A couple of studies evaluating scientific research (found here and here) published in 2024 concluded that many challenges remain in AI speech translation, such as shortcomings in nuance, prosody, and latency; lack of a single, comprehensive evaluation metric; and a scarcity of annotated training data.
But a lot of progress has been made, and AI speech-to-speech and speech-to-text translation are increasingly used in situations previously reserved for human interpreters or human voice-over actors, and money keeps pouring into AI startups that offer the synthetic version of both.
In January 2025, for example, ElevenLabs closed a Series C round of USD 180m, bringing the company’s valuation to USD 3-3.3bn. And there are quite a few more examples, like Pocketalk, EzDubs, Synthesia, and Lingopal.ai, which have also raised funds in the same time period.
It is remarkable then to find a data point showing that human interpreting is also growing, with LinkedIn placing interpreting as one of the fastest growing jobs in the UK in its Job Trends 2025 report at #22 out of 25.
We asked readers what they thought about this ranking, and two-thirds of respondents either believe there must be something to it (30.6%) or that it is unlikely but possible (30.6%). A little over a quarter are more skeptical (26.6%) and think it is impossible according to what they are seeing. The rest (12.2%) find it completely feasible and agree with the ranking.
Cast a Language on my Pod, STAT
Continuing on the subject of AI speech tech, ElevenLab’s Dubbing Studio was used for an episode of the Lex Fridman podcast featuring Ukrainian President Volodymyr Zelenskyy. It was also used for a five-minute segment of a recent SlatorPod episode, and Slator peeps found the voice cloning and AI translation to be quite good.
YouTube, a hub for podcasts, officially deployed automated multilingual dubbing for hundreds of thousands of channels. At publication date, the feature allows creators to automatically dub recorded English audio into and from French, German, Hindi, Indonesian, Italian, Japanese, Portuguese, and Spanish.
Google, which owns YouTube, also released NotebookLM in September 2024, a large language model (LLM) interface that, among other things, allows users to create synthetic podcasts based on one or more content sources.
AI-generated podcasts are an impressive technology, but non-English versions will need to wait a bit longer, at least on NotebookLM, according to Kelly Schaefer, Director of Product and Domain Lead at Google Labs. Schaefer told TechCrunch in February 2025 that the company was thinking about how to prioritize the languages and make sure they feel “really genuine and just as seamless and natural as our current Audio Overviews do.”
We asked readers if they think podcasts will become the key use case for AI dubbing, and opinions are quite divided. A third (33.3%) see it as a possibility, while close to a quarter (24.2%) see it as unlikely. The rest of the respondents are represented in two equal groups (18.2% each) who either think it will definitely be the case or that it will be so in the short term, and a small cohort (6.1%) that does not think it will happen at all.
Appetite for Foundation LLMs
While venture capitalists continue pouring billions into promising AI startups, established companies hurry to develop, adopt, or partner with others to incorporate AI into their operations or offerings, including translation-as-a-feature. What about public institutions? Some of them are also partnering with others, as is the case with the European Union (EU).
The European Commission (EC) is in fact collaborating with 20 institutions, including startups, labs, and supercomputing centers, to develop a set of foundation models known as the OpenEuroLLM.
The EC has initially allocated EUR 37.4m (USD 39.4m) to the program. The expectation is that more funding will follow in the organization’s quest for Europe’s “digital sovereignty” and open source availability by making the program a part of the Strategic Technologies for Europe Platform (STEP).
We asked readers if they believe the EU should be funding the building of LLM Foundation models, and for the majority (52.4%) the answer is “Yes, of course.” Close to a third of respondents (28.6%) believe this endeavor should be left to private companies. The rest either think it is a possibility (16.7%) or a [negative] probability (2.3%).
Vintage Procurement
AI automation is already shortening cycles in all sorts of processes, absorbing and connecting tasks and workflows into seemingly invisible sequences via orchestration. AI procurement systems have existed for some time, but public institutions like the EU and the UK’s National Health Service (NHS) are still going about RFPs the old-fashioned way.
Public procurement still largely relies on the old and trusted request for proposals (RFP) cycle, with its complex, long, and tedious spreadsheets, forms, and supporting documentation requirements. For example, the NHS published a six-section tender for USD 90m for translation and interpretation services in January 2025.
Coveted, multimillion-dollar long-term contracts are great. Who would not like the assurance of capital over multiple years? But do people actually like RFPs? Most Slator readers (55.6%) do not really like them, but understand they are part of the game. About 1 in 5 (22.2%) just hate them. Some readers (13.3%) would rather craft a bespoke solution, and the smallest group (8.9%) said that of course they like it, winning is great.