A voice agent can sound excellent in a quiet English demo and still struggle with the conversations a business actually receives. Callers interrupt, switch languages, speak over background noise, and use local names or vocabulary. These conditions make transcription and turn-taking design just as important as the voice’s naturalness.
In April 2026, Deepgram announced general availability of Flux Multilingual, positioning it as a conversational speech-recognition model across ten languages. The company highlights turn-taking and multilingual performance. Those are vendor claims and an invitation to test, not evidence that every deployment will work equally well.
Why turn-taking matters
Traditional speech recognition aims to produce accurate text. A live voice agent has another problem: it must decide when the caller has finished and whether an apparent pause is merely a breath. If it responds too early, it interrupts people. If it waits too long, the conversation feels awkward. Deepgram’s original Flux introduction focused on that conversational problem.
For a real deployment, evaluate the full interaction. A word error rate in isolation does not tell you whether the agent correctly understands the customer’s intent, handles a correction, and completes the task.
Multilingual is more than a language list
A published language list does not automatically settle questions about accents, dialects, or code-switching. A caller might move between Hindi and English in one sentence, pronounce a brand name differently, or give an address with local place names. The only reliable evaluation is with consenting test callers and examples that resemble your expected traffic.
Create a test set that includes language switches, numbers and dates, names, low-quality phone audio, interruptions, and incomplete sentences. Review both the transcript and what the agent did next. If the system confidently books the wrong date, the practical failure is larger than a transcription metric suggests.
A sensible evaluation plan
Compare candidate speech models on the same recordings and live scenarios. Record task completion, recovery after interruptions, median and slow-end response time, and per-call cost. Include the rest of the stack: model reasoning, text-to-speech, network path, and telephony can all add delay. Design a fallback for unsupported or low-confidence conversations, including a straightforward transfer to a person.
Do not infer that a product uses a newly announced model merely because the model is relevant to its category. Provider selection and configuration change over time. InferScale Labs evaluates the components of a custom voice-agent solution against the business’s languages, call types, and operational constraints. You can also experience InferScale Voice and discuss a pilot based on your own calls.
Sources: Deepgram’s Flux Multilingual announcement and original Flux overview. Product capabilities and performance statements are attributed to Deepgram.