--- title: "Building Voice Input: When Travelers Talk to the AI" description: "Adding voice as an input modality — speech-to-text integration, noisy airport environments, multi-language support, and mixing voice with text in bookings." canonical: https://nowah.xyz/blog/launching-voice-input-travelers-talk-ai lastModified: "2026-08-07T08:24:30.623Z" --- # Building Voice Input: When Travelers Talk to the AI Adding voice as an input modality — speech-to-text integration, noisy airport environments, multi-language support, and mixing voice with text in bookings. A traveler in an airport lounge dictated a complex [multi-city](/blog/multi-city-flight-booking-ai-agents) booking faster than anyone could type it. "Find me a direct flight from JFK to NRT leaving next Thursday morning, business class, with good legroom. Then I need a hotel in Shibuya for five nights. And check if I need a visa." That is 30 seconds of speech. It would have been two minutes of typing and three separate searches on a traditional booking site. Voice is a natural interface for travel. Travelers often have their hands full, literally, carrying luggage, holding passports, wrangling children. And travel requests are often complex enough that speaking them is faster and more natural than typing them. ## The speech-to-text pipeline ![Illustration for this section](https://pics.nowah.xyz/website-media/launch-017-img-1.webp) Voice input flows through a real-time transcription pipeline. The traveler taps the microphone button, speaks, and the speech is converted to text that the agent processes identically to typed messages. The agent does not know or care whether the input came from voice or keyboard. It receives text and responds with text, flight cards, and hotel options. The transcription pipeline handles noise filtering, which matters enormously for travel contexts. Airports are loud. Gate announcements, other conversations, rolling luggage, and background music all compete with the traveler's voice. Our pipeline applies ambient noise filtering before transcription, and we tuned it specifically for the acoustic environments where travelers are most likely to use the app. ## Multi-language voice Travel is inherently multilingual. A Japanese traveler booking a flight to Paris might think in Japanese, speak in English, and want results in Japanese. Our voice input supports the same language set as our text translation, covering dozens of languages with automatic language detection. Automatic detection is important because travelers switch languages mid-conversation without thinking about it. They might start in English and drop in the name of a hotel or neighborhood in the local language. The transcription system needs to handle this gracefully, recognizing that "book me a hotel near Shinjuku-gyoenmae" is not a transcription error but a natural mix of English and Japanese. ## Mixing modalities ![Supporting diagram](https://pics.nowah.xyz/website-media/launch-017-img-2.webp) The most common usage pattern is not pure voice or pure text. It is a hybrid. Travelers start with voice for the initial complex request because speaking is faster. Then they switch to text for confirmations and modifications because tapping "yes" or "change to window seat" is quicker than speaking. Then they might switch back to voice for a follow-up question. We designed the interface to support this seamless switching. The microphone button is always available. The text input is always available. The conversation thread does not distinguish between voice-originated and text-originated messages. The traveler picks whichever modality is most convenient in the moment. ## When voice fails Voice input will fail sometimes. The airport is too loud. The traveler's accent trips up the transcription. A proper noun is misheard. When voice fails, the experience needs to degrade gracefully to text, not to frustration. Our fallback design shows the transcribed text before sending it to the agent, giving the traveler a chance to correct any transcription errors. If the transcription is clearly garbled, a single tap cancels the voice input and switches to the keyboard. The traveler is never trapped in a voice-only interaction that is not working. Voice-specific metrics tell us how the feature is performing: transcription accuracy rate, correction rate when travelers edit the transcription before sending, and booking completion rate for voice-originated conversations versus text-only. Sessions that combine voice and text show higher booking completion than text-only sessions, suggesting that the convenience of voice for complex initial requests removes friction from the booking process. Voice input was designed for one-handed use while carrying luggage. That design constraint drove every decision, from the microphone button placement to the transcription preview to the fallback to text. Travel is a physical activity, and the best travel app adapts to how travelers actually use their phones in the real world. --- Nowah is an AI travel agent that searches and books real flights and hotels through conversation — no filters, no thirty open tabs. [Plan your next trip](https://app.nowah.xyz).