Real-time translation for travel and hospitality contact centers

Chapter 01

Where can real-time translation improve guest service?

Real-time translation can connect guests with agents who have the expertise to resolve their request, even when they speak different languages. For travel and hospitality businesses, that opens a path to broader service coverage without staffing every language across every shift.

International tourism reached an estimated 1.52 billion arrivals in 2025, up 4% from 2024, according to UN Tourism¹. Serving travelers across markets means supporting questions about availability, booking changes, cancellation terms, and unexpected disruptions. The guest needs a clear answer from someone who can act on the request.

Language-based routing can separate that guest from an available specialist. An overnight agent may know exactly how to change a stay but need another person to explain the options. An interpreter can bridge the exchange, while adding a separate connection and coordination step.

Speech-to-speech translation offers another route. The guest speaks in their preferred language, the agent hears a translation, and the agent’s reply is translated back. The agent continues to use their service skills and booking tools while the technology handles the language exchange.

The opportunity is especially practical for repeatable tasks with clear completion criteria. A reservation change is complete when the correct dates, room, rate, and conditions have been confirmed. A cancellation is complete when the guest understands any charges and receives the appropriate confirmation. These tasks provide a concrete starting point for evaluating both quality and business value.

For operators, the potential benefits include serving more language demand through existing teams, reducing dependence on language-specific transfers, and making specialist knowledge available to more guests. The best initial use case addresses a visible operating constraint, such as an underserved language, limited evening coverage, or repeated handoffs during booking support.

Chapter 02

How does real-time speech translation work?

Real-time speech translation converts spoken input into speech in another language during a live conversation. It must capture what was said, preserve the intended meaning, and deliver an understandable spoken response quickly enough for the exchange to continue.

A common architecture connects automatic speech recognition, machine translation, and speech synthesis. Speech recognition turns the incoming audio into text; translation produces the corresponding wording in the target language; synthesis generates the spoken output. Other architectures translate more directly from speech. Research on low-latency speech translation emphasizes evaluating the complete system, including how it divides incoming speech and how long each processing stage takes. For contact center buyers, that means testing how quickly an accurate translation reaches the listener during a full exchange, rather than judging translation quality or processing speed in isolation².

These stages give buyers a useful way to understand performance. A property name misheard at the input stage can become an incorrect name in the translated response. A correctly recognized sentence can still be translated poorly, and an accurate translation can become difficult to follow if the spoken delivery is unclear or badly timed.

Streaming translation works on incoming speech incrementally. The system has to determine how much speech to hear before it starts translating. Beginning sooner reduces the listener’s wait, but a phrase’s meaning may depend on words that come later. Waiting for those words can improve accuracy while adding delay. While this determination has historically been based on a static “policy module”, research on simultaneous translation shows the advantages of systems that can adapt that timing to the incoming speech, identifying meaningful units rather than translating at fixed intervals³.

In a booking conversation, a guest might begin with one arrival date and correct it before finishing the request. The system needs to carry that correction into the translated exchange. Likewise, an agent explaining a cancellation policy must convey the condition attached to a refund, rather than producing a fluent but incomplete offer.

Chapter 03

How much delay can a guest conversation accommodate?

Delay changes how people manage a conversation. A 2025 study found that added latency increased overlapping speech and pauses between speakers, even when participants did not recognize the delay⁴. For translated guest service, the practical test is whether guests and agents can ask follow-up questions, correct details, and confirm the outcome without repeatedly interrupting each other.

A useful latency target preserves a responsive exchange without sacrificing meaning. Evaluate how long participants wait for translated audio, how the delay behaves during longer explanations, and whether they can recover naturally after an interruption.

For hospitality, the experience should be tested through complete tasks. Can the guest hear two room options and ask a follow-up question? Can an agent confirm a corrected departure date without competing with earlier translated speech? Does a cancellation explanation arrive in a sequence the guest can follow?

