Nova-3 : it sometimes transcribes audio in the incorrect language (in the middle of a conversation), which gets us to loose the entire meaning of what has been said. Also, the quality of transcription in French and Dutch, compared to gemini, is quite weak for a similar latency (e.g. short utterances like "oui", "non", names, medicine names etc). I've suggest the following on deepgram discord: - Allow to set-up a subset of the languages, to make sure it doesn't transcribe erroneously in a language that will never occur (e.g. "en, fr, nl" would avoid it to transcribe anything in german, for example, although german and dutch would be pretty close and without that setting it might confuse) - In streaming mode, make it more difficult to change language as the conversation evolves. Most use cases will have it detect the right language at the start and shouldn't change after 10 minutes. Might be an option / parameter, to keep it open for conversations that are hold in different languages.