04 · Call detection
Answering-machine detection
Identify when a call reaches an answering machine so the workflow can choose the appropriate next step.
PersonAnswering machineWorkflow decision
Built for the way people actually speak.
Mixed languages. Financial terminology. Background noise. We build models for the details that make a business conversation work.
01 · Language intelligence
Floe
Follow the speaker. Understand the switch.
Floe distinguishes borrowed words from a real change in language preference. It reads conversational context across turns, so an English phrase inside a Hindi sentence does not automatically change the agent’s language.
Its output helps steer speech recognition, conversational responses and text-to-speech during an interaction.
11
Supported languages
<10 ms
Internally measured CPU inference
English, Hindi, Tamil, Telugu, Gujarati, Kannada, Malayalam, Marathi, Bengali, Odia and Punjabi. Latency varies with configuration and workload.
Illustrative conversation
“Mujhe apni payment date check karni hai.”
Hindi with familiar English termsKeep the conversation in Hindi
“Let’s continue in English, please.”
A clear change in language preferenceSwitch the response to English
Context → Language preference → Response
02 · Speech intelligence
Aurora
Catch the details that move the work forward.
Speech recognition built for India’s banking, financial services and insurance conversations. Aurora handles Indian English, Hindi and Hinglish, with attention to amounts, policy numbers, account references and product terminology.
Adapt it with customer-authorised enterprise data. Deploy through managed cloud, a private VPC or on-premises infrastructure.
Results below are Blue Machines internal benchmarks on BFSI datasets, not guarantees for every deployment. Recognition quality depends on audio, language and workload.
Internal evaluation · BFSI speech
1.51%
English semantic WER
2.43%
Hindi BFSI semantic WER
5.52%
Multilingual semantic WER
4.23%
BFSI entity error rate
WER = word error rate. Lower is better.
03 · Conversational intelligence
Polaris
Keep the conversation moving.
Our conversation LLM uses a mixture-of-experts architecture, activating a subset of experts for each token to support fast, cost-efficient responses in live conversations.
For complex reasoning, the conversation stack can route to frontier models and switch models mid-call.
The right model for the moment
Everyday conversation
Polaris · Responsive dialogueComplex reasoning
Route to a frontier model when neededA connected conversation, across model choices.
Built around the live call
A clearer start. A clearer signal.
05 · Audio processing
Noise cancellation
Reduce background noise in conversational audio to help keep speech clear for the conversation stack.
Incoming audioNoise reductionClearer speech
Built for your conversations.
Bring your languages, audio and business requirements. We’ll help you find the right model and deployment.
Talk to our team →