Blue Machines AI Launches Floe to Detect Language Switching in Conversations

Voice-based AI startup Blue Machines AI has launched Floe, a language detection model that identifies whether a customer has actually switched languages during a conversation, rather than treating individual English or regional-language words as a signal to change the agent’s response language.

The model supports 11 languages including English, Hindi, Tamil, Telugu, Gujarati, Kannada, Malayalam, Marathi, Bengali, Odia and Punjabi. It is built for voice and chat interactions across channels such as WhatsApp and SMS.

The company said Floe analyses words, parts of speech, sentence structure, short utterances and previous conversation turns to determine the language an agent should use.

This matters in Indian conversations, where customers often mix English terms such as product names, financial terminology and acronyms into regional-language sentences. For example, “Mera credit card block ho gaya hai” contains the English term “credit card”, but the sentence remains primarily Hindi.

Blue Machines AI said Floe is intended to distinguish between the presence of a language and a customer’s actual language preference. It also considers short responses such as “haan”, “okay”, “correct” and “theek hai” without treating them as automatic language switches.

“In enterprise conversations, language is not a static setting; it is a decision that can change during an interaction,” said Nirmit Parikh, Founder and CEO, Blue Machines AI. “The challenge is not merely to identify the languages being spoken, but to understand which language the customer expects the agent to use.”

Floe is integrated into the company’s conversational AI platform, where its output can inform speech recognition, conversational models, text-to-speech, pronunciation, regional terminology, language-specific prompts, compliance disclosures, escalation, routing and conversation analytics.

The company said the model has recorded internally measured latency of less than 10 milliseconds under production-scale conditions. It is optimised for CPU inference, allowing language detection to run within the real-time conversation without relying on GPU infrastructure for routine inference.

“For enterprise AI, language switching has to work inside the live conversational path without slowing the interaction down,” said Abhishek Ranjan, Chief Technology Officer, Blue Machines AI. “Our model is optimised for CPU inference and has demonstrated latency of less than 10 milliseconds under internally measured production-scale conditions.”

By tracking language across multiple turns instead of classifying each utterance independently, Floe is intended to reduce unnecessary language changes and clarifications while keeping the agent’s responses consistent throughout an interaction.