Cartesia · Indexed content

Learn.Compare speech tools, understand the tradeoffs, and build voice applications.Magnifying Glass Streamline Icon: https://streamlinehq.comWord error rate: what WER misses in voice AILearn what word error rate measures, why it can mislead TTS evaluations, and how to test the speech quality that matters in a live conversation.Sep 18, 2026Conversational IVR: how to replace phone menusLearn how conversational IVR routes spoken requests, where voice agents fit, and what to test before replacing a phone menu.Sep 17, 2026What a frozen acoustic judge changes about ASR hallucinationA new speech-recognition study reduces insertion errors on meeting audio. Its results show why error types deserve attention alongside overall word error rate.Sep 16, 2026AI call center: how to pilot voice agentsPlan an AI call center pilot: choose a support queue, connect tools, test human handoffs, and measure resolution and cost before expanding.Sep 16, 2026How to compare TTS APIs using endpoint-level benchmarksCompare TTS APIs using endpoint-level evidence. A closer look at Coval's Qwen3 entries shows what latency numbers can tell you and how to test a shortlist for production.Sep 15, 2026AI receptionist: …

Word error rate (WER) counts differences between a transcript and a reference. It can help detect missing or incorrect words. It cannot tell you whether a voice is good for your product. Two generated clips can say exactly the same words and still create very different experiences. One sounds reassuring; the other sounds delighted that your flight was canceled. The transcript passes. The conversation does not. Our approach, explained in Is this TTS model good?, starts with the experience we want people to have. For real-time voice agents, that means understanding the response the first time, hearing delivery that fits the situation, and having a conversation without awkward delays or interruptions. WER is one diagnostic within that evaluation, not a score to optimize in isolation. What word error rate actually measures WER uses the minimum number of word edits needed to align a transcript with its reference: WER = (substitutions + deletions + insertions) / reference words Multiply by 100 to express it as a percentage. In this example, the transcript substitutes “Thursday” for “Tuesday”: Reference: please move my appointment to next Tuesday morning Transcript: please move my appoint…

Conversational IVR lets callers say “I need to change my delivery address” instead of working through “press 1 for sales, press 2 for support.” It uses speech recognition and language understanding to choose a route or an approved action. Replacing the menu helps only if callers reach the right place with less effort. A pleasant voice that sends them to the wrong queue has not fixed the problem. This guide is for engineers and contact-center owners replacing an inbound phone menu. It covers the routing decision, ambiguity, keypad fallbacks, and a test plan. If you already know which queue needs a task-completing agent, use our AI call center pilot guide. How conversational IVR works IVR stands for interactive voice response. A conventional system collects input, applies routing rules, and either plays an answer or transfers the call. That input can already include speech: Twilio’s Gather verb, for example, accepts keypad digits, speech, or both. Recognizing a spoken digit alone does not make a menu conversational. In a conversational flow, the caller can describe the problem without knowing the company’s department names. The system transcribes the speech, interprets the request, a…

ASR hallucination can put words into a transcript that the speaker never said. A new paper, Grounded in Sound, tests whether acoustic feedback during training can reduce those insertion errors. Its results on meeting audio make a useful case for measuring insertions separately from overall word error rate.What changed in the ASR hallucination results?The authors report relative insertion reductions of 28.3% on close-talking AMI-IHM meeting speech and 22.3% on far-field AMI-SDM. On AMI-SDM, word error rate fell from 35.89% to 34.71%. The other five evaluation tiers showed no statistically detectable WER difference against the schedule-matched WER-only baseline.Those percentages describe different measurements. A 22.3% reduction in insertions is not a 22.3% reduction in all transcription errors. The far-field WER improvement was 1.18 percentage points. A product evaluation that records only total WER would miss much of the change in error composition.The scope is specific. The experiments use a Qwen2-Audio-7B policy adapted with LoRA and a frozen 0.3B wav2vec2-CTC judge. The authors describe a mechanism demonstrated in one model setup. The results do not establish equivalent gains fo…

