Speechmatics Unveils Oak: The Multilingual AI Fixing Medical Transcription

CAMBRIDGE, UK – October 14, 2026 – The modern medical consultation is rarely a monolithic exchange of textbook English. In the United States alone, more than one in five people speak a language other than English at home. The clinicians treating them increasingly mirror this linguistic diversity, with recent studies indicating that nearly 72 percent of first-year resident physicians speak a second language.

This demographic shift has created a complex linguistic environment where doctors and patients frequently "code-switch"—seamlessly transitioning between languages mid-sentence. While this fluid communication builds patient trust and cultural rapport, it has become a critical vulnerability for the artificial intelligence systems tasked with documenting these high-stakes conversations.

Today, Cambridge-based voice AI developer Speechmatics released Oak, a specialized medical speech recognition model engineered to solve this precise problem. By accurately transcribing clinical conversations across 55 languages without requiring pre-set language configurations, Oak represents a significant leap forward in healthcare AI. More importantly, it highlights a crucial shift in how enterprise AI must be evaluated: moving away from generic benchmarks to focus on domain-specific accuracy that directly impacts patient safety.

Breaking the Language Barrier in High-Stakes Healthcare

For years, the artificial intelligence industry has struggled with the intersection of complex medical terminology and multilingual dialogue. General-purpose speech-to-text models often fail to recognize specialized drug names or diagnoses, while dedicated monolingual medical models hallucinate or drop words entirely when a patient suddenly switches to Spanish to describe their symptoms.

"Our mission is to understand every voice and the risk of getting that wrong in a medical environment is huge," said Katy Wigdahl, CEO of Speechmatics. "A consultation can move from Spanish to English and back in one sentence, with a drug name in the middle. That switch is the precise moment a clinician can least afford an error."

The clinical consequences of such errors are severe. A mistranscribed medication dosage or a misunderstood contraindication can lead to life-threatening adverse events. Oak was developed specifically to mitigate this risk. In internal evaluations, the model achieved a 91 percent medical term recall rate across English and ten additional languages.

The performance gap is particularly notable in Spanish, the most widely spoken non-English language in US homes with over 42 million speakers. During testing on Spanish medical consultations, Oak transcribed 96.3 percent of medical terms accurately, compared to an 86.8 percent average across eight competing providers. In practical terms, for every 100 drug names, conditions, and side effects spoken, Oak misses approximately four. Competing models miss an average of 13—more than three times the error rate.

Beyond the Word Error Rate Illusion

The release of Oak also spotlights a growing debate within the AI development community regarding how machine learning models are benchmarked. Historically, the industry standard for speech recognition has been Word Error Rate (WER). This metric calculates the percentage of words transcribed incorrectly, treating every word as equally important.

However, in a clinical setting, all words are not created equal.

"Evaluating a medical scribe AI using standard WER is fundamentally flawed," noted a clinical informatics specialist familiar with the technology's deployment. "Mistranscribing the word 'the' is a harmless typo. Mistranscribing 'amlodipine' as 'amoxicillin' puts a lethal error directly into a patient's electronic health record. We need metrics that measure clinical safety, not just general linguistic proximity."

Speechmatics has actively pushed back against the WER standard, choosing instead to evaluate Oak based on Medical Term Recall (MTR). This domain-specific metric focuses exclusively on the system's ability to correctly capture critical clinical terminology: pharmaceuticals, procedures, conditions, and side effects. By testing against AWS, Google, Microsoft, OpenAI, and specialized voice providers across 19 medical test sets, the company demonstrated that general models—even those with high overall accuracy—routinely fail when encountering complex medical vernacular mixed with background noise.

This shift from broad, context-agnostic metrics to highly specialized, entity-level evaluation is a trend that will likely redefine enterprise AI procurement. Business leaders and hospital administrators are increasingly demanding proof of value in their specific operational contexts, rejecting generalized leaderboards in favor of safety-critical benchmarks.

Proven in the Field: From the Olympics to the NHS

While Oak is a newly launched architecture, it builds upon a robust foundation of enterprise deployments that validate the company's approach to high-volume, mission-critical environments. The technology's ability to handle ambiguous medical pronunciations and rapid dialogue has already secured significant partnerships across the global healthcare ecosystem.

During the Paris 2024 Olympics, health-tech provider Humetrix utilized Speechmatics' underlying technology for multilingual patient intake, successfully processing clinical interactions for patients from 97 different countries. In the United Kingdom, tens of millions of NHS 111 triage calls are transcribed annually using Speechmatics through their partner Content Guru, demonstrating the infrastructure's capacity to handle massive, real-time audio processing at a national scale.

Furthermore, Sully.ai, a platform providing clinical scribes and AI receptionists to over 500 healthcare providers, selected Speechmatics to power its autonomous agents. The decision was reportedly driven by the system's superior handling of medication names in noisy, real-world clinical environments—a notorious pain point for automated scribing tools.

The Sovereign Edge: NVIDIA and the Privacy Imperative

Beyond accuracy, the adoption of AI in healthcare is heavily constrained by strict data privacy regulations, including HIPAA in the United States and GDPR in Europe. Hospitals are notoriously hesitant to send sensitive, unencrypted patient audio to third-party cloud servers for processing.

To navigate these regulatory hurdles, Speechmatics engineered Oak with highly flexible deployment architecture. While available in the cloud, the model can also be entirely self-hosted within a healthcare organization's private environment or deployed at the edge.

This on-premises capability is heavily supported by a strategic partnership with NVIDIA. Speechmatics utilizes the NVIDIA Triton Inference Server to deploy Oak, granting hospital IT departments granular control over where and how patient data is processed. To ensure that this localized processing does not sacrifice speed, the system leverages quantized NVIDIA CUDA kernels. This optimization technique reduces the computational footprint of the deep learning model, allowing it to process an hour of complex, multilingual audio in just 18 seconds without requiring massive, cloud-based server farms.

"The future of healthcare AI isn't just about building smarter models; it's about building models that can legally and securely operate inside the firewall of a hospital," explained a health-tech infrastructure analyst. "By combining high-speed edge compute with multilingual medical accuracy, you remove the two biggest bottlenecks to enterprise adoption: privacy concerns and latency."

As the healthcare industry continues to grapple with administrative burnout and diverse patient demographics, tools like Oak represent the next generation of business disruption. By addressing the nuanced realities of how humans actually speak, and securing the infrastructure required to process that speech safely, AI is moving out of the realm of generalized novelty and into the core workflows of life-saving industries.