Agent coaching can make the interaction easier. Clear phrasing, deliberate pauses, and confirmation of consequential details create a workable rhythm. The goal is a conversation in which both participants know when to listen, when to respond, and whether the request has been understood.

Chapter 04

What should translation preserve to earn guest trust?

Effective translation preserves what the guest and agent mean, including the details and qualifications that determine the outcome. Natural delivery then makes that information easier to follow and keeps the exchange personable.

When making and taking reservations, small wording changes can create different commitments. A nightly rate differs from a total stay price. “Refundable until Friday” differs from “refundable on Friday.” An omitted “not” can reverse a policy explanation. These details need direct evaluation alongside general fluency.

When a guest disputes a cancellation fee, an agent may need to acknowledge frustration while explaining a firm policy. The translated reply should preserve both the acknowledgment and the boundary. Warm delivery cannot compensate for an incorrect policy, while accurate wording still needs to be understandable.

Terminology preparation can also support a smoother exchange. Property names, room categories, and loyalty terms should be included in testing, along with surnames and reservation codes.

Finally, confirm the outcome explicitly. The agent should restate the agreed dates, rate, and applicable conditions before completing a change. That familiar service practice gives both participants an opportunity to correct a misunderstanding while the booking is still in progress.

Chapter 05

How does translation fit the contact center’s technology?

Translation needs to fit the actual call path, agent workspace, and transfer process. Successful deployment keeps the agent focused on serving the guest while making language settings and support controls accessible.

Map a reservation call from start to finish. Identify how guest audio reaches translation, where translated audio enters the agent’s headset, and how the agent’s reply returns to the guest. Include call recording, captions, and any handoff to a property or specialist team. That map makes implementation requirements concrete for IT and operations.

For agents, test the softphone, headset, booking application, and virtual desktop setup they will use in production. A language setting should be easy to verify before a policy explanation begins. A transfer should preserve an understandable experience for the guest. If translation becomes unavailable, the agent needs a defined route to continue service.

For IT and security teams, document where audio is processed and what happens to transcripts or diagnostic information. Separate the translation service’s behavior from recordings retained in the operator’s own systems. Establish the applicable access controls, retention settings, and regional processing arrangements for the selected deployment.

Real-time translation needs to fit the contact center’s existing call flow, from the guest’s first words through transfers and final confirmation. During evaluation, teams should establish where audio is processed, whether audio or transcripts are retained, and how translation connects to their telephony and recording systems. They should also understand how an agent can continue serving the guest if translation becomes unavailable.

These questions give operations, IT, and security teams a shared basis for assessing deployment. Answering them early can reveal requirements that affect cost, rollout effort, or the guest experience before a pilot begins.

Chapter 06

What does research show about speech translation quality?

Recent research shows progress in both preserving meaning and producing natural-sounding translated speech. These are complementary measures of quality; guests need accurate information delivered in a way they can comfortably follow.

Translated speech is becoming more natural

A 2025 study of the Hibiki speech translation model evaluated French-to-English translations for naturalness, including the flow of speech, intonation, and placement of pauses. Human listeners rated the model’s naturalness at 3.73 out of 5, compared with 2.18 for the comparison system and 4.12 for human interpretation⁵.

Those results show how much the listening experience can differ between systems. A translated explanation may preserve every booking detail yet remain difficult to follow if pauses interrupt phrases or the delivery sounds abrupt. More natural speech can make the information easier to absorb while retaining the personal character of a spoken exchange.

For hospitality teams, this makes listening quality a practical evaluation criterion. When an agent explains room options or a cancellation condition, the guest needs to follow the explanation well enough to ask questions and make a decision.

Accuracy and delivery need separate evaluation

A 2026 UTHealth Houston study found that AI translation came close to certified interpreters in conveying meaning and using the right terminology. Human interpreters, however, delivered clearer, more natural speech⁶. This highlights the fact that a translation can get the information right and still be harder to follow.