An AI call center uses AI to answer customer phone calls, assist human representatives, or review conversations after they end. If you want an AI voice agent to resolve support requests, start with one queue and a result you can verify in your backend. Keep a route to a person. Expand only after the agent passes the same task and permission checks you expect from your support team. This guide is for contact-center operations leads and engineers adding voice agents to an existing support operation. It covers a pilot for order-status calls, including a scorecard you can adapt. For a small business choosing an answering and scheduling service, start with our AI receptionist guide. Choose which part of the call center needs AI “AI call center” can describe several different products. Separate them before comparing vendors. Need What to evaluate What still needs an owner Answer calls and complete routine tasks A voice agent with access to approved tools Tool permissions, customer verification, escalations Help a human representative during a call Transcription and agent-assist software The representative’s workflow and approval of suggestions Find problems in completed calls Conversatio…

A useful TTS API comparison starts with the endpoint your application will call. A model-family name leaves too much unspecified: the exact model, hosting arrangement, region, voice, and measurement method can all differ between entries on a leaderboard.Coval's Qwen3 TTS entries illustrate the problem. They share a family name, but the published endpoint descriptions do not establish an experiment in which identical weights run under otherwise identical conditions. The practical lesson is to compare complete services and keep the limits of the evidence visible.What do the Qwen3 TTS benchmark entries actually compare?The figures below are the 30-day average TTFA values displayed together on Coval's TTS ranking when viewed on September 15, 2026. The linked model pages supply the hosting and licensing details. These are measurements of separate endpoints, not a controlled test of identical systems.Qwen3 TTS Fast, hosted by Nari, reports 71 ms mean TTFA. Coval lists it as open-weight, shared inference in the US.Qwen3 TTS 1.7b, hosted by Baseten, reports 106 ms mean TTFA. Its entry describes open-weight, dedicated inference in the US.Qwen3 TTS Flash Realtime, hosted by Alibaba Cloud, re…

An AI receptionist answers inbound phone calls and handles tasks such as checking business hours, routing callers, or booking appointments. Buy a ready-made receptionist if it already works with your scheduling software. Build on a voice-agent platform if your call flows require custom business rules or connections to your own systems. This guide is for the team responsible for that choice: an operations lead defining what calls should accomplish, and a developer responsible for the systems behind them. Cartesia Managed Agents is a platform for building the agent. It is not a preconfigured receptionist with every industry’s scheduling integration included. What an AI receptionist needs to do Start with a narrow job. An after-hours agent that answers approved questions and collects callback requests has fewer ways to fail than one that can reschedule appointments, take payments, and change customer records. Write down which actions the agent can complete and which require a person. For example: Caller request What the agent needs Safe fallback “When do you close?” Approved hours, including holiday exceptions Say the hours are unavailable and offer a callback “Can I come in tomorrow?…

To turn text into speech in JavaScript, create a SpeechSynthesisUtterance and pass it to window.speechSynthesis.speak(). The browser handles playback. You don’t need an API key, a server, or a JavaScript package. This works well for a prototype or a read-aloud control where device-provided voices are acceptable. If you’re building a voice agent that needs the same voice across devices or audio streamed to a phone call, use a hosted text-to-speech API instead. Browser SpeechSynthesis and Cartesia’s API are separate ways to generate speech. How browser speech synthesis works window.speechSynthesis is the Web Speech API’s controller for synthesized speech. An utterance holds the text and settings for one request. Calling speak() adds the utterance to a queue; it doesn’t return an audio file.1 The browser exposes the voices available on the current device through getVoices(). That list can load after your script runs, so read it immediately and refresh it when voiceschanged fires.2 Don’t hard-code a voice name from your laptop and expect it to exist on a customer’s phone. 1. Save a working text-to-speech demo Save the following as speech.html and open it in a browser. Enter a short sen…

I think Coval's streaming STT benchmark is useful because it puts the inference stack back into the conversation, where conversational products actually live. If you only compare model weights, you miss the part of latency that users feel most: the deployment path, the region, and the endpoint you put behind the model.What Coval's streaming STT benchmark actually measuresAs checked on September 14, 2026, Coval's WER page lists 30 STT models and 28 TTS models over a rolling 30-day window. The STT table reports word error rate alongside transcript latency and sample counts. Some rows have tens of thousands of samples; others have far fewer. Those differences matter when interpreting a ranking. Coval WER leaderboardThe page also exposes more than one latency view, including Time to Final Segment, Time to First Token, and Time to Final Transcript. That is useful because different products care about different moments: a live captioning tool cares about final transcript speed, while a voice agent cares about when it can safely start speaking again. Coval benchmarks overview Coval WER leaderboardCoval runs the public benchmark from a fixed region. That makes the comparison cleaner, becau…

