SlatorPod

In the 2025 year-end episode of SlatorPod, hosts Florian Faes and Esther Bond reflect on a year defined by rapid AI investment, shifting policy, and structural change across the language industry.Esther opens the year-in-review by highlighting January’s twin funding milestones in the language AI and product space. Florian follows with February, which saw hyperscalers and AI labs release data highly relevant to the way AI translation is being used.March, April, and May saw major developments both on the regulatory side and in terms of bolt-on acquisition deals.Past the mid-year point, OpenAI’s decision to hire a localization manager was what grabbed the industry’s collective attention. The AI lab’s decision contrasted with September’s news, which saw the closure of one of the world’s most recognized academic programs for localization.The year closed on publicly listed LSIs releasing mixed results and major announcements in AI translation for literature and live speech translation rollouts.The duo closes with 2026 predictions!

Josh Miller and Christopher Antunes, Co-Founders and co-CEOs of 3Play Media, join SlatorPod to talk about the company’s trajectory as a leading language solutions integrator (LSI) in multilingual video accessibility.The duo explains how the two met at MIT, where an early challenge from OpenCourseWare revealed that captioning thousands of technical videos was financially impossible, leading to the company’s founding, where they developed proprietary tooling, leveraged AI, and incorporated expert-in-the-loop solutions.Josh describes how their platform evolved into a dual system supporting both customers and large-scale operations. Chris notes that the LSI now serves media and entertainment, higher education, e-learning, and corporate clients.Chris explains that three major trends — the European Accessibility Act (EAA), advancements in voice technology, and the rise of live events — drove their expansion into global localization. The co-CEOs detail their dubbing journey, noting rapid learning over the last 18 months and the emergence of a big mid-tier market between high-end theatrical dubbing and low-cost AI-only output.Josh explains how the EAA is pushing companies to prepare for large columns of multilingual captioning and audio description. He notes that interpretations of the law still vary, but major media firms are already investing to avoid disruption. The duo shares findings from their 2025 State of ASR Report, where they found that accuracy initially improved sharply with generative models but has now plateaued.Looking to the future, the co-CEOs are working on shifting their model to incorporate AI-generated scores and analytics, allowing customers to decide on the level of expert intervention.

Florian and Esther discuss the language industry news of the past few weeks, reflecting on SlatorCon Remote and announcing that SlatorCon London 2026 is open for registration.The duo touch on IMDb’s decision to recognize dubbing artists as part of new professional credit categories, explaining how this expands visibility for multilingual voice talent. They then move on to Coursera’s strategy shift and outline how its new CEO is betting on AI translation and AI dubbing to revive slowing growth. Florian and Esther talk about Amazon’s rollout of AI-translated Kindle eBooks, and question authors' willingness to rely on automated translation despite Amazon’s promise of fast turnarounds, in as little as 72 hours.Florian highlights research on spatial audio improving AI live speech translation, and reflects on how clearer speaker differentiation could enhance comprehension. Although he stresses ongoing challenges in live settings, like latency and overlapping speech.In Esther’s M&A and funding corner, healthcare AI technology startup No Barrier raises USD 2.7m, Cisco acquires EZ Dubs to enhance WebEx’s real-time speech translation capabilities, and audio AI startup AudioShake raises USD 14m. Florian analyzes OneMeta’s financials and notes its rapid revenue growth despite significant ongoing and limited marketing presence. Esther details the landmark UK NHS framework agreement for language services, including scope and the number of awarded vendors.Florian concludes with updates on interpreting performances at Teleperformance and AMN Healthcare, noting mixed results.

Amir Haramaty, Co-Founder and President of aiOla, joins SlatorPod to talk about how spoken, multilingual data can transform enterprise workflows and unlock real ROI.The Co-Founder introduces himself not as a serial entrepreneur but as a serial problem solver, focused on one core challenge: most enterprise data remains uncaptured, unstructured, and unused.Amir emphasizes that traditional speech tech fails in real-world conditions, where accents, noise, and hyper-specific jargon dominate. He illustrates how he tackles this challenge by building workflow-specific language models that extract only the data relevant to a process.Amir says aiOla converts speech not into text but into structured, schema-ready data, allowing organizations to automate workflows, improve compliance, and identify trends long before humans can. He explains that the company focuses on narrow processes rather than general conversation, enabling precision in niche environments.Amir shares how aiOla routinely cuts multi-hour procedures down to minutes, drives efficiency across frontline roles, and creates previously unavailable datasets that feed enterprise intelligence. He highlights ROI examples from supermarkets, airlines, manufacturing, and automotive industries.Amir explains that after proving aiOla’s value, he realized the fastest way to scale was through firms already embedded in enterprise digital transformation. He notes that aiOla now partners with UST, Accenture, Salesforce, and Nvidia, creating a distribution engine capable of replicating wins across thousands of clients. He calls this channel strategy a force multiplier that shortens sales cycles and embeds aiOla inside broader modernization initiatives. Amir adds that these partners not only bring scale but also domain expertise aiOla deliberately chose not to build in-house. Amir outlines future priorities, including product-led growth, speech-based coding, and speech-prompted AI agents. He predicts that agentic systems will rely heavily on high-quality spoken data, making aiOla’s role even more central.

Tom Kocmi, Researcher at Cohere, and Alon Lavie, Distinguished Career Professor at Carnegie Mellon University, join Florian and Slator language AI Research Analyst, Maria Stasimioti, on SlatorPod to talk about the state-of-the-art in AI translation and what the latest WMT25 results reveal about progress and remaining challenges.Tom outlines how the WMT conference has become a crucial annual benchmark for assessing AI translation quality and ensuring systems are tested on fresh, demanding datasets. He notes that systems now face literary text, social-media language, ASR-noisy speech transcripts, and data selected through a difficulty-sampling algorithm. He stresses that these harder inputs expose far more system weaknesses than in previous years.He adds that human translators also struggle as they face fatigue, time pressure, and constraints such as not being allowed to post-edit. He emphasizes that human parity claims are unreliable and highlights the need for improved human evaluation design.Alon underscores that harder test data also challenges evaluators. He explains that segment-level scoring is now more difficult, and even human evaluators miss different subsets of errors. He highlights that automated metrics built on earlier-era training data underperformed, particularly COMET, because they absorbed their own biases.He reports that the strongest performers in the evaluation task were reasoning-capable large language models (LLMs), either lightly prompted or submitted with elaborate evaluation-specific prompting. He notes that while these LLM-as-judge setups outperformed traditional neural metrics overall, their segment-level performance varied.Tom points out that the translation task also revealed notable progress from smaller academic models around 9B parameters, some ranking near trillion-parameter frontier models. He sees this as a sign that competitive research is still widely accessible.The duo concludes that they must carefully choose evaluation methods, avoid assessing models with the same metric used during training, and adopt LLM-based judging for more reliable assessments.