The distinction is useful for evaluating guest service. Correctly translating a price or departure date addresses one requirement. Delivering that information clearly enough for the guest to understand and confirm it addresses another. A single overall quality score can obscure where a system performs well and where the experience needs improvement.

It’s critical to evaluate translated conversations from both perspectives. Check whether the information survives the translation, then listen to how it reaches the guest. Those complementary assessments give a live pilot a stronger foundation for measuring completed bookings, resolved requests, and guest satisfaction.

Chapter 07

How should a hospitality team evaluate translation?

Evaluate translation through four stages:

1. Define the service opportunity
2. Test realistic conversations
3. Run a controlled pilot
4. Build the expansion case.

Each stage should connect language performance to a booking or service outcome.

7.1 Define the service opportunity

Choose a language pair and service task where limited language coverage creates an observable problem, such as repeated transfers for reservation changes or gaps in evening support. Select a queue with enough call volume to measure performance and work that is reasonably consistent over the evaluation period.

Establish how those requests are handled today and what improvement would justify deployment. For reservations, the primary outcome might be completed bookings; for support, requests resolved without another contact. Identify the communication problem you expect translation to reduce, such as repeated clarification or language-related callbacks.

Record the baseline before launch, including the comparison period and source of each metric. Agree on definitions across participating teams or BPO partners so that a “resolved request” or “completed booking” means the same thing throughout the pilot.

7.2 Test realistic conversations

Before introducing translation on live calls, test the requests agents will handle in the pilot. Include routine exchanges alongside corrected dates, similar-sounding property names, background noise, and explanations containing multiple prices or cancellation conditions.

Have qualified bilingual reviewers assess whether each translation preserves the intended meaning and the details needed to complete the request. Distinguish awkward wording from consequential errors, such as changing a departure date or omitting a refund condition. Use shared examples to keep scoring consistent.

Evaluate complete exchanges, including interruptions and corrections. Can the guest compare room options, ask a follow-up question, and confirm the final booking details? Use these tests to identify terminology needs, refine agent guidance, and resolve quality issues before the live pilot begins.

7.3 Run a controlled live pilot

Launch with a defined agent group and enough call volume to assess performance across the requests you plan to support. Prepare agents with guidance on pacing, confirmation, and escalation, plus a straightforward way to report translation problems.

Compare results with similar agents handling comparable calls during the same period, or use a clearly defined pre-pilot baseline. Account for changes in staffing, booking demand, and call mix. A quieter travel period or a shift toward simpler requests can improve service metrics independently of translation.

Track translation quality alongside completed bookings, repeat contacts, and transfers. Include guest feedback on understanding and ease of communication. A longer call may be worthwhile if it resolves a request that previously required a callback.

Agree on the decision criteria before launch. Define what would justify expansion, further testing, or stopping, including unacceptable errors in dates, prices, or cancellation terms. The pilot should show whether translation can expand language coverage while meeting your standards for guest service and cost.

7.4 Build the expansion case

Use the pilot results to decide where translation should expand and what needs further work. Specify the next language pair, service task, or shift, supported by the quality and business outcomes observed during testing.

Calculate cost per resolved request or completed booking, including agent time, translation charges, and ongoing quality review. Include any continuing interpreter or bilingual support costs. Record coverage gains, such as additional supported hours or fewer language-related callbacks, separately from efficiency improvements.

For projected annual savings, state the assumptions behind call volume, adoption, pricing, and staffing. Identify which costs the business could actually avoid and what further deployment would require. The expansion case should explain what improved during the pilot, what it cost, and why the proposed next step is justified.

Chapter 08

What are the limitations of real-time language translation?

Real-time language translation can struggle with unfamiliar terminology, some language varieties, and exchanges where the intended meaning remains unclear. It also cannot resolve conflicting booking information or make service decisions for an agent. Recognizing these limits gives teams a clear basis for deciding when to clarify, involve a specialist, or switch to another form of language support.