To convert text to speech in Python, send a transcript to a speech model and save the returned audio. This guide uses Cartesia’s Python SDK to generate a WAV file with Sonic. You’ll need Python 3.9 or later, an internet connection, and a Cartesia account with API access. The example below targets SDK version 4.2.0.1 It calls a hosted API: your text leaves your machine. If you need offline synthesis, use a local speech engine instead. Install the Python SDK Create a virtual environment in your project directory: python -m venv .venv On macOS or Linux, activate it with: source .venv/bin/activate On Windows PowerShell, use: .venv\Scripts\Activate.ps1 Then install the version used in this guide: python -m pip install "cartesia==4.2.0" If your system uses python3 rather than python, use that command to create the environment. Pinning the SDK keeps the example’s method names stable. When you upgrade, check the SDK release notes for changes. Set your API key Create an API key in the Cartesia Playground. Give it the permissions needed for text-to-speech generation. Set the key in the terminal where you’ll run Python. On macOS or Linux: export CARTESIA_API_KEY="your-api-key" On Windows Powe…

Cartesia blog[Product]TTS Model vs. Voice: They're Not the SameZubin Pratap Sep 10, 2026AllEngineeringResearchNewsProductAllEngineeringResearchNewsProductTailless Line Arrow Down 1 Streamline Icon: https://streamlinehq.comMagnifying Glass Streamline Icon: https://streamlinehq.comWhy Precision / Recall Matters So MuchAccuracy sounds like the right metric until your model scores 90% by doing nothing. Here's why precision and recall tell a more honest story — and why it matters for voice AI.[Product]Sep 3, 2026Introducing Sonic-3.6Sonic-3.6 makes huge strides compared to Sonic-3.5 in naturalness and quality, with listeners preferring it in up to 93% of blind head-to-head tests across fifteen locales.[Product]Aug 27, 2026New to Ink-2: keyterm prompting and configurable turn detectionTwo new Ink-2 features: keyterm prompting, which improves transcription accuracy on domain-specific entities, and configurable turn detection, which tunes endpointing for speed or accuracy.[Product]Aug 11, 2026Is this TTS model good?A practical look at why text-to-speech evaluation breaks down as models improve, and how to measure what really matters.[Research]Jul 28, 2026Introducing Ink-2: The #1-ranked ST…

A text-to-speech (TTS) model generates speech. But the model and the voice it speaks with are not the same thing. If you know LLMs, a voice is loosely analogous to a system prompt: the model supplies the general capability, and the input shapes what it generates. In TTS, voice is an input often separate from the model, and it shapes the model’s response. And the model’s response - audio - is much more complex than “mere” text. Models that generate speech had to know what should be said, how it should be said, at what pace, how loud, what sort of intonation, prosody, emphasis…all the many characteristics and metrics for a “good voice”. These are instructions you’d normally want to give the model, but most TTS models allow you to specify one or two inputs - the transcript and a selected voice. Voice Model Inputs A fixed speaker model takes a single input - the transcript. The model is trained to speak in a single voice (encoded in the model’s learned weights). It does not take a voice as an input. So for all transcripts supplied, it returns audio that sounds like the same speaker every time. > Fixed-speaker model + Transcript = speech in one voice When the only voice you hear is the …

BlogZubin PratapArticles by Zubin PratapMagnifying Glass Streamline Icon: https://streamlinehq.comTTS Model vs. Voice: They're Not the SameA TTS model generates speech. But the model and the voice it speaks with are not the same thing — and confusing them leads to poor voice AI design decisions.[Product]Sep 10, 2026Why Precision / Recall Matters So MuchAccuracy sounds like the right metric until your model scores 90% by doing nothing. Here's why precision and recall tell a more honest story — and why it matters for voice AI.[Product]Sep 3, 2026A Guide to Choosing Voice AI ModelsA practical framework for evaluating ASR, TTS, and turn-detection models against your real-world use case — not lab conditions.[Product]Jun 30, 2026A Beginner's Guide to Voice AI Terminology A plain-language glossary of the 20 key terms you need to understand, evaluate, and build conversational voice AI.[Product]Jun 30, 2026Get started todayTalk to an expert. Connect with a member of our team and learn how Cartesia can help you build world-class voice experiences.Contact SalesStart building. Access our models via API and bring a voice agent into production in minutes.Try Cartesia{"@context":"https://schema.o…

