Cartesia Ink-2 Claims to Be Built for Production. Here Is What Production Actually Requires.

On July 9, 2026, Cartesia launched Ink-2, a new speech-to-text model built for real-time voice agents. The headline numbers are strong: 8% word error rate on multi-accent call-center audio, outpacing Deepgram Flux at 10% and ElevenLabs Scribe v2 at 12% on the same benchmark. Ink-2 adds built-in semantic turn detection that reads meaning rather than silence to determine when a speaker has finished. Its time-to-final-transcript clocks at 0.1 seconds.

What makes Ink-2 worth examining is how Cartesia frames accuracy. Rather than measuring word error rate on clean studio audio, they sample live voice agent calls with non-native English speakers, background noise, and degraded audio from poor network conditions. Ink-2 reaches 6.5% WER on that dataset versus 9.2% for ElevenLabs Scribe v2 and 9.4% for Deepgram Flux. The methodology is more honest than most STT benchmark disclosures in the industry.

Producing reliable voice agents requires more than an accurate STT model. Production readiness means your accuracy holds on your domain vocabulary, your language mix, and your specific audio conditions, not a generic call-center benchmark. The output side of the pipeline, where TTS models speak to users, carries an entirely separate set of production requirements that no STT launch addresses.

"Built for production" and "ready for your production" are not the same claim. The gap between those two phrases is where most voice agent failures originate.

What Does Cartesia Mean by "Production Conditions"?

Cartesia's production benchmark samples generic call-center dialogue across 14 English accents under real-world audio conditions. This is a meaningful improvement over clean speech benchmarks and it reflects common failure modes in real deployments.

What it does not cover is your domain. The benchmark vocabulary is generic. It does not include medical terminology, legal entity names, financial instrument identifiers, proprietary product names, or the specific vocabulary your voice agent actually handles. A model that transcribes "I'd like to cancel my subscription" at 6.5% WER may transcribe "the contraindication for metformin at 500mg twice daily" or a 9-digit CUSIP at 15 to 20% WER. These are different environments, and a call-center benchmark does not reveal the difference.

Most voice teams discover this the first time a customer calls about a product category the model has never encountered. The benchmark WER is not wrong. It is measuring the wrong content for your use case.

Is Ink-2 Ready for Non-English Voice Agents?

Ink-2 launches as an English-only model. Cartesia states that multilingual support is on the way.

If your voice agents serve Spanish, Mandarin, Portuguese, Japanese, Korean, or any combination of non-English languages, Ink-2 is not yet viable as your primary STT layer. You still need a routing mechanism to direct non-English audio to a model that handles it. That routing logic is not something Ink-2 provides at launch.

For teams building globally, roadmap commitments introduce deployment risk. You cannot lock your infrastructure to a single STT provider that does not cover your full language surface. The multilingual gap also means that any quality baseline you establish in English today requires reestablishment and revalidation when multilingual support ships. Quality scores across language families do not transfer.

What Happens After Ink-2 Transcribes the Audio?

This is the gap the Ink-2 launch does not address, and it is where voice agent pipelines fail without triggering an alert.

Ink-2 converts speech to text. That text goes to an LLM for response generation. The LLM response goes to a TTS model: Cartesia's own Sonic, ElevenLabs, Deepgram Aura-2, or any of 100+ available engines. The TTS model generates audio. That audio ships to the end user.

At no point in that chain does any model validate its own output. The TTS model does not score the pronunciation of domain-specific terms. It does not flag when generated audio drifts from the established voice profile. It does not check format compliance against the telephony or playback environment. It does not lock its own model version so that audio generated at call 1 and call 100,000 came from the same model state.

Ink-2 may transcribe the input correctly. If the TTS output that follows is mispronounced, inconsistent, or format-noncompliant, the conversation still fails. The problem shifts from the input layer to the output layer. It does not disappear.

How Should Voice AI Teams Evaluate the Full Pipeline?

A voice agent has two audio quality surfaces: what it hears via STT and what it says via TTS. Ink-2 addresses the input surface with genuine engineering rigor. No model addresses the output surface natively.

Teams that focus exclusively on input accuracy treat output quality as an implicit assumption. That assumption holds in demos and breaks in production. The signals are predictable: brand name mispronunciations that go undetected until a customer complaint surfaces, voice drift across long-running agents as models update silently, format compliance failures when audio hits a telephony system expecting G.711 at 8kHz, and no audit trail when a regulatory review asks which model version produced a specific output.

Before committing to any STT provider, evaluate four things. First, measure domain accuracy on your content specifically, not on the provider's benchmark. Pull 500 real samples from your use case and measure WER on those. Second, confirm language coverage maps to your deployed languages today rather than a product roadmap. Third, test turn detection on your conversational patterns rather than on generic demos. Fourth, identify clearly where STT responsibility ends and TTS output responsibility begins, and who owns the gap.

How Does a Voice AI Production Layer Close This Gap?

Onepin is a voice workflow platform that orchestrates, validates, and ships production-ready audio across 100+ TTS models. It handles quality scoring, model version locking, format compliance, retry logic, and the per-output audit trail that regulated industries require. It sits above the model layer so that accuracy at transcription and accuracy at output are both measured rather than assumed.

Ink-2 raises the standard for what STT providers owe their customers. The investment in production-condition benchmarks and semantic turn detection moves the category forward in meaningful ways. Serious infrastructure thinking applied to the input side of voice agents is overdue.

The output side still needs the same seriousness. Every voice agent pipeline needs an owner for what ships to the user, not just what arrives from the user.

Start with Onepin to close the output validation gap before it becomes a customer complaint.