Recognize when clarification is no longer working

A guest repeatedly correcting the same date, giving answers that do not match the question, or expressing uncertainty about an agreed charge may signal a communication problem. Agents do not need to diagnose a translation error to recognize that the conversation is failing to reach a clear outcome.

The American Translators Association cautions that AI-generated translations can appear accurate even when they contain errors that require linguistic expertise to identify, also noting that performance varies across language pairs⁷. Confirming the guest’s intended action and the relevant booking details therefore provides a stronger check than relying on how fluent the translated speech sounds.

Match the support to the problem

Some conversations need a bilingual agent or qualified interpreter. Others need a supervisor who can authorize an exception, a property employee who can verify local information, or an established emergency response procedure. A routine booking call may change direction when a guest reports a safety concern or disputes what they previously agreed to.

The escalation path should address the actual obstacle. Translation can support communication, but it cannot resolve conflicting booking records or give an agent authority to waive a charge. Guests who prefer another form of language assistance should also have a clear route to request it.

Carry the request through the handoff

Before transferring, record the guest’s preferred language, the request, and what remains unresolved. Distinguish confirmed details from information that still needs checking, particularly dates, amounts, and policy conditions. This gives the next person a useful starting point without treating a potentially misunderstood detail as settled.

Review recurring escalations for patterns. Repeated difficulty with a property name may point to a glossary update; problems with a particular language variety or service task may require further testing. Those findings can guide improvements while keeping the scope of deployment aligned with demonstrated performance.

Chapter 09

How can you get started with Sanas?

Sanas Language Translation gives travel and hospitality teams a way to serve more guests through the agents they already employ. Reservations specialists can apply their booking expertise across languages, while guests receive spoken support in their preferred language.

Start with a language need your current service model struggles to meet. That might be a queue with frequent language-related transfers, an evening shift with limited bilingual coverage, or a market where guests regularly wait for a callback. A specific service gap gives your team a clear purpose for evaluating translation and a practical way to measure its value.

Bring that use case into a conversation with Sanas. Identify the language pair, the requests agents need to handle, and the systems they use. From there, define a focused pilot around the outcome you want to improve, whether that is more completed bookings or more requests resolved during the first call.

The business case should reflect the opportunity. Extending support into an underserved shift may be valuable even if average handling time stays the same. Reducing language-related transfers may give guests faster access to someone who can resolve their request. Evaluate those gains alongside translation quality and the total cost of delivering service.

As results support expansion, extend the deployment to additional shifts or service tasks, and evaluate new language pairs against the same standards. That gives your team a practical path to broader coverage, grounded in the conversations your guests need to have.

References

1. UN Tourism. “International tourist arrivals up 4% in 2025 reflecting strong travel demand around the world.” January 2026. Accessed 10.26.26
2. Huber, C., et al. (2023). End-to-end evaluation for low-latency simultaneous speech translation. arXiv:2308.03415. Accessed 10.26.26
3. Zhang, R., He, Z., Wu, H., and Wang, H. “Learning Adaptive Segmentation Policy for End-to-End Simultaneous Translation.” ACL 2022. Accessed 10.26.26
4. Edwards, D. W. (2025). “Impacts of telecommunications latency on the timing of speaker transitions.” Speech Communication, 171, 103226. Accessed 10.26.26
5. Labiausse, T., et al. (2025). “High-Fidelity Simultaneous Speech-To-Speech Translation.” Proceedings of the 42nd International Conference on Machine Learning, 267, 32116–32129. Accessed 10.29.26
6. Singh, U. P., et al. (2026). “Evaluating LingualAI: a prospective validation of AI-based real-time translation against certified human interpreters.” npj Health Systems, 3, 29. Accessed 10.29.30
7. American Translators Association. (2025, May 3). ATA statement on artificial intelligence. Accessed 10.30.26