Two ways to be wrong Precision vs recall is obvious if you evaluate ML models for a living. For everyone else, not so much. It’s best understood with an example. Let’s say you’re building a RAG pipeline. And you’re really trying for highly reliable retrieval from your vector store. One of the first questions: how many chunks should you fetch (your top_k)? If you set it to fetch 3 chunks they’re all likely to be highly relevant. But you might miss relevant chunks ranked 4th or 10th. That means relevant material got left out. You could try and fix this by increasing your top_k to 50. With that setting you catch everything relevant, but there was also plenty of irrelevant stuff which increases the odds that the LLM returns “noisy” info that has some non-relevant information, even though you retrieved all the relevant chunks from your vector store. This is a trade off, and it’s referred to as Precision vs Recall. The top-3 retrieval has high precision because what you got back was high relevance. It was on point. But it also has low recall because you missed stuff. On the other hand, the top-50 retrieval has high recall (you caught everything) but lower precision (you caught a lot of j…

At Cartesia, we’re deeply invested in making foundational advances to AI architectures and algorithms, because that’s what drives step changes in model capabilities. Sonic-3.6 is the culmination of our latest advances in architectures that learn efficiently from multilingual audio across pre- and post-training. Sonic-3.6 makes huge strides compared to Sonic-3.5 in naturalness and quality, with listeners preferring it in up to 93% of blind head-to-head tests across fifteen locales. At the start of this year, we made a bet that instead of incrementally adjusting the existing paradigm, the way to do the best research in the world was to rethink everything from first principles. Since then, we’ve rebuilt our data, model architecture, training and evaluation from the ground up with novel ideas, and we’re seeing those efforts pay off. Sonic-3.6 lands just two months after Sonic-3.5, and it once again takes #1 on the Artificial Analysis leaderboard across both the controlled and provider voice boards. On the controlled board, the incumbent model we beat is Sonic-3.5. Overtaking our own model while no other provider has closed the gap is a testament to our accelerating research velocity. C…

Last month we released Ink-2, the most accurate streaming STT model for English. Today, we’re releasing two new features: keyterm prompting, which improves transcription accuracy on complex domain-specific entities, and configurable turn detection, which tunes endpointing for speed or accuracy. With these improvements, agents can transcribe names and technical terms more reliably, and configure latency and endpointing accuracy. Keyterm prompting: boost transcription accuracy on difficult entities astro-island,astro-slot,astro-static-slot{display:contents}(()=>{var a=(s,i,o)=>{let r=async()=>{await(await s())()},t=typeof i.value=="object"?i.value:void 0,c={rootMargin:t==null?void 0:t.rootMargin},n=new IntersectionObserver(e=>{for(let l of e)if(l.isIntersecting){n.disconnect(),r();break}},c);for(let e of o.children)n.observe(e)};(self.Astro||(self.Astro={})).visible=a;window.dispatchEvent(new Event("astro:visible"));})();(()=>{var g=Object.defineProperty;var w=(c,s,d)=>s in c?g(c,s,{enumerable:!0,configurable:!0,writable:!0,value:d}):c[s]=d;var l=(c,s,d)=>w(c,typeof s!="symbol"?s+"":s,d);var E=new Set(["__proto__","constructor","prototype"]);{let c={0:t=>y(t),1:t=>d(t),2:t=>new RegEx…

BlogLisa BedinArticles by Lisa BedinMagnifying Glass Streamline Icon: https://streamlinehq.comNew to Ink-2: keyterm prompting and configurable turn detectionTwo new Ink-2 features: keyterm prompting, which improves transcription accuracy on domain-specific entities, and configurable turn detection, which tunes endpointing for speed or accuracy.[Product]Aug 11, 2026Get started todayTalk to an expert. Connect with a member of our team and learn how Cartesia can help you build world-class voice experiences.Contact SalesStart building. Access our models via API and bring a voice agent into production in minutes.Try Cartesia{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Cartesia","item":"https://www.cartesia.ai/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https://www.cartesia.ai/blog"},{"@type":"ListItem","position":3,"name":"Lisa Bedin","item":"https://www.cartesia.ai/authors/lisa-bedin"}]}{"@context":"https://schema.org","@type":"ProfilePage","mainEntity":{"@type":"Person","name":"Lisa Bedin","url":"https://www.cartesia.ai/authors/lisa-bedin","worksFor":{"@type":"Organization","name":"Cartesia","url":"…

