What's Happening?
The Gemini Live API, specifically models/gemini-3.8-live, is facing issues with Finnish input transcription, as reported by developers of the SuomiNavi+ navigation app. Despite setting `generationConfig.speechConfig.languageCode` and `inputAudioTranscription.languageCodes`
to "fi-FI" since October 12, 2026, Finnish speech is frequently transcribed incorrectly, sometimes in other languages. An example from November 13, 2026, showed the Finnish phrase "Lopeta navigointi" ("stop navigation") being transcribed as "la patata la vengo a prendere," leading to the model responding, "Sorry, I didn’t catch that." Measurements conducted between November 9-20, 2026, across 92 recorded sessions and 134 user turns, revealed that Gemini's input transcription had wrong meanings in 29% of turns (37% with CarPlay microphones), significantly higher than Apple's on-device fi-FI speech recognizer, which showed a 7% error rate (8% via CarPlay).
Why It's Important?
This issue with the Gemini Live API has significant implications for applications relying on accurate multilingual speech-to-text capabilities, particularly in regions with less common languages like Finnish. For developers, the unreliability of transcription directly impacts user experience and the functionality of their products, as demonstrated by the SuomiNavi+ app. The discrepancy in performance between Gemini Live and Apple's on-device recognizer highlights a potential gap in Google's AI language processing for specific languages. This could lead to reduced adoption of Gemini Live for applications requiring precise voice commands or interactions in non-English languages, potentially ceding market share to competitors with more robust multilingual support. Businesses and developers investing in AI-powered voice interfaces need consistent and accurate transcription to ensure their services are accessible and effective for a global user base.
What's Next?
Developers are seeking clarification on whether `languageCodes` in `AudioTranscriptionConfig` is honored by `gemini-3.8-live` and if there's a method to strictly restrict input recognition to Finnish, preventing transcription into other languages. They also inquire about the support for `customVocabulary` in Live sessions and its impact on the model's understanding of audio beyond just transcription. Google's response to these queries will be crucial for developers using the Gemini Live API. Addressing these issues could involve updates to the API's language processing models, improved documentation, or the introduction of new features to enhance language specificity and custom vocabulary support. The resolution of these problems will determine the API's viability for applications targeting diverse linguistic markets.
Beyond the Headlines
The challenges faced by the Gemini Live API in accurately transcribing Finnish speech underscore a broader issue in the development of AI language models: the equitable and effective support for all languages, not just dominant ones. While major languages often receive extensive training data and optimization, smaller language communities can experience less accurate or reliable AI services. This can create a digital divide, where the benefits of advanced AI technologies are not equally distributed. The problem also highlights the complexity of speech recognition, where nuances in pronunciation, dialect, and context can significantly impact accuracy. Ensuring that AI models can handle a wide array of languages with high fidelity is not just a technical challenge but also an ethical imperative to promote inclusivity and accessibility in the digital age.













