As communication becomes more global and immediate, language is becoming a more visible part of the experience. Customers expect quick responses, teams collaborate across borders, and live events bring together people who may speak different languages. In these moments, waiting for translation can mean losing the flow of the conversation.
Real-time AI translation is changing that dynamic by bringing translation into the interaction itself. Speech can be recognised, translated and delivered as a conversation unfolds, allowing people to communicate more naturally across languages. For enterprises, the opportunity is not simply to translate more quickly, but to make multilingual communication work at the pace of the experience.
From translation to communication infrastructure
Language still shapes how people access information, engage with businesses and take part in digital experiences. The opportunity is greatest in live interactions such as customer support, team meetings, virtual events and online communities, where people need to communicate in their preferred language without losing the flow of conversation.
This marks a shift where translation moves from something applied to finished content to part of the communication infrastructure itself. For enterprises, though, that is not simply a matter of adding an AI model. The infrastructure must process speech, translate it and deliver the output fast enough to preserve the natural rhythm of a conversation.
Making real-time truly real-time
In a live conversation, even small delays can disrupt the experience. A translation pipeline typically chains Automatic Speech Recognition (ASR), Machine Translation (MT) and Text-To-Speech (TTS), and each stage adds time. To feel natural, end-to-end latency needs to stay within a few hundred milliseconds, with sub-500ms often cited as a benchmark.
Meeting that target depends less on any single model and more on how the stages work together. Streaming ASR can pass partial results onward, translation can begin before a sentence ends, and speech synthesis can start as soon as output is ready, allowing the different stages to overlap. Network transport matters just as much, since a fast pipeline gains little if audio takes a slow route to reach it.
Cost is the other half of the equation. Continuous, low-latency inference is compute-intensive, and GPU costs can climb quickly across thousands of simultaneous conversations, especially in on-premises or local data-centre deployments. Businesses can manage this by matching model size to the task, improving GPU utilisation through shared capacity, translating selectively and tracking the cost of each conversation.
Deciding where each workload runs matters just as much. Centralised cloud infrastructure can add latency and data transfer costs, while dedicated compute closer to users raises infrastructure spend. A mix of cloud and edge can help organisations balance latency, scalability, cost and data requirements for each use case.
India’s multilingual opportunity
India shows what real-time multilingual communication can offer. People move between languages depending on location, profession, community and context, so enterprise conversations are rarely confined to one language. Code-switching, such as Hinglish, is routine, and a system that translates word by word will quickly sound mechanical. Speech recognition and translation need to understand mixed-language input within a single sentence.
Accuracy also depends on vocabulary. Generic models can stumble on BFSI compliance terms, manufacturing shorthand, IT services jargon and industry acronyms, and in mission-critical workflows, those errors carry consequences. Domain adaptation, including custom glossaries and organisation-specific terminology, can help systems account for the way a particular business actually communicates.
The opportunity extends to customer service, education and collaboration, as well as to Indian businesses expanding abroad that need to engage customers, partners and employees across geographies. For India, it goes beyond translation. It is about creating digital experiences that are inclusive by design and aligned with how people naturally communicate.
Building for enterprise realities
For large enterprises, adoption raises questions of governance, deployment and integration. Industries handling sensitive data may not be able to send every interaction to a public cloud, while regulated organisations may require private or on-premise deployments, including air-gapped environments with strict network isolation.
Those deployments carry their own challenges. Models need regular updates to stay accurate, and in isolated environments that means controlled, validated update cycles and ongoing maintenance rather than continuous cloud delivery. Hybrid architectures can help by allowing enterprises to decide where data is processed while maintaining privacy, security and regulatory controls, including those required under India’s Digital Personal Data Protection framework.
Integration is another common friction point. Translation outputs often need to flow into CRM ticketing, audit logs and analytics dashboards, but doing this can require significant custom middleware. If the translated interaction remains isolated from the rest of the workflow, much of its value is lost.
Real-time communication platforms such as Agora can help bring these pieces together, combining live voice and video with AI capabilities such as translation and transcription. The aim is to make these capabilities part of the communication experience while allowing the resulting data and insights to work within the systems enterprises already use.
Trust will determine adoption
Technology alone will not decide adoption. If a system repeatedly mistranslates specialised terms, misses context or produces awkward phrasing, support agents and employees will revert to manual workarounds. In high-stakes interactions, even a small misunderstanding can damage customer relationships.
Context awareness, domain adaptation and continuous feedback can help. Systems that learn from user corrections can align over time with an organisation’s terminology and communication patterns, while showing users how their feedback improves results can build confidence.
The goal is simple: translation should become reliable enough that people do not have to think about the technology behind it.
What comes next
The future of real-time translation may be one where people barely notice it. Instead of opening a separate tool, you join a meeting, contact support, or attend a live event and communicate in the language you are most comfortable with.
For Agora, the value will come not from simply adding AI features to a platform, but from making them work naturally within live human interactions while meeting enterprise requirements for latency, cost, security, compliance and integration.
The larger opportunity is not just to translate more languages. It is to reduce the friction that keeps people from understanding one another. If real-time AI can do that without getting in the way of the conversation, language can become less of a barrier to participation, customer experience, business growth and human connection.
Authored by Ranga Jagannath, Senior Director – Growth, Agora
(The views expressed are solely those of the author(s) and do not necessarily reflect those of ET Edge CIO&Leader, its management, or members. The author(s) are solely responsible for the originality, accuracy, and authenticity of the content, including proper attribution of third-party material and compliance with applicable copyright and intellectual property laws.)