BlogEli PughArticles by Eli PughMagnifying Glass Streamline Icon: https://streamlinehq.comNew to Ink-2: keyterm prompting and configurable turn detectionTwo new Ink-2 features: keyterm prompting, which improves transcription accuracy on domain-specific entities, and configurable turn detection, which tunes endpointing for speed or accuracy.[Product]Aug 11, 2026Introducing Ink-2: The #1-ranked STT built for voice agentsWe're excited to release Ink-2: our speech-to-text model built for real-time voice agents, ranked #1 on Artificial Analysis's streaming leaderboard with the most accurate built-in turn detection of any provider.[News]Jul 9, 2026Get started todayTalk to an expert. Connect with a member of our team and learn how Cartesia can help you build world-class voice experiences.Contact SalesStart building. Access our models via API and bring a voice agent into production in minutes.Try Cartesia{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Cartesia","item":"https://www.cartesia.ai/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https://www.cartesia.ai/blog"},{"@type":"ListItem","position":3,"name":"…

Speko is a router for voice AI. Developers build and run voice agents through one Speko API, and Speko picks the speech-to-text, LLM, and text-to-speech models for each call based on its language and use case. Those picks come from a benchmark Speko re-runs every two to three days across roughly a dozen providers. On the text-to-speech side, the model that comes out on top most often is Cartesia’s Sonic 3.5. “We are building an open router of voice AI basically. We are trying to be this neutral layer where we help people to pick the right model stack.Beknazar AbdikamalovFounder, Speko The challenge: three model decisions and nobody neutral to ask Roughly 95% of voice AI runs cascaded: speech-to-text, then an LLM, then text-to-speech. That is three separate model decisions per agent, and most of the providers with an opinion on them have a stake in the answer. New providers keep arriving, so there are more options every month without more clarity about which ones hold up. It gets harder outside English. A model that reads well in English can drift or flatten in another language, and there is very little public data on which ones do. Companies large enough to run their own evaluation…

George Rybintsev built ActOnCue for actors recording auditions at home. An actor uploads a script, picks the character they are playing, and rehearses against AI voices reading the other parts. Cartesia is the default provider for those voices, and Sonic 3 is what moved ActOnCue off ElevenLabs. The challenge: the solo job that needs a second person Self-tapes were a pandemic workaround that stuck. Casting still runs on them, which leaves actors to produce their own auditions. A self-tape is the only solo endeavor that requires a crew. One person can run the camera and read the second part, but that doubles the effort, the time, and the editing. In practice, a self-tape means calling your roommate at 10pm and asking them to play Juliette. ActOnCue takes the second person out of the equation without handing the work back to the actor. Drop in a script, select your character, and start rehearsing. The platform finds the remaining characters and assigns voices to them automatically, and any of those voices can be changed after the first run. An actor should not have to configure an AI system before an audition. “You just want to go from A to B really fast and really efficiently.George …

Most of our interactions with digital systems still happen through typing into boxes or tapping on screens. It works, but it is not exactly frictionless. Voice can make that interaction feel faster, more natural, and closer to the way people already communicate. If we can make it work well, we can unlock a more interactive way to learn, get help, move across languages, and get things done more seamlessly. When people interact through voice, the generated speech becomes the experience. Small differences in delivery can change whether the interaction feels helpful, natural, or broken. That is what makes text-to-speech such an important piece of the stack: it turns written input into spoken output, and it shapes the feel of the entire interaction. Voice, though, is not a one-size-fits-all experience. For example, when someone says an audio generated by TTS model is “good,” they might mean one or several things: Check Circle Streamline Icon: https://streamlinehq.comCorrectnessIs every word spoken correctly?Medal Badge Star Streamline Icon: https://streamlinehq.comAudio qualityIs the audio free of static, background noise, or distortion?Brain Streamline Icon: https://streamlinehq.comCor…

