The Financial Times reported that OpenAI’s chatbot replied to some UK-based users’ English queries in Welsh while using its voice mode. A reporter for the newspaper also had a similar experience.
OpenAI told the FT the issue was caused by Whisper, its automatic speech recognition system, which sometimes gets confused.
A research paper published by the company in 2022 said Welsh was an “outlier with much worse than expected performance” and that it had about “9000 hours of translation data.” The paper also says the language identification system “misclassified” English audio as Welsh translation data.
ChatGPT users ran into a similar bug in February when it responded to questions in “Spanglish” — a combination of Spanish and English.
OpenAI first debuted its chatbot’s voice mode feature at its GPT-4o launch event in May and was set to release it in June. However, the company said in June it was delaying the rollout to improve its ability to “detect and refuse certain content.”
OpenAI then gave a small group with premium subscriptions access to the feature in July.
Taya Christianson, a spokesperson for OpenAI, told The Verge it will release the new mode to all of its Plus subscribers in the fall. It appears the company is being cautious about the release after its approach to safety was criticised by former employees.
Last week, OpenAI published a report on the safety research and frontier risks evaluations it carried out on GPT-4o ahead of its release.
One key risk it outlined was “unauthorized voice generation,” which included the ability for GTP-4o to “unintentionally generate an output emulating the user’s voice” in some “rare instances.”
OpenAI didn’t immediately respond to a request for comment from Business Insider, made outside normal working hours.
