Multilingual Voice AI for Pilgrim Assistance: STT Reliability, Bias by Invisibility, and Ethical Deployment in Sacred Environments

The deployment of voice AI assistants in multilingual, safety-critical environments presents challenges that extend beyond prompt engineering. Building on prior work that proposed the SACRED Framework through two proof-of-concept voice assistants deployed at Masjid al-Haram (Fareed, 2026), this study addresses two previously identified gaps: multilingual speech recognition reliability and deployment oversight. Three commercial speech-to-text providers were evaluated within a production voice orchestration platform. Structured language routing achieved reliable transcription for English, Urdu, and Bahasa Indonesia, while Arabic remained unresolvable across four configurations and eight test calls. A structured field level evaluation across 41 calls in Urdu and Bahasa Indonesia found that identification numbers, the field most directly tied to a pilgrim's identity, were correctly relayed to staff in only 15 percent of Urdu calls and 52 percent of Bahasa Indonesia calls at final confirmation, with many of these errors traced specifically to the assistant's response generation rather than to speech recognition itself. A supplementary check using a second speaker produced results in a similar range, offering limited but direct evidence that this unreliability is not solely an artifact of one individual's accent or pronunciation. Cross-referencing language priorities against Hajj demographic data revealed that Bahasa Indonesia, representing the largest single national pilgrim quota of 221,000 pilgrims, was entirely omitted from the original design. This paper conceptualizes this as bias by invisibility, distinct from technical exclusion, and proposes empirical demographic verification as its remedy. Concrete ethical mechanisms, including a data retention policy, safety-critical escalation paths, and a spiritual distress acknowledgment mechanism, are proposed and tested within Shifa. A three-layer governance framework is introduced. Together these findings motivate four revisions to the SACRED Framework and the introduction of Oversight as a seventh principle, extending the framework to SACREDO.