BlogAparajita SarafMachine Learning EngineerArticles by Aparajita SarafMagnifying Glass Streamline Icon: https://streamlinehq.comIs this TTS model good?A practical look at why text-to-speech evaluation breaks down as models improve, and how to measure what really matters.[Research]Jul 28, 2026Get started todayTalk to an expert. Connect with a member of our team and learn how Cartesia can help you build world-class voice experiences.Contact SalesStart building. Access our models via API and bring a voice agent into production in minutes.Try Cartesia{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Cartesia","item":"https://www.cartesia.ai/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https://www.cartesia.ai/blog"},{"@type":"ListItem","position":3,"name":"Aparajita Saraf","item":"https://www.cartesia.ai/authors/aparajita-saraf"}]}{"@context":"https://schema.org","@type":"ProfilePage","mainEntity":{"@type":"Person","name":"Aparajita Saraf","url":"https://www.cartesia.ai/authors/aparajita-saraf","worksFor":{"@type":"Organization","name":"Cartesia","url":"https://www.cartesia.ai/"},"jobTitle":"Machine Learni…

We’re excited to release Ink-2: a speech-to-text model built for real-time voice agents. It’s ranked #1 on Artificial Analysis’s streaming leaderboard for lowest word error rate†, with the most accurate built-in turn detection of any provider, so the model knows precisely when to listen and when to respond. For voice agents, speech-to-text has to get three things right: accuracy, turn detection, and latency. If any one of these falls short, the experience breaks down. The agent may misunderstand the user, interrupt at the wrong time, respond too slowly, or make the conversation feel unnatural. Ink-2 was built to lead on all three. Accuracy: Getting every word right We’ve done extensive work on structured entity recognition like phone numbers, email addresses, alphanumerics, and dates. Ink-2 understands when it’s mid-entity and waits for the full sequence before committing, with no special prompting needed. We built Ink-2 to be robust across a range of accents, which reflects real voice agent calls. On AppTek, a multi-accent benchmark spanning 14 English accents on real call-center dialogue, Ink-2 is the strongest streaming STT provider at 8% WER, vs. 10% for Deepgram Flux and 12% f…

Voice AI Concepts You Must Know When your voice agent interrupts the user at the wrong moment, is that a Speech to Text (STT) problem? Or is that a Voice Activity Detection (VAD) problem? Or could it be that your Text to Speech AI model wasn’t stopped mid-generation? They’re three different things, and the fix is different for each. They even live in different parts of your pipeline. That’s why building truly conversational voice AI agents is much harder than just picking a STT or TTS model. At Cartesia, we are building AI for humans to express themselves, and be understood. Understanding starts with a common vocabulary. This blog gives you the vocabulary you need to understand, explore and reason about Conversational AI Agents, and it will help you understand the benchmark results that seem to get released every few days! I’ve grouped the terms into sections so they’re easier to contextualize. You know what they say - context is everything (especially in AI!). Core Voice Pipeline STT (Speech-to-Text). This is the conversion of speech input into text output. In practice, STT and ASR (Automatic Speech Recognition) are often used interchangeably. Any time you dictate a text message i…

Choosing Models for Enterprise Voice AI Many teams assume building an AI Voice Agent is as easy as picking a Text-To-Speech (TTS) voice that “sounds professional,” hooking it up to their existing chat bot, wiring it to an LLM and adding some Automatic Speech Recognition (ASR) to kick it all off. After a few months they ship a prototype that benchmarks at 600–800ms round-trip in a quiet room on macOS. But then in real life, this system hits 2 - 4 seconds on real telephony infrastructure, with P95 (95th percentile latency – the maximum latency number that 95% of calls experience) spiking to 5 seconds. Add to this lower quality that comes with 8khz telephony infrastructure, and the “IRL” experience of voice agents can be starkly different from eval conditions. It is tempting to choose the models based on benchmarks that measure them in laboratory conditions. But real-life conversations are messy, and that makes enterprise-grade voice agents a hard engineering problem. We cannot solve these hard problems by treating voice models as interchangeable commodities. Instead, they must be picked, analysed and measured against their performance for their intended use case. The same model that …