&format=webp)](https://chat.openai.com/?q=What makes Telnyx's global network different, breakdown call quality for normal phone calls and voice AI calls.) * [![Claude](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/claude_icon-c7c8eef3.webp?width=92&format=webp)](https://claude.ai/new?q=What makes Telnyx's global network different, breakdown call quality for normal phone calls and voice AI calls.) * [![Perplexity](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/perplexity_icon-cd5b74f1.webp?width=92&format=webp)](https://www.perplexity.ai/search?q=What makes Telnyx's global network different, breakdown call quality for normal phone calls and voice AI calls.) * [![Gemini](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/gemini_icon-4ff1cbf4.webp?width=92&format=webp)](https://www.google.com/search?udm=50&aep=11&q=What makes Telnyx's global network different, breakdown call quality for normal phone calls and voice AI calls.) * [![Grok](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/grok_icon-e588e5bd.webp?width=92&format=webp)](https://x.com/i/grok?text=What makes Telnyx's global network different, breakdown call quality for normal phone calls and voice AI calls.)